mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/mm_inline.h>
  16 #include <linux/kthread.h>
  17 #include <linux/khugepaged.h>
  18 #include <linux/freezer.h>
  19 #include <linux/mman.h>
  20 #include <asm/tlb.h>
  21 #include <asm/pgalloc.h>
  22 #include "internal.h"
  23
  24 /*
  25  * By default transparent hugepage support is enabled for all mappings
  26  * and khugepaged scans all mappings. Defrag is only invoked by
  27  * khugepaged hugepage allocations and by page faults inside
  28  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  29  * allocations.
  30  */
  31 unsigned long transparent_hugepage_flags __read_mostly =
  32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  33         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  34 #endif
  35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  36         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  37 #endif
  38         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  39         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
  40
  41 /* default scan 8*512 pte (or vmas) every 30 second */
  42 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  43 static unsigned int khugepaged_pages_collapsed;
  44 static unsigned int khugepaged_full_scans;
  45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  46 /* during fragmentation poll the hugepage allocator once every minute */
  47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  48 static struct task_struct *khugepaged_thread __read_mostly;
  49 static DEFINE_MUTEX(khugepaged_mutex);
  50 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  52 /*
  53  * default collapse hugepages if there is at least one pte mapped like
  54  * it would have happened if the vma was large enough during page
  55  * fault.
  56  */
  57 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  58
  59 static int khugepaged(void *none);
  60 static int mm_slots_hash_init(void);
  61 static int khugepaged_slab_init(void);
  62 static void khugepaged_slab_free(void);
  63
  64 #define MM_SLOTS_HASH_HEADS 1024
  65 static struct hlist_head *mm_slots_hash __read_mostly;
  66 static struct kmem_cache *mm_slot_cache __read_mostly;
  67
  68 /**
  69  * struct mm_slot - hash lookup from mm to mm_slot
  70  * @hash: hash collision list
  71  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  72  * @mm: the mm that this information is valid for
  73  */
  74 struct mm_slot {
  75         struct hlist_node hash;
  76         struct list_head mm_node;
  77         struct mm_struct *mm;
  78 };
  79
  80 /**
  81  * struct khugepaged_scan - cursor for scanning
  82  * @mm_head: the head of the mm list to scan
  83  * @mm_slot: the current mm_slot we are scanning
  84  * @address: the next address inside that to be scanned
  85  *
  86  * There is only the one khugepaged_scan instance of this cursor structure.
  87  */
  88 struct khugepaged_scan {
  89         struct list_head mm_head;
  90         struct mm_slot *mm_slot;
  91         unsigned long address;
  92 };
  93 static struct khugepaged_scan khugepaged_scan = {
  94         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
  95 };
  96
  97
  98 static int set_recommended_min_free_kbytes(void)
  99 {
 100         struct zone *zone;
 101         int nr_zones = 0;
 102         unsigned long recommended_min;
 103         extern int min_free_kbytes;
 104
 105         if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
 106                       &transparent_hugepage_flags) &&
 107             !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 108                       &transparent_hugepage_flags))
 109                 return 0;
 110
 111         for_each_populated_zone(zone)
 112                 nr_zones++;
 113
 114         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 115         recommended_min = pageblock_nr_pages * nr_zones * 2;
 116
 117         /*
 118          * Make sure that on average at least two pageblocks are almost free
 119          * of another type, one for a migratetype to fall back to and a
 120          * second to avoid subsequent fallbacks of other types There are 3
 121          * MIGRATE_TYPES we care about.
 122          */
 123         recommended_min += pageblock_nr_pages * nr_zones *
 124                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 125
 126         /* don't ever allow to reserve more than 5% of the lowmem */
 127         recommended_min = min(recommended_min,
 128                               (unsigned long) nr_free_buffer_pages() / 20);
 129         recommended_min <<= (PAGE_SHIFT-10);
 130
 131         if (recommended_min > min_free_kbytes)
 132                 min_free_kbytes = recommended_min;
 133         setup_per_zone_wmarks();
 134         return 0;
 135 }
 136 late_initcall(set_recommended_min_free_kbytes);
 137
 138 static int start_khugepaged(void)
 139 {
 140         int err = 0;
 141         if (khugepaged_enabled()) {
 142                 int wakeup;
 143                 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
 144                         err = -ENOMEM;
 145                         goto out;
 146                 }
 147                 mutex_lock(&khugepaged_mutex);
 148                 if (!khugepaged_thread)
 149                         khugepaged_thread = kthread_run(khugepaged, NULL,
 150                                                         "khugepaged");
 151                 if (unlikely(IS_ERR(khugepaged_thread))) {
 152                         printk(KERN_ERR
 153                                "khugepaged: kthread_run(khugepaged) failed\n");
 154                         err = PTR_ERR(khugepaged_thread);
 155                         khugepaged_thread = NULL;
 156                 }
 157                 wakeup = !list_empty(&khugepaged_scan.mm_head);
 158                 mutex_unlock(&khugepaged_mutex);
 159                 if (wakeup)
 160                         wake_up_interruptible(&khugepaged_wait);
 161
 162                 set_recommended_min_free_kbytes();
 163         } else
 164                 /* wakeup to exit */
 165                 wake_up_interruptible(&khugepaged_wait);
 166 out:
 167         return err;
 168 }
 169
 170 #ifdef CONFIG_SYSFS
 171
 172 static ssize_t double_flag_show(struct kobject *kobj,
 173                                 struct kobj_attribute *attr, char *buf,
 174                                 enum transparent_hugepage_flag enabled,
 175                                 enum transparent_hugepage_flag req_madv)
 176 {
 177         if (test_bit(enabled, &transparent_hugepage_flags)) {
 178                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 179                 return sprintf(buf, "[always] madvise never\n");
 180         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 181                 return sprintf(buf, "always [madvise] never\n");
 182         else
 183                 return sprintf(buf, "always madvise [never]\n");
 184 }
 185 static ssize_t double_flag_store(struct kobject *kobj,
 186                                  struct kobj_attribute *attr,
 187                                  const char *buf, size_t count,
 188                                  enum transparent_hugepage_flag enabled,
 189                                  enum transparent_hugepage_flag req_madv)
 190 {
 191         if (!memcmp("always", buf,
 192                     min(sizeof("always")-1, count))) {
 193                 set_bit(enabled, &transparent_hugepage_flags);
 194                 clear_bit(req_madv, &transparent_hugepage_flags);
 195         } else if (!memcmp("madvise", buf,
 196                            min(sizeof("madvise")-1, count))) {
 197                 clear_bit(enabled, &transparent_hugepage_flags);
 198                 set_bit(req_madv, &transparent_hugepage_flags);
 199         } else if (!memcmp("never", buf,
 200                            min(sizeof("never")-1, count))) {
 201                 clear_bit(enabled, &transparent_hugepage_flags);
 202                 clear_bit(req_madv, &transparent_hugepage_flags);
 203         } else
 204                 return -EINVAL;
 205
 206         return count;
 207 }
 208
 209 static ssize_t enabled_show(struct kobject *kobj,
 210                             struct kobj_attribute *attr, char *buf)
 211 {
 212         return double_flag_show(kobj, attr, buf,
 213                                 TRANSPARENT_HUGEPAGE_FLAG,
 214                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 215 }
 216 static ssize_t enabled_store(struct kobject *kobj,
 217                              struct kobj_attribute *attr,
 218                              const char *buf, size_t count)
 219 {
 220         ssize_t ret;
 221
 222         ret = double_flag_store(kobj, attr, buf, count,
 223                                 TRANSPARENT_HUGEPAGE_FLAG,
 224                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 225
 226         if (ret > 0) {
 227                 int err = start_khugepaged();
 228                 if (err)
 229                         ret = err;
 230         }
 231
 232         if (ret > 0 &&
 233             (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
 234                       &transparent_hugepage_flags) ||
 235              test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 236                       &transparent_hugepage_flags)))
 237                 set_recommended_min_free_kbytes();
 238
 239         return ret;
 240 }
 241 static struct kobj_attribute enabled_attr =
 242         __ATTR(enabled, 0644, enabled_show, enabled_store);
 243
 244 static ssize_t single_flag_show(struct kobject *kobj,
 245                                 struct kobj_attribute *attr, char *buf,
 246                                 enum transparent_hugepage_flag flag)
 247 {
 248         return sprintf(buf, "%d\n",
 249                        !!test_bit(flag, &transparent_hugepage_flags));
 250 }
 251
 252 static ssize_t single_flag_store(struct kobject *kobj,
 253                                  struct kobj_attribute *attr,
 254                                  const char *buf, size_t count,
 255                                  enum transparent_hugepage_flag flag)
 256 {
 257         unsigned long value;
 258         int ret;
 259
 260         ret = kstrtoul(buf, 10, &value);
 261         if (ret < 0)
 262                 return ret;
 263         if (value > 1)
 264                 return -EINVAL;
 265
 266         if (value)
 267                 set_bit(flag, &transparent_hugepage_flags);
 268         else
 269                 clear_bit(flag, &transparent_hugepage_flags);
 270
 271         return count;
 272 }
 273
 274 /*
 275  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 276  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 277  * memory just to allocate one more hugepage.
 278  */
 279 static ssize_t defrag_show(struct kobject *kobj,
 280                            struct kobj_attribute *attr, char *buf)
 281 {
 282         return double_flag_show(kobj, attr, buf,
 283                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 284                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 285 }
 286 static ssize_t defrag_store(struct kobject *kobj,
 287                             struct kobj_attribute *attr,
 288                             const char *buf, size_t count)
 289 {
 290         return double_flag_store(kobj, attr, buf, count,
 291                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 292                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 293 }
 294 static struct kobj_attribute defrag_attr =
 295         __ATTR(defrag, 0644, defrag_show, defrag_store);
 296
 297 #ifdef CONFIG_DEBUG_VM
 298 static ssize_t debug_cow_show(struct kobject *kobj,
 299                                 struct kobj_attribute *attr, char *buf)
 300 {
 301         return single_flag_show(kobj, attr, buf,
 302                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 303 }
 304 static ssize_t debug_cow_store(struct kobject *kobj,
 305                                struct kobj_attribute *attr,
 306                                const char *buf, size_t count)
 307 {
 308         return single_flag_store(kobj, attr, buf, count,
 309                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 310 }
 311 static struct kobj_attribute debug_cow_attr =
 312         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 313 #endif /* CONFIG_DEBUG_VM */
 314
 315 static struct attribute *hugepage_attr[] = {
 316         &enabled_attr.attr,
 317         &defrag_attr.attr,
 318 #ifdef CONFIG_DEBUG_VM
 319         &debug_cow_attr.attr,
 320 #endif
 321         NULL,
 322 };
 323
 324 static struct attribute_group hugepage_attr_group = {
 325         .attrs = hugepage_attr,
 326 };
 327
 328 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 329                                          struct kobj_attribute *attr,
 330                                          char *buf)
 331 {
 332         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 333 }
 334
 335 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 336                                           struct kobj_attribute *attr,
 337                                           const char *buf, size_t count)
 338 {
 339         unsigned long msecs;
 340         int err;
 341
 342         err = strict_strtoul(buf, 10, &msecs);
 343         if (err || msecs > UINT_MAX)
 344                 return -EINVAL;
 345
 346         khugepaged_scan_sleep_millisecs = msecs;
 347         wake_up_interruptible(&khugepaged_wait);
 348
 349         return count;
 350 }
 351 static struct kobj_attribute scan_sleep_millisecs_attr =
 352         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 353                scan_sleep_millisecs_store);
 354
 355 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 356                                           struct kobj_attribute *attr,
 357                                           char *buf)
 358 {
 359         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 360 }
 361
 362 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 363                                            struct kobj_attribute *attr,
 364                                            const char *buf, size_t count)
 365 {
 366         unsigned long msecs;
 367         int err;
 368
 369         err = strict_strtoul(buf, 10, &msecs);
 370         if (err || msecs > UINT_MAX)
 371                 return -EINVAL;
 372
 373         khugepaged_alloc_sleep_millisecs = msecs;
 374         wake_up_interruptible(&khugepaged_wait);
 375
 376         return count;
 377 }
 378 static struct kobj_attribute alloc_sleep_millisecs_attr =
 379         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 380                alloc_sleep_millisecs_store);
 381
 382 static ssize_t pages_to_scan_show(struct kobject *kobj,
 383                                   struct kobj_attribute *attr,
 384                                   char *buf)
 385 {
 386         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 387 }
 388 static ssize_t pages_to_scan_store(struct kobject *kobj,
 389                                    struct kobj_attribute *attr,
 390                                    const char *buf, size_t count)
 391 {
 392         int err;
 393         unsigned long pages;
 394
 395         err = strict_strtoul(buf, 10, &pages);
 396         if (err || !pages || pages > UINT_MAX)
 397                 return -EINVAL;
 398
 399         khugepaged_pages_to_scan = pages;
 400
 401         return count;
 402 }
 403 static struct kobj_attribute pages_to_scan_attr =
 404         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 405                pages_to_scan_store);
 406
 407 static ssize_t pages_collapsed_show(struct kobject *kobj,
 408                                     struct kobj_attribute *attr,
 409                                     char *buf)
 410 {
 411         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 412 }
 413 static struct kobj_attribute pages_collapsed_attr =
 414         __ATTR_RO(pages_collapsed);
 415
 416 static ssize_t full_scans_show(struct kobject *kobj,
 417                                struct kobj_attribute *attr,
 418                                char *buf)
 419 {
 420         return sprintf(buf, "%u\n", khugepaged_full_scans);
 421 }
 422 static struct kobj_attribute full_scans_attr =
 423         __ATTR_RO(full_scans);
 424
 425 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 426                                       struct kobj_attribute *attr, char *buf)
 427 {
 428         return single_flag_show(kobj, attr, buf,
 429                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 430 }
 431 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 432                                        struct kobj_attribute *attr,
 433                                        const char *buf, size_t count)
 434 {
 435         return single_flag_store(kobj, attr, buf, count,
 436                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 437 }
 438 static struct kobj_attribute khugepaged_defrag_attr =
 439         __ATTR(defrag, 0644, khugepaged_defrag_show,
 440                khugepaged_defrag_store);
 441
 442 /*
 443  * max_ptes_none controls if khugepaged should collapse hugepages over
 444  * any unmapped ptes in turn potentially increasing the memory
 445  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 446  * reduce the available free memory in the system as it
 447  * runs. Increasing max_ptes_none will instead potentially reduce the
 448  * free memory in the system during the khugepaged scan.
 449  */
 450 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 451                                              struct kobj_attribute *attr,
 452                                              char *buf)
 453 {
 454         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 455 }
 456 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 457                                               struct kobj_attribute *attr,
 458                                               const char *buf, size_t count)
 459 {
 460         int err;
 461         unsigned long max_ptes_none;
 462
 463         err = strict_strtoul(buf, 10, &max_ptes_none);
 464         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 465                 return -EINVAL;
 466
 467         khugepaged_max_ptes_none = max_ptes_none;
 468
 469         return count;
 470 }
 471 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 472         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 473                khugepaged_max_ptes_none_store);
 474
 475 static struct attribute *khugepaged_attr[] = {
 476         &khugepaged_defrag_attr.attr,
 477         &khugepaged_max_ptes_none_attr.attr,
 478         &pages_to_scan_attr.attr,
 479         &pages_collapsed_attr.attr,
 480         &full_scans_attr.attr,
 481         &scan_sleep_millisecs_attr.attr,
 482         &alloc_sleep_millisecs_attr.attr,
 483         NULL,
 484 };
 485
 486 static struct attribute_group khugepaged_attr_group = {
 487         .attrs = khugepaged_attr,
 488         .name = "khugepaged",
 489 };
 490
 491 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 492 {
 493         int err;
 494
 495         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 496         if (unlikely(!*hugepage_kobj)) {
 497                 printk(KERN_ERR "hugepage: failed kobject create\n");
 498                 return -ENOMEM;
 499         }
 500
 501         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 502         if (err) {
 503                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 504                 goto delete_obj;
 505         }
 506
 507         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 508         if (err) {
 509                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 510                 goto remove_hp_group;
 511         }
 512
 513         return 0;
 514
 515 remove_hp_group:
 516         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 517 delete_obj:
 518         kobject_put(*hugepage_kobj);
 519         return err;
 520 }
 521
 522 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 523 {
 524         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 525         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 526         kobject_put(hugepage_kobj);
 527 }
 528 #else
 529 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 530 {
 531         return 0;
 532 }
 533
 534 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 535 {
 536 }
 537 #endif /* CONFIG_SYSFS */
 538
 539 static int __init hugepage_init(void)
 540 {
 541         int err;
 542         struct kobject *hugepage_kobj;
 543
 544         if (!has_transparent_hugepage()) {
 545                 transparent_hugepage_flags = 0;
 546                 return -EINVAL;
 547         }
 548
 549         err = hugepage_init_sysfs(&hugepage_kobj);
 550         if (err)
 551                 return err;
 552
 553         err = khugepaged_slab_init();
 554         if (err)
 555                 goto out;
 556
 557         err = mm_slots_hash_init();
 558         if (err) {
 559                 khugepaged_slab_free();
 560                 goto out;
 561         }
 562
 563         /*
 564          * By default disable transparent hugepages on smaller systems,
 565          * where the extra memory used could hurt more than TLB overhead
 566          * is likely to save.  The admin can still enable it through /sys.
 567          */
 568         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 569                 transparent_hugepage_flags = 0;
 570
 571         start_khugepaged();
 572
 573         set_recommended_min_free_kbytes();
 574
 575         return 0;
 576 out:
 577         hugepage_exit_sysfs(hugepage_kobj);
 578         return err;
 579 }
 580 module_init(hugepage_init)
 581
 582 static int __init setup_transparent_hugepage(char *str)
 583 {
 584         int ret = 0;
 585         if (!str)
 586                 goto out;
 587         if (!strcmp(str, "always")) {
 588                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 589                         &transparent_hugepage_flags);
 590                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 591                           &transparent_hugepage_flags);
 592                 ret = 1;
 593         } else if (!strcmp(str, "madvise")) {
 594                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 595                           &transparent_hugepage_flags);
 596                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 597                         &transparent_hugepage_flags);
 598                 ret = 1;
 599         } else if (!strcmp(str, "never")) {
 600                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 601                           &transparent_hugepage_flags);
 602                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 603                           &transparent_hugepage_flags);
 604                 ret = 1;
 605         }
 606 out:
 607         if (!ret)
 608                 printk(KERN_WARNING
 609                        "transparent_hugepage= cannot parse, ignored\n");
 610         return ret;
 611 }
 612 __setup("transparent_hugepage=", setup_transparent_hugepage);
 613
 614 static void prepare_pmd_huge_pte(pgtable_t pgtable,
 615                                  struct mm_struct *mm)
 616 {
 617         assert_spin_locked(&mm->page_table_lock);
 618
 619         /* FIFO */
 620         if (!mm->pmd_huge_pte)
 621                 INIT_LIST_HEAD(&pgtable->lru);
 622         else
 623                 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
 624         mm->pmd_huge_pte = pgtable;
 625 }
 626
 627 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 628 {
 629         if (likely(vma->vm_flags & VM_WRITE))
 630                 pmd = pmd_mkwrite(pmd);
 631         return pmd;
 632 }
 633
 634 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 635                                         struct vm_area_struct *vma,
 636                                         unsigned long haddr, pmd_t *pmd,
 637                                         struct page *page)
 638 {
 639         pgtable_t pgtable;
 640
 641         VM_BUG_ON(!PageCompound(page));
 642         pgtable = pte_alloc_one(mm, haddr);
 643         if (unlikely(!pgtable))
 644                 return VM_FAULT_OOM;
 645
 646         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 647         __SetPageUptodate(page);
 648
 649         spin_lock(&mm->page_table_lock);
 650         if (unlikely(!pmd_none(*pmd))) {
 651                 spin_unlock(&mm->page_table_lock);
 652                 mem_cgroup_uncharge_page(page);
 653                 put_page(page);
 654                 pte_free(mm, pgtable);
 655         } else {
 656                 pmd_t entry;
 657                 entry = mk_pmd(page, vma->vm_page_prot);
 658                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 659                 entry = pmd_mkhuge(entry);
 660                 /*
 661                  * The spinlocking to take the lru_lock inside
 662                  * page_add_new_anon_rmap() acts as a full memory
 663                  * barrier to be sure clear_huge_page writes become
 664                  * visible after the set_pmd_at() write.
 665                  */
 666                 page_add_new_anon_rmap(page, vma, haddr);
 667                 set_pmd_at(mm, haddr, pmd, entry);
 668                 prepare_pmd_huge_pte(pgtable, mm);
 669                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 670                 mm->nr_ptes++;
 671                 spin_unlock(&mm->page_table_lock);
 672         }
 673
 674         return 0;
 675 }
 676
 677 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 678 {
 679         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 680 }
 681
 682 static inline struct page *alloc_hugepage_vma(int defrag,
 683                                               struct vm_area_struct *vma,
 684                                               unsigned long haddr, int nd,
 685                                               gfp_t extra_gfp)
 686 {
 687         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 688                                HPAGE_PMD_ORDER, vma, haddr, nd);
 689 }
 690
 691 #ifndef CONFIG_NUMA
 692 static inline struct page *alloc_hugepage(int defrag)
 693 {
 694         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 695                            HPAGE_PMD_ORDER);
 696 }
 697 #endif
 698
 699 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 700                                unsigned long address, pmd_t *pmd,
 701                                unsigned int flags)
 702 {
 703         struct page *page;
 704         unsigned long haddr = address & HPAGE_PMD_MASK;
 705         pte_t *pte;
 706
 707         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 708                 if (unlikely(anon_vma_prepare(vma)))
 709                         return VM_FAULT_OOM;
 710                 if (unlikely(khugepaged_enter(vma)))
 711                         return VM_FAULT_OOM;
 712                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 713                                           vma, haddr, numa_node_id(), 0);
 714                 if (unlikely(!page)) {
 715                         count_vm_event(THP_FAULT_FALLBACK);
 716                         goto out;
 717                 }
 718                 count_vm_event(THP_FAULT_ALLOC);
 719                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 720                         put_page(page);
 721                         goto out;
 722                 }
 723                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 724                                                           page))) {
 725                         mem_cgroup_uncharge_page(page);
 726                         put_page(page);
 727                         goto out;
 728                 }
 729
 730                 return 0;
 731         }
 732 out:
 733         /*
 734          * Use __pte_alloc instead of pte_alloc_map, because we can't
 735          * run pte_offset_map on the pmd, if an huge pmd could
 736          * materialize from under us from a different thread.
 737          */
 738         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
 739                 return VM_FAULT_OOM;
 740         /* if an huge pmd materialized from under us just retry later */
 741         if (unlikely(pmd_trans_huge(*pmd)))
 742                 return 0;
 743         /*
 744          * A regular pmd is established and it can't morph into a huge pmd
 745          * from under us anymore at this point because we hold the mmap_sem
 746          * read mode and khugepaged takes it in write mode. So now it's
 747          * safe to run pte_offset_map().
 748          */
 749         pte = pte_offset_map(pmd, address);
 750         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 751 }
 752
 753 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 754                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 755                   struct vm_area_struct *vma)
 756 {
 757         struct page *src_page;
 758         pmd_t pmd;
 759         pgtable_t pgtable;
 760         int ret;
 761
 762         ret = -ENOMEM;
 763         pgtable = pte_alloc_one(dst_mm, addr);
 764         if (unlikely(!pgtable))
 765                 goto out;
 766
 767         spin_lock(&dst_mm->page_table_lock);
 768         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 769
 770         ret = -EAGAIN;
 771         pmd = *src_pmd;
 772         if (unlikely(!pmd_trans_huge(pmd))) {
 773                 pte_free(dst_mm, pgtable);
 774                 goto out_unlock;
 775         }
 776         if (unlikely(pmd_trans_splitting(pmd))) {
 777                 /* split huge page running from under us */
 778                 spin_unlock(&src_mm->page_table_lock);
 779                 spin_unlock(&dst_mm->page_table_lock);
 780                 pte_free(dst_mm, pgtable);
 781
 782                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 783                 goto out;
 784         }
 785         src_page = pmd_page(pmd);
 786         VM_BUG_ON(!PageHead(src_page));
 787         get_page(src_page);
 788         page_dup_rmap(src_page);
 789         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 790
 791         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 792         pmd = pmd_mkold(pmd_wrprotect(pmd));
 793         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 794         prepare_pmd_huge_pte(pgtable, dst_mm);
 795         dst_mm->nr_ptes++;
 796
 797         ret = 0;
 798 out_unlock:
 799         spin_unlock(&src_mm->page_table_lock);
 800         spin_unlock(&dst_mm->page_table_lock);
 801 out:
 802         return ret;
 803 }
 804
 805 /* no "address" argument so destroys page coloring of some arch */
 806 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
 807 {
 808         pgtable_t pgtable;
 809
 810         assert_spin_locked(&mm->page_table_lock);
 811
 812         /* FIFO */
 813         pgtable = mm->pmd_huge_pte;
 814         if (list_empty(&pgtable->lru))
 815                 mm->pmd_huge_pte = NULL;
 816         else {
 817                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
 818                                               struct page, lru);
 819                 list_del(&pgtable->lru);
 820         }
 821         return pgtable;
 822 }
 823
 824 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 825                                         struct vm_area_struct *vma,
 826                                         unsigned long address,
 827                                         pmd_t *pmd, pmd_t orig_pmd,
 828                                         struct page *page,
 829                                         unsigned long haddr)
 830 {
 831         pgtable_t pgtable;
 832         pmd_t _pmd;
 833         int ret = 0, i;
 834         struct page **pages;
 835
 836         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 837                         GFP_KERNEL);
 838         if (unlikely(!pages)) {
 839                 ret |= VM_FAULT_OOM;
 840                 goto out;
 841         }
 842
 843         for (i = 0; i < HPAGE_PMD_NR; i++) {
 844                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 845                                                __GFP_OTHER_NODE,
 846                                                vma, address, page_to_nid(page));
 847                 if (unlikely(!pages[i] ||
 848                              mem_cgroup_newpage_charge(pages[i], mm,
 849                                                        GFP_KERNEL))) {
 850                         if (pages[i])
 851                                 put_page(pages[i]);
 852                         mem_cgroup_uncharge_start();
 853                         while (--i >= 0) {
 854                                 mem_cgroup_uncharge_page(pages[i]);
 855                                 put_page(pages[i]);
 856                         }
 857                         mem_cgroup_uncharge_end();
 858                         kfree(pages);
 859                         ret |= VM_FAULT_OOM;
 860                         goto out;
 861                 }
 862         }
 863
 864         for (i = 0; i < HPAGE_PMD_NR; i++) {
 865                 copy_user_highpage(pages[i], page + i,
 866                                    haddr + PAGE_SIZE * i, vma);
 867                 __SetPageUptodate(pages[i]);
 868                 cond_resched();
 869         }
 870
 871         spin_lock(&mm->page_table_lock);
 872         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 873                 goto out_free_pages;
 874         VM_BUG_ON(!PageHead(page));
 875
 876         pmdp_clear_flush_notify(vma, haddr, pmd);
 877         /* leave pmd empty until pte is filled */
 878
 879         pgtable = get_pmd_huge_pte(mm);
 880         pmd_populate(mm, &_pmd, pgtable);
 881
 882         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 883                 pte_t *pte, entry;
 884                 entry = mk_pte(pages[i], vma->vm_page_prot);
 885                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 886                 page_add_new_anon_rmap(pages[i], vma, haddr);
 887                 pte = pte_offset_map(&_pmd, haddr);
 888                 VM_BUG_ON(!pte_none(*pte));
 889                 set_pte_at(mm, haddr, pte, entry);
 890                 pte_unmap(pte);
 891         }
 892         kfree(pages);
 893
 894         smp_wmb(); /* make pte visible before pmd */
 895         pmd_populate(mm, pmd, pgtable);
 896         page_remove_rmap(page);
 897         spin_unlock(&mm->page_table_lock);
 898
 899         ret |= VM_FAULT_WRITE;
 900         put_page(page);
 901
 902 out:
 903         return ret;
 904
 905 out_free_pages:
 906         spin_unlock(&mm->page_table_lock);
 907         mem_cgroup_uncharge_start();
 908         for (i = 0; i < HPAGE_PMD_NR; i++) {
 909                 mem_cgroup_uncharge_page(pages[i]);
 910                 put_page(pages[i]);
 911         }
 912         mem_cgroup_uncharge_end();
 913         kfree(pages);
 914         goto out;
 915 }
 916
 917 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 918                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 919 {
 920         int ret = 0;
 921         struct page *page, *new_page;
 922         unsigned long haddr;
 923
 924         VM_BUG_ON(!vma->anon_vma);
 925         spin_lock(&mm->page_table_lock);
 926         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 927                 goto out_unlock;
 928
 929         page = pmd_page(orig_pmd);
 930         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 931         haddr = address & HPAGE_PMD_MASK;
 932         if (page_mapcount(page) == 1) {
 933                 pmd_t entry;
 934                 entry = pmd_mkyoung(orig_pmd);
 935                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 936                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 937                         update_mmu_cache(vma, address, entry);
 938                 ret |= VM_FAULT_WRITE;
 939                 goto out_unlock;
 940         }
 941         get_page(page);
 942         spin_unlock(&mm->page_table_lock);
 943
 944         if (transparent_hugepage_enabled(vma) &&
 945             !transparent_hugepage_debug_cow())
 946                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 947                                               vma, haddr, numa_node_id(), 0);
 948         else
 949                 new_page = NULL;
 950
 951         if (unlikely(!new_page)) {
 952                 count_vm_event(THP_FAULT_FALLBACK);
 953                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 954                                                    pmd, orig_pmd, page, haddr);
 955                 if (ret & VM_FAULT_OOM)
 956                         split_huge_page(page);
 957                 put_page(page);
 958                 goto out;
 959         }
 960         count_vm_event(THP_FAULT_ALLOC);
 961
 962         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 963                 put_page(new_page);
 964                 split_huge_page(page);
 965                 put_page(page);
 966                 ret |= VM_FAULT_OOM;
 967                 goto out;
 968         }
 969
 970         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 971         __SetPageUptodate(new_page);
 972
 973         spin_lock(&mm->page_table_lock);
 974         put_page(page);
 975         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 976                 spin_unlock(&mm->page_table_lock);
 977                 mem_cgroup_uncharge_page(new_page);
 978                 put_page(new_page);
 979                 goto out;
 980         } else {
 981                 pmd_t entry;
 982                 VM_BUG_ON(!PageHead(page));
 983                 entry = mk_pmd(new_page, vma->vm_page_prot);
 984                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 985                 entry = pmd_mkhuge(entry);
 986                 pmdp_clear_flush_notify(vma, haddr, pmd);
 987                 page_add_new_anon_rmap(new_page, vma, haddr);
 988                 set_pmd_at(mm, haddr, pmd, entry);
 989                 update_mmu_cache(vma, address, entry);
 990                 page_remove_rmap(page);
 991                 put_page(page);
 992                 ret |= VM_FAULT_WRITE;
 993         }
 994 out_unlock:
 995         spin_unlock(&mm->page_table_lock);
 996 out:
 997         return ret;
 998 }
 999
1000 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
1001                                    unsigned long addr,
1002                                    pmd_t *pmd,
1003                                    unsigned int flags)
1004 {
1005         struct page *page = NULL;
1006
1007         assert_spin_locked(&mm->page_table_lock);
1008
1009         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1010                 goto out;
1011
1012         page = pmd_page(*pmd);
1013         VM_BUG_ON(!PageHead(page));
1014         if (flags & FOLL_TOUCH) {
1015                 pmd_t _pmd;
1016                 /*
1017                  * We should set the dirty bit only for FOLL_WRITE but
1018                  * for now the dirty bit in the pmd is meaningless.
1019                  * And if the dirty bit will become meaningful and
1020                  * we'll only set it with FOLL_WRITE, an atomic
1021                  * set_bit will be required on the pmd to set the
1022                  * young bit, instead of the current set_pmd_at.
1023                  */
1024                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1025                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1026         }
1027         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1028         VM_BUG_ON(!PageCompound(page));
1029         if (flags & FOLL_GET)
1030                 get_page_foll(page);
1031
1032 out:
1033         return page;
1034 }
1035
1036 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1037                  pmd_t *pmd, unsigned long addr)
1038 {
1039         int ret = 0;
1040
1041         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1042                 struct page *page;
1043                 pgtable_t pgtable;
1044                 pgtable = get_pmd_huge_pte(tlb->mm);
1045                 page = pmd_page(*pmd);
1046                 pmd_clear(pmd);
1047                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1048                 page_remove_rmap(page);
1049                 VM_BUG_ON(page_mapcount(page) < 0);
1050                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1051                 VM_BUG_ON(!PageHead(page));
1052                 tlb->mm->nr_ptes--;
1053                 spin_unlock(&tlb->mm->page_table_lock);
1054                 tlb_remove_page(tlb, page);
1055                 pte_free(tlb->mm, pgtable);
1056                 ret = 1;
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         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1068                 /*
1069                  * All logical pages in the range are present
1070                  * if backed by a huge page.
1071                  */
1072                 spin_unlock(&vma->vm_mm->page_table_lock);
1073                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1074                 ret = 1;
1075         }
1076
1077         return ret;
1078 }
1079
1080 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1081                   unsigned long old_addr,
1082                   unsigned long new_addr, unsigned long old_end,
1083                   pmd_t *old_pmd, pmd_t *new_pmd)
1084 {
1085         int ret = 0;
1086         pmd_t pmd;
1087
1088         struct mm_struct *mm = vma->vm_mm;
1089
1090         if ((old_addr & ~HPAGE_PMD_MASK) ||
1091             (new_addr & ~HPAGE_PMD_MASK) ||
1092             old_end - old_addr < HPAGE_PMD_SIZE ||
1093             (new_vma->vm_flags & VM_NOHUGEPAGE))
1094                 goto out;
1095
1096         /*
1097          * The destination pmd shouldn't be established, free_pgtables()
1098          * should have release it.
1099          */
1100         if (WARN_ON(!pmd_none(*new_pmd))) {
1101                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1102                 goto out;
1103         }
1104
1105         ret = __pmd_trans_huge_lock(old_pmd, vma);
1106         if (ret == 1) {
1107                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1108                 VM_BUG_ON(!pmd_none(*new_pmd));
1109                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1110                 spin_unlock(&mm->page_table_lock);
1111         }
1112 out:
1113         return ret;
1114 }
1115
1116 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1117                 unsigned long addr, pgprot_t newprot)
1118 {
1119         struct mm_struct *mm = vma->vm_mm;
1120         int ret = 0;
1121
1122         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1123                 pmd_t entry;
1124                 entry = pmdp_get_and_clear(mm, addr, pmd);
1125                 entry = pmd_modify(entry, newprot);
1126                 set_pmd_at(mm, addr, pmd, entry);
1127                 spin_unlock(&vma->vm_mm->page_table_lock);
1128                 ret = 1;
1129         }
1130
1131         return ret;
1132 }
1133
1134 /*
1135  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1136  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1137  *
1138  * Note that if it returns 1, this routine returns without unlocking page
1139  * table locks. So callers must unlock them.
1140  */
1141 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1142 {
1143         spin_lock(&vma->vm_mm->page_table_lock);
1144         if (likely(pmd_trans_huge(*pmd))) {
1145                 if (unlikely(pmd_trans_splitting(*pmd))) {
1146                         spin_unlock(&vma->vm_mm->page_table_lock);
1147                         wait_split_huge_page(vma->anon_vma, pmd);
1148                         return -1;
1149                 } else {
1150                         /* Thp mapped by 'pmd' is stable, so we can
1151                          * handle it as it is. */
1152                         return 1;
1153                 }
1154         }
1155         spin_unlock(&vma->vm_mm->page_table_lock);
1156         return 0;
1157 }
1158
1159 pmd_t *page_check_address_pmd(struct page *page,
1160                               struct mm_struct *mm,
1161                               unsigned long address,
1162                               enum page_check_address_pmd_flag flag)
1163 {
1164         pgd_t *pgd;
1165         pud_t *pud;
1166         pmd_t *pmd, *ret = NULL;
1167
1168         if (address & ~HPAGE_PMD_MASK)
1169                 goto out;
1170
1171         pgd = pgd_offset(mm, address);
1172         if (!pgd_present(*pgd))
1173                 goto out;
1174
1175         pud = pud_offset(pgd, address);
1176         if (!pud_present(*pud))
1177                 goto out;
1178
1179         pmd = pmd_offset(pud, address);
1180         if (pmd_none(*pmd))
1181                 goto out;
1182         if (pmd_page(*pmd) != page)
1183                 goto out;
1184         /*
1185          * split_vma() may create temporary aliased mappings. There is
1186          * no risk as long as all huge pmd are found and have their
1187          * splitting bit set before __split_huge_page_refcount
1188          * runs. Finding the same huge pmd more than once during the
1189          * same rmap walk is not a problem.
1190          */
1191         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1192             pmd_trans_splitting(*pmd))
1193                 goto out;
1194         if (pmd_trans_huge(*pmd)) {
1195                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1196                           !pmd_trans_splitting(*pmd));
1197                 ret = pmd;
1198         }
1199 out:
1200         return ret;
1201 }
1202
1203 static int __split_huge_page_splitting(struct page *page,
1204                                        struct vm_area_struct *vma,
1205                                        unsigned long address)
1206 {
1207         struct mm_struct *mm = vma->vm_mm;
1208         pmd_t *pmd;
1209         int ret = 0;
1210
1211         spin_lock(&mm->page_table_lock);
1212         pmd = page_check_address_pmd(page, mm, address,
1213                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1214         if (pmd) {
1215                 /*
1216                  * We can't temporarily set the pmd to null in order
1217                  * to split it, the pmd must remain marked huge at all
1218                  * times or the VM won't take the pmd_trans_huge paths
1219                  * and it won't wait on the anon_vma->root->mutex to
1220                  * serialize against split_huge_page*.
1221                  */
1222                 pmdp_splitting_flush_notify(vma, address, pmd);
1223                 ret = 1;
1224         }
1225         spin_unlock(&mm->page_table_lock);
1226
1227         return ret;
1228 }
1229
1230 static void __split_huge_page_refcount(struct page *page)
1231 {
1232         int i;
1233         struct zone *zone = page_zone(page);
1234         struct lruvec *lruvec;
1235         int tail_count = 0;
1236
1237         /* prevent PageLRU to go away from under us, and freeze lru stats */
1238         spin_lock_irq(&zone->lru_lock);
1239         lruvec = mem_cgroup_page_lruvec(page, zone);
1240
1241         compound_lock(page);
1242         /* complete memcg works before add pages to LRU */
1243         mem_cgroup_split_huge_fixup(page);
1244
1245         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1246                 struct page *page_tail = page + i;
1247
1248                 /* tail_page->_mapcount cannot change */
1249                 BUG_ON(page_mapcount(page_tail) < 0);
1250                 tail_count += page_mapcount(page_tail);
1251                 /* check for overflow */
1252                 BUG_ON(tail_count < 0);
1253                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1254                 /*
1255                  * tail_page->_count is zero and not changing from
1256                  * under us. But get_page_unless_zero() may be running
1257                  * from under us on the tail_page. If we used
1258                  * atomic_set() below instead of atomic_add(), we
1259                  * would then run atomic_set() concurrently with
1260                  * get_page_unless_zero(), and atomic_set() is
1261                  * implemented in C not using locked ops. spin_unlock
1262                  * on x86 sometime uses locked ops because of PPro
1263                  * errata 66, 92, so unless somebody can guarantee
1264                  * atomic_set() here would be safe on all archs (and
1265                  * not only on x86), it's safer to use atomic_add().
1266                  */
1267                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1268                            &page_tail->_count);
1269
1270                 /* after clearing PageTail the gup refcount can be released */
1271                 smp_mb();
1272
1273                 /*
1274                  * retain hwpoison flag of the poisoned tail page:
1275                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1276                  *   by the memory-failure.
1277                  */
1278                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1279                 page_tail->flags |= (page->flags &
1280                                      ((1L << PG_referenced) |
1281                                       (1L << PG_swapbacked) |
1282                                       (1L << PG_mlocked) |
1283                                       (1L << PG_uptodate)));
1284                 page_tail->flags |= (1L << PG_dirty);
1285
1286                 /* clear PageTail before overwriting first_page */
1287                 smp_wmb();
1288
1289                 /*
1290                  * __split_huge_page_splitting() already set the
1291                  * splitting bit in all pmd that could map this
1292                  * hugepage, that will ensure no CPU can alter the
1293                  * mapcount on the head page. The mapcount is only
1294                  * accounted in the head page and it has to be
1295                  * transferred to all tail pages in the below code. So
1296                  * for this code to be safe, the split the mapcount
1297                  * can't change. But that doesn't mean userland can't
1298                  * keep changing and reading the page contents while
1299                  * we transfer the mapcount, so the pmd splitting
1300                  * status is achieved setting a reserved bit in the
1301                  * pmd, not by clearing the present bit.
1302                 */
1303                 page_tail->_mapcount = page->_mapcount;
1304
1305                 BUG_ON(page_tail->mapping);
1306                 page_tail->mapping = page->mapping;
1307
1308                 page_tail->index = page->index + i;
1309
1310                 BUG_ON(!PageAnon(page_tail));
1311                 BUG_ON(!PageUptodate(page_tail));
1312                 BUG_ON(!PageDirty(page_tail));
1313                 BUG_ON(!PageSwapBacked(page_tail));
1314
1315                 lru_add_page_tail(page, page_tail, lruvec);
1316         }
1317         atomic_sub(tail_count, &page->_count);
1318         BUG_ON(atomic_read(&page->_count) <= 0);
1319
1320         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1321         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1322
1323         ClearPageCompound(page);
1324         compound_unlock(page);
1325         spin_unlock_irq(&zone->lru_lock);
1326
1327         for (i = 1; i < HPAGE_PMD_NR; i++) {
1328                 struct page *page_tail = page + i;
1329                 BUG_ON(page_count(page_tail) <= 0);
1330                 /*
1331                  * Tail pages may be freed if there wasn't any mapping
1332                  * like if add_to_swap() is running on a lru page that
1333                  * had its mapping zapped. And freeing these pages
1334                  * requires taking the lru_lock so we do the put_page
1335                  * of the tail pages after the split is complete.
1336                  */
1337                 put_page(page_tail);
1338         }
1339
1340         /*
1341          * Only the head page (now become a regular page) is required
1342          * to be pinned by the caller.
1343          */
1344         BUG_ON(page_count(page) <= 0);
1345 }
1346
1347 static int __split_huge_page_map(struct page *page,
1348                                  struct vm_area_struct *vma,
1349                                  unsigned long address)
1350 {
1351         struct mm_struct *mm = vma->vm_mm;
1352         pmd_t *pmd, _pmd;
1353         int ret = 0, i;
1354         pgtable_t pgtable;
1355         unsigned long haddr;
1356
1357         spin_lock(&mm->page_table_lock);
1358         pmd = page_check_address_pmd(page, mm, address,
1359                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1360         if (pmd) {
1361                 pgtable = get_pmd_huge_pte(mm);
1362                 pmd_populate(mm, &_pmd, pgtable);
1363
1364                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1365                      i++, haddr += PAGE_SIZE) {
1366                         pte_t *pte, entry;
1367                         BUG_ON(PageCompound(page+i));
1368                         entry = mk_pte(page + i, vma->vm_page_prot);
1369                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1370                         if (!pmd_write(*pmd))
1371                                 entry = pte_wrprotect(entry);
1372                         else
1373                                 BUG_ON(page_mapcount(page) != 1);
1374                         if (!pmd_young(*pmd))
1375                                 entry = pte_mkold(entry);
1376                         pte = pte_offset_map(&_pmd, haddr);
1377                         BUG_ON(!pte_none(*pte));
1378                         set_pte_at(mm, haddr, pte, entry);
1379                         pte_unmap(pte);
1380                 }
1381
1382                 smp_wmb(); /* make pte visible before pmd */
1383                 /*
1384                  * Up to this point the pmd is present and huge and
1385                  * userland has the whole access to the hugepage
1386                  * during the split (which happens in place). If we
1387                  * overwrite the pmd with the not-huge version
1388                  * pointing to the pte here (which of course we could
1389                  * if all CPUs were bug free), userland could trigger
1390                  * a small page size TLB miss on the small sized TLB
1391                  * while the hugepage TLB entry is still established
1392                  * in the huge TLB. Some CPU doesn't like that. See
1393                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1394                  * Erratum 383 on page 93. Intel should be safe but is
1395                  * also warns that it's only safe if the permission
1396                  * and cache attributes of the two entries loaded in
1397                  * the two TLB is identical (which should be the case
1398                  * here). But it is generally safer to never allow
1399                  * small and huge TLB entries for the same virtual
1400                  * address to be loaded simultaneously. So instead of
1401                  * doing "pmd_populate(); flush_tlb_range();" we first
1402                  * mark the current pmd notpresent (atomically because
1403                  * here the pmd_trans_huge and pmd_trans_splitting
1404                  * must remain set at all times on the pmd until the
1405                  * split is complete for this pmd), then we flush the
1406                  * SMP TLB and finally we write the non-huge version
1407                  * of the pmd entry with pmd_populate.
1408                  */
1409                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1410                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1411                 pmd_populate(mm, pmd, pgtable);
1412                 ret = 1;
1413         }
1414         spin_unlock(&mm->page_table_lock);
1415
1416         return ret;
1417 }
1418
1419 /* must be called with anon_vma->root->mutex hold */
1420 static void __split_huge_page(struct page *page,
1421                               struct anon_vma *anon_vma)
1422 {
1423         int mapcount, mapcount2;
1424         struct anon_vma_chain *avc;
1425
1426         BUG_ON(!PageHead(page));
1427         BUG_ON(PageTail(page));
1428
1429         mapcount = 0;
1430         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1431                 struct vm_area_struct *vma = avc->vma;
1432                 unsigned long addr = vma_address(page, vma);
1433                 BUG_ON(is_vma_temporary_stack(vma));
1434                 if (addr == -EFAULT)
1435                         continue;
1436                 mapcount += __split_huge_page_splitting(page, vma, addr);
1437         }
1438         /*
1439          * It is critical that new vmas are added to the tail of the
1440          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1441          * and establishes a child pmd before
1442          * __split_huge_page_splitting() freezes the parent pmd (so if
1443          * we fail to prevent copy_huge_pmd() from running until the
1444          * whole __split_huge_page() is complete), we will still see
1445          * the newly established pmd of the child later during the
1446          * walk, to be able to set it as pmd_trans_splitting too.
1447          */
1448         if (mapcount != page_mapcount(page))
1449                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1450                        mapcount, page_mapcount(page));
1451         BUG_ON(mapcount != page_mapcount(page));
1452
1453         __split_huge_page_refcount(page);
1454
1455         mapcount2 = 0;
1456         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1457                 struct vm_area_struct *vma = avc->vma;
1458                 unsigned long addr = vma_address(page, vma);
1459                 BUG_ON(is_vma_temporary_stack(vma));
1460                 if (addr == -EFAULT)
1461                         continue;
1462                 mapcount2 += __split_huge_page_map(page, vma, addr);
1463         }
1464         if (mapcount != mapcount2)
1465                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1466                        mapcount, mapcount2, page_mapcount(page));
1467         BUG_ON(mapcount != mapcount2);
1468 }
1469
1470 int split_huge_page(struct page *page)
1471 {
1472         struct anon_vma *anon_vma;
1473         int ret = 1;
1474
1475         BUG_ON(!PageAnon(page));
1476         anon_vma = page_lock_anon_vma(page);
1477         if (!anon_vma)
1478                 goto out;
1479         ret = 0;
1480         if (!PageCompound(page))
1481                 goto out_unlock;
1482
1483         BUG_ON(!PageSwapBacked(page));
1484         __split_huge_page(page, anon_vma);
1485         count_vm_event(THP_SPLIT);
1486
1487         BUG_ON(PageCompound(page));
1488 out_unlock:
1489         page_unlock_anon_vma(anon_vma);
1490 out:
1491         return ret;
1492 }
1493
1494 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1495                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1496
1497 int hugepage_madvise(struct vm_area_struct *vma,
1498                      unsigned long *vm_flags, int advice)
1499 {
1500         switch (advice) {
1501         case MADV_HUGEPAGE:
1502                 /*
1503                  * Be somewhat over-protective like KSM for now!
1504                  */
1505                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1506                         return -EINVAL;
1507                 *vm_flags &= ~VM_NOHUGEPAGE;
1508                 *vm_flags |= VM_HUGEPAGE;
1509                 /*
1510                  * If the vma become good for khugepaged to scan,
1511                  * register it here without waiting a page fault that
1512                  * may not happen any time soon.
1513                  */
1514                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1515                         return -ENOMEM;
1516                 break;
1517         case MADV_NOHUGEPAGE:
1518                 /*
1519                  * Be somewhat over-protective like KSM for now!
1520                  */
1521                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1522                         return -EINVAL;
1523                 *vm_flags &= ~VM_HUGEPAGE;
1524                 *vm_flags |= VM_NOHUGEPAGE;
1525                 /*
1526                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1527                  * this vma even if we leave the mm registered in khugepaged if
1528                  * it got registered before VM_NOHUGEPAGE was set.
1529                  */
1530                 break;
1531         }
1532
1533         return 0;
1534 }
1535
1536 static int __init khugepaged_slab_init(void)
1537 {
1538         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1539                                           sizeof(struct mm_slot),
1540                                           __alignof__(struct mm_slot), 0, NULL);
1541         if (!mm_slot_cache)
1542                 return -ENOMEM;
1543
1544         return 0;
1545 }
1546
1547 static void __init khugepaged_slab_free(void)
1548 {
1549         kmem_cache_destroy(mm_slot_cache);
1550         mm_slot_cache = NULL;
1551 }
1552
1553 static inline struct mm_slot *alloc_mm_slot(void)
1554 {
1555         if (!mm_slot_cache)     /* initialization failed */
1556                 return NULL;
1557         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1558 }
1559
1560 static inline void free_mm_slot(struct mm_slot *mm_slot)
1561 {
1562         kmem_cache_free(mm_slot_cache, mm_slot);
1563 }
1564
1565 static int __init mm_slots_hash_init(void)
1566 {
1567         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1568                                 GFP_KERNEL);
1569         if (!mm_slots_hash)
1570                 return -ENOMEM;
1571         return 0;
1572 }
1573
1574 #if 0
1575 static void __init mm_slots_hash_free(void)
1576 {
1577         kfree(mm_slots_hash);
1578         mm_slots_hash = NULL;
1579 }
1580 #endif
1581
1582 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1583 {
1584         struct mm_slot *mm_slot;
1585         struct hlist_head *bucket;
1586         struct hlist_node *node;
1587
1588         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1589                                 % MM_SLOTS_HASH_HEADS];
1590         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1591                 if (mm == mm_slot->mm)
1592                         return mm_slot;
1593         }
1594         return NULL;
1595 }
1596
1597 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1598                                     struct mm_slot *mm_slot)
1599 {
1600         struct hlist_head *bucket;
1601
1602         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1603                                 % MM_SLOTS_HASH_HEADS];
1604         mm_slot->mm = mm;
1605         hlist_add_head(&mm_slot->hash, bucket);
1606 }
1607
1608 static inline int khugepaged_test_exit(struct mm_struct *mm)
1609 {
1610         return atomic_read(&mm->mm_users) == 0;
1611 }
1612
1613 int __khugepaged_enter(struct mm_struct *mm)
1614 {
1615         struct mm_slot *mm_slot;
1616         int wakeup;
1617
1618         mm_slot = alloc_mm_slot();
1619         if (!mm_slot)
1620                 return -ENOMEM;
1621
1622         /* __khugepaged_exit() must not run from under us */
1623         VM_BUG_ON(khugepaged_test_exit(mm));
1624         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1625                 free_mm_slot(mm_slot);
1626                 return 0;
1627         }
1628
1629         spin_lock(&khugepaged_mm_lock);
1630         insert_to_mm_slots_hash(mm, mm_slot);
1631         /*
1632          * Insert just behind the scanning cursor, to let the area settle
1633          * down a little.
1634          */
1635         wakeup = list_empty(&khugepaged_scan.mm_head);
1636         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1637         spin_unlock(&khugepaged_mm_lock);
1638
1639         atomic_inc(&mm->mm_count);
1640         if (wakeup)
1641                 wake_up_interruptible(&khugepaged_wait);
1642
1643         return 0;
1644 }
1645
1646 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1647 {
1648         unsigned long hstart, hend;
1649         if (!vma->anon_vma)
1650                 /*
1651                  * Not yet faulted in so we will register later in the
1652                  * page fault if needed.
1653                  */
1654                 return 0;
1655         if (vma->vm_ops)
1656                 /* khugepaged not yet working on file or special mappings */
1657                 return 0;
1658         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1659         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1660         hend = vma->vm_end & HPAGE_PMD_MASK;
1661         if (hstart < hend)
1662                 return khugepaged_enter(vma);
1663         return 0;
1664 }
1665
1666 void __khugepaged_exit(struct mm_struct *mm)
1667 {
1668         struct mm_slot *mm_slot;
1669         int free = 0;
1670
1671         spin_lock(&khugepaged_mm_lock);
1672         mm_slot = get_mm_slot(mm);
1673         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1674                 hlist_del(&mm_slot->hash);
1675                 list_del(&mm_slot->mm_node);
1676                 free = 1;
1677         }
1678         spin_unlock(&khugepaged_mm_lock);
1679
1680         if (free) {
1681                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1682                 free_mm_slot(mm_slot);
1683                 mmdrop(mm);
1684         } else if (mm_slot) {
1685                 /*
1686                  * This is required to serialize against
1687                  * khugepaged_test_exit() (which is guaranteed to run
1688                  * under mmap sem read mode). Stop here (after we
1689                  * return all pagetables will be destroyed) until
1690                  * khugepaged has finished working on the pagetables
1691                  * under the mmap_sem.
1692                  */
1693                 down_write(&mm->mmap_sem);
1694                 up_write(&mm->mmap_sem);
1695         }
1696 }
1697
1698 static void release_pte_page(struct page *page)
1699 {
1700         /* 0 stands for page_is_file_cache(page) == false */
1701         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1702         unlock_page(page);
1703         putback_lru_page(page);
1704 }
1705
1706 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1707 {
1708         while (--_pte >= pte) {
1709                 pte_t pteval = *_pte;
1710                 if (!pte_none(pteval))
1711                         release_pte_page(pte_page(pteval));
1712         }
1713 }
1714
1715 static void release_all_pte_pages(pte_t *pte)
1716 {
1717         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1718 }
1719
1720 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1721                                         unsigned long address,
1722                                         pte_t *pte)
1723 {
1724         struct page *page;
1725         pte_t *_pte;
1726         int referenced = 0, isolated = 0, none = 0;
1727         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1728              _pte++, address += PAGE_SIZE) {
1729                 pte_t pteval = *_pte;
1730                 if (pte_none(pteval)) {
1731                         if (++none <= khugepaged_max_ptes_none)
1732                                 continue;
1733                         else {
1734                                 release_pte_pages(pte, _pte);
1735                                 goto out;
1736                         }
1737                 }
1738                 if (!pte_present(pteval) || !pte_write(pteval)) {
1739                         release_pte_pages(pte, _pte);
1740                         goto out;
1741                 }
1742                 page = vm_normal_page(vma, address, pteval);
1743                 if (unlikely(!page)) {
1744                         release_pte_pages(pte, _pte);
1745                         goto out;
1746                 }
1747                 VM_BUG_ON(PageCompound(page));
1748                 BUG_ON(!PageAnon(page));
1749                 VM_BUG_ON(!PageSwapBacked(page));
1750
1751                 /* cannot use mapcount: can't collapse if there's a gup pin */
1752                 if (page_count(page) != 1) {
1753                         release_pte_pages(pte, _pte);
1754                         goto out;
1755                 }
1756                 /*
1757                  * We can do it before isolate_lru_page because the
1758                  * page can't be freed from under us. NOTE: PG_lock
1759                  * is needed to serialize against split_huge_page
1760                  * when invoked from the VM.
1761                  */
1762                 if (!trylock_page(page)) {
1763                         release_pte_pages(pte, _pte);
1764                         goto out;
1765                 }
1766                 /*
1767                  * Isolate the page to avoid collapsing an hugepage
1768                  * currently in use by the VM.
1769                  */
1770                 if (isolate_lru_page(page)) {
1771                         unlock_page(page);
1772                         release_pte_pages(pte, _pte);
1773                         goto out;
1774                 }
1775                 /* 0 stands for page_is_file_cache(page) == false */
1776                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1777                 VM_BUG_ON(!PageLocked(page));
1778                 VM_BUG_ON(PageLRU(page));
1779
1780                 /* If there is no mapped pte young don't collapse the page */
1781                 if (pte_young(pteval) || PageReferenced(page) ||
1782                     mmu_notifier_test_young(vma->vm_mm, address))
1783                         referenced = 1;
1784         }
1785         if (unlikely(!referenced))
1786                 release_all_pte_pages(pte);
1787         else
1788                 isolated = 1;
1789 out:
1790         return isolated;
1791 }
1792
1793 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1794                                       struct vm_area_struct *vma,
1795                                       unsigned long address,
1796                                       spinlock_t *ptl)
1797 {
1798         pte_t *_pte;
1799         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1800                 pte_t pteval = *_pte;
1801                 struct page *src_page;
1802
1803                 if (pte_none(pteval)) {
1804                         clear_user_highpage(page, address);
1805                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1806                 } else {
1807                         src_page = pte_page(pteval);
1808                         copy_user_highpage(page, src_page, address, vma);
1809                         VM_BUG_ON(page_mapcount(src_page) != 1);
1810                         release_pte_page(src_page);
1811                         /*
1812                          * ptl mostly unnecessary, but preempt has to
1813                          * be disabled to update the per-cpu stats
1814                          * inside page_remove_rmap().
1815                          */
1816                         spin_lock(ptl);
1817                         /*
1818                          * paravirt calls inside pte_clear here are
1819                          * superfluous.
1820                          */
1821                         pte_clear(vma->vm_mm, address, _pte);
1822                         page_remove_rmap(src_page);
1823                         spin_unlock(ptl);
1824                         free_page_and_swap_cache(src_page);
1825                 }
1826
1827                 address += PAGE_SIZE;
1828                 page++;
1829         }
1830 }
1831
1832 static void collapse_huge_page(struct mm_struct *mm,
1833                                unsigned long address,
1834                                struct page **hpage,
1835                                struct vm_area_struct *vma,
1836                                int node)
1837 {
1838         pgd_t *pgd;
1839         pud_t *pud;
1840         pmd_t *pmd, _pmd;
1841         pte_t *pte;
1842         pgtable_t pgtable;
1843         struct page *new_page;
1844         spinlock_t *ptl;
1845         int isolated;
1846         unsigned long hstart, hend;
1847
1848         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1849 #ifndef CONFIG_NUMA
1850         up_read(&mm->mmap_sem);
1851         VM_BUG_ON(!*hpage);
1852         new_page = *hpage;
1853 #else
1854         VM_BUG_ON(*hpage);
1855         /*
1856          * Allocate the page while the vma is still valid and under
1857          * the mmap_sem read mode so there is no memory allocation
1858          * later when we take the mmap_sem in write mode. This is more
1859          * friendly behavior (OTOH it may actually hide bugs) to
1860          * filesystems in userland with daemons allocating memory in
1861          * the userland I/O paths.  Allocating memory with the
1862          * mmap_sem in read mode is good idea also to allow greater
1863          * scalability.
1864          */
1865         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1866                                       node, __GFP_OTHER_NODE);
1867
1868         /*
1869          * After allocating the hugepage, release the mmap_sem read lock in
1870          * preparation for taking it in write mode.
1871          */
1872         up_read(&mm->mmap_sem);
1873         if (unlikely(!new_page)) {
1874                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1875                 *hpage = ERR_PTR(-ENOMEM);
1876                 return;
1877         }
1878 #endif
1879
1880         count_vm_event(THP_COLLAPSE_ALLOC);
1881         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1882 #ifdef CONFIG_NUMA
1883                 put_page(new_page);
1884 #endif
1885                 return;
1886         }
1887
1888         /*
1889          * Prevent all access to pagetables with the exception of
1890          * gup_fast later hanlded by the ptep_clear_flush and the VM
1891          * handled by the anon_vma lock + PG_lock.
1892          */
1893         down_write(&mm->mmap_sem);
1894         if (unlikely(khugepaged_test_exit(mm)))
1895                 goto out;
1896
1897         vma = find_vma(mm, address);
1898         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1899         hend = vma->vm_end & HPAGE_PMD_MASK;
1900         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1901                 goto out;
1902
1903         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1904             (vma->vm_flags & VM_NOHUGEPAGE))
1905                 goto out;
1906
1907         if (!vma->anon_vma || vma->vm_ops)
1908                 goto out;
1909         if (is_vma_temporary_stack(vma))
1910                 goto out;
1911         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1912
1913         pgd = pgd_offset(mm, address);
1914         if (!pgd_present(*pgd))
1915                 goto out;
1916
1917         pud = pud_offset(pgd, address);
1918         if (!pud_present(*pud))
1919                 goto out;
1920
1921         pmd = pmd_offset(pud, address);
1922         /* pmd can't go away or become huge under us */
1923         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1924                 goto out;
1925
1926         anon_vma_lock(vma->anon_vma);
1927
1928         pte = pte_offset_map(pmd, address);
1929         ptl = pte_lockptr(mm, pmd);
1930
1931         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1932         /*
1933          * After this gup_fast can't run anymore. This also removes
1934          * any huge TLB entry from the CPU so we won't allow
1935          * huge and small TLB entries for the same virtual address
1936          * to avoid the risk of CPU bugs in that area.
1937          */
1938         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1939         spin_unlock(&mm->page_table_lock);
1940
1941         spin_lock(ptl);
1942         isolated = __collapse_huge_page_isolate(vma, address, pte);
1943         spin_unlock(ptl);
1944
1945         if (unlikely(!isolated)) {
1946                 pte_unmap(pte);
1947                 spin_lock(&mm->page_table_lock);
1948                 BUG_ON(!pmd_none(*pmd));
1949                 set_pmd_at(mm, address, pmd, _pmd);
1950                 spin_unlock(&mm->page_table_lock);
1951                 anon_vma_unlock(vma->anon_vma);
1952                 goto out;
1953         }
1954
1955         /*
1956          * All pages are isolated and locked so anon_vma rmap
1957          * can't run anymore.
1958          */
1959         anon_vma_unlock(vma->anon_vma);
1960
1961         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1962         pte_unmap(pte);
1963         __SetPageUptodate(new_page);
1964         pgtable = pmd_pgtable(_pmd);
1965         VM_BUG_ON(page_count(pgtable) != 1);
1966         VM_BUG_ON(page_mapcount(pgtable) != 0);
1967
1968         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1969         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1970         _pmd = pmd_mkhuge(_pmd);
1971
1972         /*
1973          * spin_lock() below is not the equivalent of smp_wmb(), so
1974          * this is needed to avoid the copy_huge_page writes to become
1975          * visible after the set_pmd_at() write.
1976          */
1977         smp_wmb();
1978
1979         spin_lock(&mm->page_table_lock);
1980         BUG_ON(!pmd_none(*pmd));
1981         page_add_new_anon_rmap(new_page, vma, address);
1982         set_pmd_at(mm, address, pmd, _pmd);
1983         update_mmu_cache(vma, address, _pmd);
1984         prepare_pmd_huge_pte(pgtable, mm);
1985         spin_unlock(&mm->page_table_lock);
1986
1987 #ifndef CONFIG_NUMA
1988         *hpage = NULL;
1989 #endif
1990         khugepaged_pages_collapsed++;
1991 out_up_write:
1992         up_write(&mm->mmap_sem);
1993         return;
1994
1995 out:
1996         mem_cgroup_uncharge_page(new_page);
1997 #ifdef CONFIG_NUMA
1998         put_page(new_page);
1999 #endif
2000         goto out_up_write;
2001 }
2002
2003 static int khugepaged_scan_pmd(struct mm_struct *mm,
2004                                struct vm_area_struct *vma,
2005                                unsigned long address,
2006                                struct page **hpage)
2007 {
2008         pgd_t *pgd;
2009         pud_t *pud;
2010         pmd_t *pmd;
2011         pte_t *pte, *_pte;
2012         int ret = 0, referenced = 0, none = 0;
2013         struct page *page;
2014         unsigned long _address;
2015         spinlock_t *ptl;
2016         int node = -1;
2017
2018         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2019
2020         pgd = pgd_offset(mm, address);
2021         if (!pgd_present(*pgd))
2022                 goto out;
2023
2024         pud = pud_offset(pgd, address);
2025         if (!pud_present(*pud))
2026                 goto out;
2027
2028         pmd = pmd_offset(pud, address);
2029         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2030                 goto out;
2031
2032         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2033         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2034              _pte++, _address += PAGE_SIZE) {
2035                 pte_t pteval = *_pte;
2036                 if (pte_none(pteval)) {
2037                         if (++none <= khugepaged_max_ptes_none)
2038                                 continue;
2039                         else
2040                                 goto out_unmap;
2041                 }
2042                 if (!pte_present(pteval) || !pte_write(pteval))
2043                         goto out_unmap;
2044                 page = vm_normal_page(vma, _address, pteval);
2045                 if (unlikely(!page))
2046                         goto out_unmap;
2047                 /*
2048                  * Chose the node of the first page. This could
2049                  * be more sophisticated and look at more pages,
2050                  * but isn't for now.
2051                  */
2052                 if (node == -1)
2053                         node = page_to_nid(page);
2054                 VM_BUG_ON(PageCompound(page));
2055                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2056                         goto out_unmap;
2057                 /* cannot use mapcount: can't collapse if there's a gup pin */
2058                 if (page_count(page) != 1)
2059                         goto out_unmap;
2060                 if (pte_young(pteval) || PageReferenced(page) ||
2061                     mmu_notifier_test_young(vma->vm_mm, address))
2062                         referenced = 1;
2063         }
2064         if (referenced)
2065                 ret = 1;
2066 out_unmap:
2067         pte_unmap_unlock(pte, ptl);
2068         if (ret)
2069                 /* collapse_huge_page will return with the mmap_sem released */
2070                 collapse_huge_page(mm, address, hpage, vma, node);
2071 out:
2072         return ret;
2073 }
2074
2075 static void collect_mm_slot(struct mm_slot *mm_slot)
2076 {
2077         struct mm_struct *mm = mm_slot->mm;
2078
2079         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2080
2081         if (khugepaged_test_exit(mm)) {
2082                 /* free mm_slot */
2083                 hlist_del(&mm_slot->hash);
2084                 list_del(&mm_slot->mm_node);
2085
2086                 /*
2087                  * Not strictly needed because the mm exited already.
2088                  *
2089                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2090                  */
2091
2092                 /* khugepaged_mm_lock actually not necessary for the below */
2093                 free_mm_slot(mm_slot);
2094                 mmdrop(mm);
2095         }
2096 }
2097
2098 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2099                                             struct page **hpage)
2100         __releases(&khugepaged_mm_lock)
2101         __acquires(&khugepaged_mm_lock)
2102 {
2103         struct mm_slot *mm_slot;
2104         struct mm_struct *mm;
2105         struct vm_area_struct *vma;
2106         int progress = 0;
2107
2108         VM_BUG_ON(!pages);
2109         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2110
2111         if (khugepaged_scan.mm_slot)
2112                 mm_slot = khugepaged_scan.mm_slot;
2113         else {
2114                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2115                                      struct mm_slot, mm_node);
2116                 khugepaged_scan.address = 0;
2117                 khugepaged_scan.mm_slot = mm_slot;
2118         }
2119         spin_unlock(&khugepaged_mm_lock);
2120
2121         mm = mm_slot->mm;
2122         down_read(&mm->mmap_sem);
2123         if (unlikely(khugepaged_test_exit(mm)))
2124                 vma = NULL;
2125         else
2126                 vma = find_vma(mm, khugepaged_scan.address);
2127
2128         progress++;
2129         for (; vma; vma = vma->vm_next) {
2130                 unsigned long hstart, hend;
2131
2132                 cond_resched();
2133                 if (unlikely(khugepaged_test_exit(mm))) {
2134                         progress++;
2135                         break;
2136                 }
2137
2138                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2139                      !khugepaged_always()) ||
2140                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2141                 skip:
2142                         progress++;
2143                         continue;
2144                 }
2145                 if (!vma->anon_vma || vma->vm_ops)
2146                         goto skip;
2147                 if (is_vma_temporary_stack(vma))
2148                         goto skip;
2149                 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2150
2151                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2152                 hend = vma->vm_end & HPAGE_PMD_MASK;
2153                 if (hstart >= hend)
2154                         goto skip;
2155                 if (khugepaged_scan.address > hend)
2156                         goto skip;
2157                 if (khugepaged_scan.address < hstart)
2158                         khugepaged_scan.address = hstart;
2159                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2160
2161                 while (khugepaged_scan.address < hend) {
2162                         int ret;
2163                         cond_resched();
2164                         if (unlikely(khugepaged_test_exit(mm)))
2165                                 goto breakouterloop;
2166
2167                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2168                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2169                                   hend);
2170                         ret = khugepaged_scan_pmd(mm, vma,
2171                                                   khugepaged_scan.address,
2172                                                   hpage);
2173                         /* move to next address */
2174                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2175                         progress += HPAGE_PMD_NR;
2176                         if (ret)
2177                                 /* we released mmap_sem so break loop */
2178                                 goto breakouterloop_mmap_sem;
2179                         if (progress >= pages)
2180                                 goto breakouterloop;
2181                 }
2182         }
2183 breakouterloop:
2184         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2185 breakouterloop_mmap_sem:
2186
2187         spin_lock(&khugepaged_mm_lock);
2188         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2189         /*
2190          * Release the current mm_slot if this mm is about to die, or
2191          * if we scanned all vmas of this mm.
2192          */
2193         if (khugepaged_test_exit(mm) || !vma) {
2194                 /*
2195                  * Make sure that if mm_users is reaching zero while
2196                  * khugepaged runs here, khugepaged_exit will find
2197                  * mm_slot not pointing to the exiting mm.
2198                  */
2199                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2200                         khugepaged_scan.mm_slot = list_entry(
2201                                 mm_slot->mm_node.next,
2202                                 struct mm_slot, mm_node);
2203                         khugepaged_scan.address = 0;
2204                 } else {
2205                         khugepaged_scan.mm_slot = NULL;
2206                         khugepaged_full_scans++;
2207                 }
2208
2209                 collect_mm_slot(mm_slot);
2210         }
2211
2212         return progress;
2213 }
2214
2215 static int khugepaged_has_work(void)
2216 {
2217         return !list_empty(&khugepaged_scan.mm_head) &&
2218                 khugepaged_enabled();
2219 }
2220
2221 static int khugepaged_wait_event(void)
2222 {
2223         return !list_empty(&khugepaged_scan.mm_head) ||
2224                 !khugepaged_enabled();
2225 }
2226
2227 static void khugepaged_do_scan(struct page **hpage)
2228 {
2229         unsigned int progress = 0, pass_through_head = 0;
2230         unsigned int pages = khugepaged_pages_to_scan;
2231
2232         barrier(); /* write khugepaged_pages_to_scan to local stack */
2233
2234         while (progress < pages) {
2235                 cond_resched();
2236
2237 #ifndef CONFIG_NUMA
2238                 if (!*hpage) {
2239                         *hpage = alloc_hugepage(khugepaged_defrag());
2240                         if (unlikely(!*hpage)) {
2241                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2242                                 break;
2243                         }
2244                         count_vm_event(THP_COLLAPSE_ALLOC);
2245                 }
2246 #else
2247                 if (IS_ERR(*hpage))
2248                         break;
2249 #endif
2250
2251                 if (unlikely(kthread_should_stop() || freezing(current)))
2252                         break;
2253
2254                 spin_lock(&khugepaged_mm_lock);
2255                 if (!khugepaged_scan.mm_slot)
2256                         pass_through_head++;
2257                 if (khugepaged_has_work() &&
2258                     pass_through_head < 2)
2259                         progress += khugepaged_scan_mm_slot(pages - progress,
2260                                                             hpage);
2261                 else
2262                         progress = pages;
2263                 spin_unlock(&khugepaged_mm_lock);
2264         }
2265 }
2266
2267 static void khugepaged_alloc_sleep(void)
2268 {
2269         wait_event_freezable_timeout(khugepaged_wait, false,
2270                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2271 }
2272
2273 #ifndef CONFIG_NUMA
2274 static struct page *khugepaged_alloc_hugepage(void)
2275 {
2276         struct page *hpage;
2277
2278         do {
2279                 hpage = alloc_hugepage(khugepaged_defrag());
2280                 if (!hpage) {
2281                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2282                         khugepaged_alloc_sleep();
2283                 } else
2284                         count_vm_event(THP_COLLAPSE_ALLOC);
2285         } while (unlikely(!hpage) &&
2286                  likely(khugepaged_enabled()));
2287         return hpage;
2288 }
2289 #endif
2290
2291 static void khugepaged_loop(void)
2292 {
2293         struct page *hpage;
2294
2295 #ifdef CONFIG_NUMA
2296         hpage = NULL;
2297 #endif
2298         while (likely(khugepaged_enabled())) {
2299 #ifndef CONFIG_NUMA
2300                 hpage = khugepaged_alloc_hugepage();
2301                 if (unlikely(!hpage))
2302                         break;
2303 #else
2304                 if (IS_ERR(hpage)) {
2305                         khugepaged_alloc_sleep();
2306                         hpage = NULL;
2307                 }
2308 #endif
2309
2310                 khugepaged_do_scan(&hpage);
2311 #ifndef CONFIG_NUMA
2312                 if (hpage)
2313                         put_page(hpage);
2314 #endif
2315                 try_to_freeze();
2316                 if (unlikely(kthread_should_stop()))
2317                         break;
2318                 if (khugepaged_has_work()) {
2319                         if (!khugepaged_scan_sleep_millisecs)
2320                                 continue;
2321                         wait_event_freezable_timeout(khugepaged_wait, false,
2322                             msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2323                 } else if (khugepaged_enabled())
2324                         wait_event_freezable(khugepaged_wait,
2325                                              khugepaged_wait_event());
2326         }
2327 }
2328
2329 static int khugepaged(void *none)
2330 {
2331         struct mm_slot *mm_slot;
2332
2333         set_freezable();
2334         set_user_nice(current, 19);
2335
2336         /* serialize with start_khugepaged() */
2337         mutex_lock(&khugepaged_mutex);
2338
2339         for (;;) {
2340                 mutex_unlock(&khugepaged_mutex);
2341                 VM_BUG_ON(khugepaged_thread != current);
2342                 khugepaged_loop();
2343                 VM_BUG_ON(khugepaged_thread != current);
2344
2345                 mutex_lock(&khugepaged_mutex);
2346                 if (!khugepaged_enabled())
2347                         break;
2348                 if (unlikely(kthread_should_stop()))
2349                         break;
2350         }
2351
2352         spin_lock(&khugepaged_mm_lock);
2353         mm_slot = khugepaged_scan.mm_slot;
2354         khugepaged_scan.mm_slot = NULL;
2355         if (mm_slot)
2356                 collect_mm_slot(mm_slot);
2357         spin_unlock(&khugepaged_mm_lock);
2358
2359         khugepaged_thread = NULL;
2360         mutex_unlock(&khugepaged_mutex);
2361
2362         return 0;
2363 }
2364
2365 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2366 {
2367         struct page *page;
2368
2369         spin_lock(&mm->page_table_lock);
2370         if (unlikely(!pmd_trans_huge(*pmd))) {
2371                 spin_unlock(&mm->page_table_lock);
2372                 return;
2373         }
2374         page = pmd_page(*pmd);
2375         VM_BUG_ON(!page_count(page));
2376         get_page(page);
2377         spin_unlock(&mm->page_table_lock);
2378
2379         split_huge_page(page);
2380
2381         put_page(page);
2382         BUG_ON(pmd_trans_huge(*pmd));
2383 }
2384
2385 static void split_huge_page_address(struct mm_struct *mm,
2386                                     unsigned long address)
2387 {
2388         pgd_t *pgd;
2389         pud_t *pud;
2390         pmd_t *pmd;
2391
2392         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2393
2394         pgd = pgd_offset(mm, address);
2395         if (!pgd_present(*pgd))
2396                 return;
2397
2398         pud = pud_offset(pgd, address);
2399         if (!pud_present(*pud))
2400                 return;
2401
2402         pmd = pmd_offset(pud, address);
2403         if (!pmd_present(*pmd))
2404                 return;
2405         /*
2406          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2407          * materialize from under us.
2408          */
2409         split_huge_page_pmd(mm, pmd);
2410 }
2411
2412 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2413                              unsigned long start,
2414                              unsigned long end,
2415                              long adjust_next)
2416 {
2417         /*
2418          * If the new start address isn't hpage aligned and it could
2419          * previously contain an hugepage: check if we need to split
2420          * an huge pmd.
2421          */
2422         if (start & ~HPAGE_PMD_MASK &&
2423             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2424             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2425                 split_huge_page_address(vma->vm_mm, start);
2426
2427         /*
2428          * If the new end address isn't hpage aligned and it could
2429          * previously contain an hugepage: check if we need to split
2430          * an huge pmd.
2431          */
2432         if (end & ~HPAGE_PMD_MASK &&
2433             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2434             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2435                 split_huge_page_address(vma->vm_mm, end);
2436
2437         /*
2438          * If we're also updating the vma->vm_next->vm_start, if the new
2439          * vm_next->vm_start isn't page aligned and it could previously
2440          * contain an hugepage: check if we need to split an huge pmd.
2441          */
2442         if (adjust_next > 0) {
2443                 struct vm_area_struct *next = vma->vm_next;
2444                 unsigned long nstart = next->vm_start;
2445                 nstart += adjust_next << PAGE_SHIFT;
2446                 if (nstart & ~HPAGE_PMD_MASK &&
2447                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2448                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2449                         split_huge_page_address(next->vm_mm, nstart);
2450         }
2451 }