mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/mm_inline.h>
  16 #include <linux/kthread.h>
  17 #include <linux/khugepaged.h>
  18 #include <linux/freezer.h>
  19 #include <linux/mman.h>
  20 #include <linux/pagemap.h>
  21 #include <asm/tlb.h>
  22 #include <asm/pgalloc.h>
  23 #include "internal.h"
  24
  25 /*
  26  * By default transparent hugepage support is enabled for all mappings
  27  * and khugepaged scans all mappings. Defrag is only invoked by
  28  * khugepaged hugepage allocations and by page faults inside
  29  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  30  * allocations.
  31  */
  32 unsigned long transparent_hugepage_flags __read_mostly =
  33 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  34         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  35 #endif
  36 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  37         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  38 #endif
  39         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  40         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
  41
  42 /* default scan 8*512 pte (or vmas) every 30 second */
  43 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  44 static unsigned int khugepaged_pages_collapsed;
  45 static unsigned int khugepaged_full_scans;
  46 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  47 /* during fragmentation poll the hugepage allocator once every minute */
  48 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  49 static struct task_struct *khugepaged_thread __read_mostly;
  50 static DEFINE_MUTEX(khugepaged_mutex);
  51 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  52 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  53 /*
  54  * default collapse hugepages if there is at least one pte mapped like
  55  * it would have happened if the vma was large enough during page
  56  * fault.
  57  */
  58 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  59
  60 static int khugepaged(void *none);
  61 static int mm_slots_hash_init(void);
  62 static int khugepaged_slab_init(void);
  63 static void khugepaged_slab_free(void);
  64
  65 #define MM_SLOTS_HASH_HEADS 1024
  66 static struct hlist_head *mm_slots_hash __read_mostly;
  67 static struct kmem_cache *mm_slot_cache __read_mostly;
  68
  69 /**
  70  * struct mm_slot - hash lookup from mm to mm_slot
  71  * @hash: hash collision list
  72  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  73  * @mm: the mm that this information is valid for
  74  */
  75 struct mm_slot {
  76         struct hlist_node hash;
  77         struct list_head mm_node;
  78         struct mm_struct *mm;
  79 };
  80
  81 /**
  82  * struct khugepaged_scan - cursor for scanning
  83  * @mm_head: the head of the mm list to scan
  84  * @mm_slot: the current mm_slot we are scanning
  85  * @address: the next address inside that to be scanned
  86  *
  87  * There is only the one khugepaged_scan instance of this cursor structure.
  88  */
  89 struct khugepaged_scan {
  90         struct list_head mm_head;
  91         struct mm_slot *mm_slot;
  92         unsigned long address;
  93 };
  94 static struct khugepaged_scan khugepaged_scan = {
  95         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
  96 };
  97
  98
  99 static int set_recommended_min_free_kbytes(void)
 100 {
 101         struct zone *zone;
 102         int nr_zones = 0;
 103         unsigned long recommended_min;
 104         extern int min_free_kbytes;
 105
 106         if (!khugepaged_enabled())
 107                 return 0;
 108
 109         for_each_populated_zone(zone)
 110                 nr_zones++;
 111
 112         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 113         recommended_min = pageblock_nr_pages * nr_zones * 2;
 114
 115         /*
 116          * Make sure that on average at least two pageblocks are almost free
 117          * of another type, one for a migratetype to fall back to and a
 118          * second to avoid subsequent fallbacks of other types There are 3
 119          * MIGRATE_TYPES we care about.
 120          */
 121         recommended_min += pageblock_nr_pages * nr_zones *
 122                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 123
 124         /* don't ever allow to reserve more than 5% of the lowmem */
 125         recommended_min = min(recommended_min,
 126                               (unsigned long) nr_free_buffer_pages() / 20);
 127         recommended_min <<= (PAGE_SHIFT-10);
 128
 129         if (recommended_min > min_free_kbytes)
 130                 min_free_kbytes = recommended_min;
 131         setup_per_zone_wmarks();
 132         return 0;
 133 }
 134 late_initcall(set_recommended_min_free_kbytes);
 135
 136 static int start_khugepaged(void)
 137 {
 138         int err = 0;
 139         if (khugepaged_enabled()) {
 140                 if (!khugepaged_thread)
 141                         khugepaged_thread = kthread_run(khugepaged, NULL,
 142                                                         "khugepaged");
 143                 if (unlikely(IS_ERR(khugepaged_thread))) {
 144                         printk(KERN_ERR
 145                                "khugepaged: kthread_run(khugepaged) failed\n");
 146                         err = PTR_ERR(khugepaged_thread);
 147                         khugepaged_thread = NULL;
 148                 }
 149
 150                 if (!list_empty(&khugepaged_scan.mm_head))
 151                         wake_up_interruptible(&khugepaged_wait);
 152
 153                 set_recommended_min_free_kbytes();
 154         } else if (khugepaged_thread) {
 155                 kthread_stop(khugepaged_thread);
 156                 khugepaged_thread = NULL;
 157         }
 158
 159         return err;
 160 }
 161
 162 #ifdef CONFIG_SYSFS
 163
 164 static ssize_t double_flag_show(struct kobject *kobj,
 165                                 struct kobj_attribute *attr, char *buf,
 166                                 enum transparent_hugepage_flag enabled,
 167                                 enum transparent_hugepage_flag req_madv)
 168 {
 169         if (test_bit(enabled, &transparent_hugepage_flags)) {
 170                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 171                 return sprintf(buf, "[always] madvise never\n");
 172         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 173                 return sprintf(buf, "always [madvise] never\n");
 174         else
 175                 return sprintf(buf, "always madvise [never]\n");
 176 }
 177 static ssize_t double_flag_store(struct kobject *kobj,
 178                                  struct kobj_attribute *attr,
 179                                  const char *buf, size_t count,
 180                                  enum transparent_hugepage_flag enabled,
 181                                  enum transparent_hugepage_flag req_madv)
 182 {
 183         if (!memcmp("always", buf,
 184                     min(sizeof("always")-1, count))) {
 185                 set_bit(enabled, &transparent_hugepage_flags);
 186                 clear_bit(req_madv, &transparent_hugepage_flags);
 187         } else if (!memcmp("madvise", buf,
 188                            min(sizeof("madvise")-1, count))) {
 189                 clear_bit(enabled, &transparent_hugepage_flags);
 190                 set_bit(req_madv, &transparent_hugepage_flags);
 191         } else if (!memcmp("never", buf,
 192                            min(sizeof("never")-1, count))) {
 193                 clear_bit(enabled, &transparent_hugepage_flags);
 194                 clear_bit(req_madv, &transparent_hugepage_flags);
 195         } else
 196                 return -EINVAL;
 197
 198         return count;
 199 }
 200
 201 static ssize_t enabled_show(struct kobject *kobj,
 202                             struct kobj_attribute *attr, char *buf)
 203 {
 204         return double_flag_show(kobj, attr, buf,
 205                                 TRANSPARENT_HUGEPAGE_FLAG,
 206                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 207 }
 208 static ssize_t enabled_store(struct kobject *kobj,
 209                              struct kobj_attribute *attr,
 210                              const char *buf, size_t count)
 211 {
 212         ssize_t ret;
 213
 214         ret = double_flag_store(kobj, attr, buf, count,
 215                                 TRANSPARENT_HUGEPAGE_FLAG,
 216                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 217
 218         if (ret > 0) {
 219                 int err;
 220
 221                 mutex_lock(&khugepaged_mutex);
 222                 err = start_khugepaged();
 223                 mutex_unlock(&khugepaged_mutex);
 224
 225                 if (err)
 226                         ret = err;
 227         }
 228
 229         return ret;
 230 }
 231 static struct kobj_attribute enabled_attr =
 232         __ATTR(enabled, 0644, enabled_show, enabled_store);
 233
 234 static ssize_t single_flag_show(struct kobject *kobj,
 235                                 struct kobj_attribute *attr, char *buf,
 236                                 enum transparent_hugepage_flag flag)
 237 {
 238         return sprintf(buf, "%d\n",
 239                        !!test_bit(flag, &transparent_hugepage_flags));
 240 }
 241
 242 static ssize_t single_flag_store(struct kobject *kobj,
 243                                  struct kobj_attribute *attr,
 244                                  const char *buf, size_t count,
 245                                  enum transparent_hugepage_flag flag)
 246 {
 247         unsigned long value;
 248         int ret;
 249
 250         ret = kstrtoul(buf, 10, &value);
 251         if (ret < 0)
 252                 return ret;
 253         if (value > 1)
 254                 return -EINVAL;
 255
 256         if (value)
 257                 set_bit(flag, &transparent_hugepage_flags);
 258         else
 259                 clear_bit(flag, &transparent_hugepage_flags);
 260
 261         return count;
 262 }
 263
 264 /*
 265  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 266  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 267  * memory just to allocate one more hugepage.
 268  */
 269 static ssize_t defrag_show(struct kobject *kobj,
 270                            struct kobj_attribute *attr, char *buf)
 271 {
 272         return double_flag_show(kobj, attr, buf,
 273                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 274                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 275 }
 276 static ssize_t defrag_store(struct kobject *kobj,
 277                             struct kobj_attribute *attr,
 278                             const char *buf, size_t count)
 279 {
 280         return double_flag_store(kobj, attr, buf, count,
 281                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 282                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 283 }
 284 static struct kobj_attribute defrag_attr =
 285         __ATTR(defrag, 0644, defrag_show, defrag_store);
 286
 287 #ifdef CONFIG_DEBUG_VM
 288 static ssize_t debug_cow_show(struct kobject *kobj,
 289                                 struct kobj_attribute *attr, char *buf)
 290 {
 291         return single_flag_show(kobj, attr, buf,
 292                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 293 }
 294 static ssize_t debug_cow_store(struct kobject *kobj,
 295                                struct kobj_attribute *attr,
 296                                const char *buf, size_t count)
 297 {
 298         return single_flag_store(kobj, attr, buf, count,
 299                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 300 }
 301 static struct kobj_attribute debug_cow_attr =
 302         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 303 #endif /* CONFIG_DEBUG_VM */
 304
 305 static struct attribute *hugepage_attr[] = {
 306         &enabled_attr.attr,
 307         &defrag_attr.attr,
 308 #ifdef CONFIG_DEBUG_VM
 309         &debug_cow_attr.attr,
 310 #endif
 311         NULL,
 312 };
 313
 314 static struct attribute_group hugepage_attr_group = {
 315         .attrs = hugepage_attr,
 316 };
 317
 318 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 319                                          struct kobj_attribute *attr,
 320                                          char *buf)
 321 {
 322         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 323 }
 324
 325 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 326                                           struct kobj_attribute *attr,
 327                                           const char *buf, size_t count)
 328 {
 329         unsigned long msecs;
 330         int err;
 331
 332         err = strict_strtoul(buf, 10, &msecs);
 333         if (err || msecs > UINT_MAX)
 334                 return -EINVAL;
 335
 336         khugepaged_scan_sleep_millisecs = msecs;
 337         wake_up_interruptible(&khugepaged_wait);
 338
 339         return count;
 340 }
 341 static struct kobj_attribute scan_sleep_millisecs_attr =
 342         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 343                scan_sleep_millisecs_store);
 344
 345 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 346                                           struct kobj_attribute *attr,
 347                                           char *buf)
 348 {
 349         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 350 }
 351
 352 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 353                                            struct kobj_attribute *attr,
 354                                            const char *buf, size_t count)
 355 {
 356         unsigned long msecs;
 357         int err;
 358
 359         err = strict_strtoul(buf, 10, &msecs);
 360         if (err || msecs > UINT_MAX)
 361                 return -EINVAL;
 362
 363         khugepaged_alloc_sleep_millisecs = msecs;
 364         wake_up_interruptible(&khugepaged_wait);
 365
 366         return count;
 367 }
 368 static struct kobj_attribute alloc_sleep_millisecs_attr =
 369         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 370                alloc_sleep_millisecs_store);
 371
 372 static ssize_t pages_to_scan_show(struct kobject *kobj,
 373                                   struct kobj_attribute *attr,
 374                                   char *buf)
 375 {
 376         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 377 }
 378 static ssize_t pages_to_scan_store(struct kobject *kobj,
 379                                    struct kobj_attribute *attr,
 380                                    const char *buf, size_t count)
 381 {
 382         int err;
 383         unsigned long pages;
 384
 385         err = strict_strtoul(buf, 10, &pages);
 386         if (err || !pages || pages > UINT_MAX)
 387                 return -EINVAL;
 388
 389         khugepaged_pages_to_scan = pages;
 390
 391         return count;
 392 }
 393 static struct kobj_attribute pages_to_scan_attr =
 394         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 395                pages_to_scan_store);
 396
 397 static ssize_t pages_collapsed_show(struct kobject *kobj,
 398                                     struct kobj_attribute *attr,
 399                                     char *buf)
 400 {
 401         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 402 }
 403 static struct kobj_attribute pages_collapsed_attr =
 404         __ATTR_RO(pages_collapsed);
 405
 406 static ssize_t full_scans_show(struct kobject *kobj,
 407                                struct kobj_attribute *attr,
 408                                char *buf)
 409 {
 410         return sprintf(buf, "%u\n", khugepaged_full_scans);
 411 }
 412 static struct kobj_attribute full_scans_attr =
 413         __ATTR_RO(full_scans);
 414
 415 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 416                                       struct kobj_attribute *attr, char *buf)
 417 {
 418         return single_flag_show(kobj, attr, buf,
 419                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 420 }
 421 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 422                                        struct kobj_attribute *attr,
 423                                        const char *buf, size_t count)
 424 {
 425         return single_flag_store(kobj, attr, buf, count,
 426                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 427 }
 428 static struct kobj_attribute khugepaged_defrag_attr =
 429         __ATTR(defrag, 0644, khugepaged_defrag_show,
 430                khugepaged_defrag_store);
 431
 432 /*
 433  * max_ptes_none controls if khugepaged should collapse hugepages over
 434  * any unmapped ptes in turn potentially increasing the memory
 435  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 436  * reduce the available free memory in the system as it
 437  * runs. Increasing max_ptes_none will instead potentially reduce the
 438  * free memory in the system during the khugepaged scan.
 439  */
 440 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 441                                              struct kobj_attribute *attr,
 442                                              char *buf)
 443 {
 444         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 445 }
 446 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 447                                               struct kobj_attribute *attr,
 448                                               const char *buf, size_t count)
 449 {
 450         int err;
 451         unsigned long max_ptes_none;
 452
 453         err = strict_strtoul(buf, 10, &max_ptes_none);
 454         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 455                 return -EINVAL;
 456
 457         khugepaged_max_ptes_none = max_ptes_none;
 458
 459         return count;
 460 }
 461 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 462         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 463                khugepaged_max_ptes_none_store);
 464
 465 static struct attribute *khugepaged_attr[] = {
 466         &khugepaged_defrag_attr.attr,
 467         &khugepaged_max_ptes_none_attr.attr,
 468         &pages_to_scan_attr.attr,
 469         &pages_collapsed_attr.attr,
 470         &full_scans_attr.attr,
 471         &scan_sleep_millisecs_attr.attr,
 472         &alloc_sleep_millisecs_attr.attr,
 473         NULL,
 474 };
 475
 476 static struct attribute_group khugepaged_attr_group = {
 477         .attrs = khugepaged_attr,
 478         .name = "khugepaged",
 479 };
 480
 481 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 482 {
 483         int err;
 484
 485         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 486         if (unlikely(!*hugepage_kobj)) {
 487                 printk(KERN_ERR "hugepage: failed kobject create\n");
 488                 return -ENOMEM;
 489         }
 490
 491         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 492         if (err) {
 493                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 494                 goto delete_obj;
 495         }
 496
 497         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 498         if (err) {
 499                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 500                 goto remove_hp_group;
 501         }
 502
 503         return 0;
 504
 505 remove_hp_group:
 506         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 507 delete_obj:
 508         kobject_put(*hugepage_kobj);
 509         return err;
 510 }
 511
 512 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 513 {
 514         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 515         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 516         kobject_put(hugepage_kobj);
 517 }
 518 #else
 519 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 520 {
 521         return 0;
 522 }
 523
 524 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 525 {
 526 }
 527 #endif /* CONFIG_SYSFS */
 528
 529 static int __init hugepage_init(void)
 530 {
 531         int err;
 532         struct kobject *hugepage_kobj;
 533
 534         if (!has_transparent_hugepage()) {
 535                 transparent_hugepage_flags = 0;
 536                 return -EINVAL;
 537         }
 538
 539         err = hugepage_init_sysfs(&hugepage_kobj);
 540         if (err)
 541                 return err;
 542
 543         err = khugepaged_slab_init();
 544         if (err)
 545                 goto out;
 546
 547         err = mm_slots_hash_init();
 548         if (err) {
 549                 khugepaged_slab_free();
 550                 goto out;
 551         }
 552
 553         /*
 554          * By default disable transparent hugepages on smaller systems,
 555          * where the extra memory used could hurt more than TLB overhead
 556          * is likely to save.  The admin can still enable it through /sys.
 557          */
 558         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 559                 transparent_hugepage_flags = 0;
 560
 561         start_khugepaged();
 562
 563         return 0;
 564 out:
 565         hugepage_exit_sysfs(hugepage_kobj);
 566         return err;
 567 }
 568 module_init(hugepage_init)
 569
 570 static int __init setup_transparent_hugepage(char *str)
 571 {
 572         int ret = 0;
 573         if (!str)
 574                 goto out;
 575         if (!strcmp(str, "always")) {
 576                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 577                         &transparent_hugepage_flags);
 578                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 579                           &transparent_hugepage_flags);
 580                 ret = 1;
 581         } else if (!strcmp(str, "madvise")) {
 582                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 583                           &transparent_hugepage_flags);
 584                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 585                         &transparent_hugepage_flags);
 586                 ret = 1;
 587         } else if (!strcmp(str, "never")) {
 588                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 589                           &transparent_hugepage_flags);
 590                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 591                           &transparent_hugepage_flags);
 592                 ret = 1;
 593         }
 594 out:
 595         if (!ret)
 596                 printk(KERN_WARNING
 597                        "transparent_hugepage= cannot parse, ignored\n");
 598         return ret;
 599 }
 600 __setup("transparent_hugepage=", setup_transparent_hugepage);
 601
 602 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 603 {
 604         if (likely(vma->vm_flags & VM_WRITE))
 605                 pmd = pmd_mkwrite(pmd);
 606         return pmd;
 607 }
 608
 609 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 610                                         struct vm_area_struct *vma,
 611                                         unsigned long haddr, pmd_t *pmd,
 612                                         struct page *page)
 613 {
 614         pgtable_t pgtable;
 615
 616         VM_BUG_ON(!PageCompound(page));
 617         pgtable = pte_alloc_one(mm, haddr);
 618         if (unlikely(!pgtable))
 619                 return VM_FAULT_OOM;
 620
 621         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 622         __SetPageUptodate(page);
 623
 624         spin_lock(&mm->page_table_lock);
 625         if (unlikely(!pmd_none(*pmd))) {
 626                 spin_unlock(&mm->page_table_lock);
 627                 mem_cgroup_uncharge_page(page);
 628                 put_page(page);
 629                 pte_free(mm, pgtable);
 630         } else {
 631                 pmd_t entry;
 632                 entry = mk_pmd(page, vma->vm_page_prot);
 633                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 634                 entry = pmd_mkhuge(entry);
 635                 /*
 636                  * The spinlocking to take the lru_lock inside
 637                  * page_add_new_anon_rmap() acts as a full memory
 638                  * barrier to be sure clear_huge_page writes become
 639                  * visible after the set_pmd_at() write.
 640                  */
 641                 page_add_new_anon_rmap(page, vma, haddr);
 642                 set_pmd_at(mm, haddr, pmd, entry);
 643                 pgtable_trans_huge_deposit(mm, pgtable);
 644                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 645                 mm->nr_ptes++;
 646                 spin_unlock(&mm->page_table_lock);
 647         }
 648
 649         return 0;
 650 }
 651
 652 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 653 {
 654         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 655 }
 656
 657 static inline struct page *alloc_hugepage_vma(int defrag,
 658                                               struct vm_area_struct *vma,
 659                                               unsigned long haddr, int nd,
 660                                               gfp_t extra_gfp)
 661 {
 662         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 663                                HPAGE_PMD_ORDER, vma, haddr, nd);
 664 }
 665
 666 #ifndef CONFIG_NUMA
 667 static inline struct page *alloc_hugepage(int defrag)
 668 {
 669         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 670                            HPAGE_PMD_ORDER);
 671 }
 672 #endif
 673
 674 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 675                                unsigned long address, pmd_t *pmd,
 676                                unsigned int flags)
 677 {
 678         struct page *page;
 679         unsigned long haddr = address & HPAGE_PMD_MASK;
 680         pte_t *pte;
 681
 682         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 683                 if (unlikely(anon_vma_prepare(vma)))
 684                         return VM_FAULT_OOM;
 685                 if (unlikely(khugepaged_enter(vma)))
 686                         return VM_FAULT_OOM;
 687                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 688                                           vma, haddr, numa_node_id(), 0);
 689                 if (unlikely(!page)) {
 690                         count_vm_event(THP_FAULT_FALLBACK);
 691                         goto out;
 692                 }
 693                 count_vm_event(THP_FAULT_ALLOC);
 694                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 695                         put_page(page);
 696                         goto out;
 697                 }
 698                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 699                                                           page))) {
 700                         mem_cgroup_uncharge_page(page);
 701                         put_page(page);
 702                         goto out;
 703                 }
 704
 705                 return 0;
 706         }
 707 out:
 708         /*
 709          * Use __pte_alloc instead of pte_alloc_map, because we can't
 710          * run pte_offset_map on the pmd, if an huge pmd could
 711          * materialize from under us from a different thread.
 712          */
 713         if (unlikely(pmd_none(*pmd)) &&
 714             unlikely(__pte_alloc(mm, vma, pmd, address)))
 715                 return VM_FAULT_OOM;
 716         /* if an huge pmd materialized from under us just retry later */
 717         if (unlikely(pmd_trans_huge(*pmd)))
 718                 return 0;
 719         /*
 720          * A regular pmd is established and it can't morph into a huge pmd
 721          * from under us anymore at this point because we hold the mmap_sem
 722          * read mode and khugepaged takes it in write mode. So now it's
 723          * safe to run pte_offset_map().
 724          */
 725         pte = pte_offset_map(pmd, address);
 726         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 727 }
 728
 729 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 730                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 731                   struct vm_area_struct *vma)
 732 {
 733         struct page *src_page;
 734         pmd_t pmd;
 735         pgtable_t pgtable;
 736         int ret;
 737
 738         ret = -ENOMEM;
 739         pgtable = pte_alloc_one(dst_mm, addr);
 740         if (unlikely(!pgtable))
 741                 goto out;
 742
 743         spin_lock(&dst_mm->page_table_lock);
 744         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 745
 746         ret = -EAGAIN;
 747         pmd = *src_pmd;
 748         if (unlikely(!pmd_trans_huge(pmd))) {
 749                 pte_free(dst_mm, pgtable);
 750                 goto out_unlock;
 751         }
 752         if (unlikely(pmd_trans_splitting(pmd))) {
 753                 /* split huge page running from under us */
 754                 spin_unlock(&src_mm->page_table_lock);
 755                 spin_unlock(&dst_mm->page_table_lock);
 756                 pte_free(dst_mm, pgtable);
 757
 758                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 759                 goto out;
 760         }
 761         src_page = pmd_page(pmd);
 762         VM_BUG_ON(!PageHead(src_page));
 763         get_page(src_page);
 764         page_dup_rmap(src_page);
 765         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 766
 767         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 768         pmd = pmd_mkold(pmd_wrprotect(pmd));
 769         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 770         pgtable_trans_huge_deposit(dst_mm, pgtable);
 771         dst_mm->nr_ptes++;
 772
 773         ret = 0;
 774 out_unlock:
 775         spin_unlock(&src_mm->page_table_lock);
 776         spin_unlock(&dst_mm->page_table_lock);
 777 out:
 778         return ret;
 779 }
 780
 781 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 782                                         struct vm_area_struct *vma,
 783                                         unsigned long address,
 784                                         pmd_t *pmd, pmd_t orig_pmd,
 785                                         struct page *page,
 786                                         unsigned long haddr)
 787 {
 788         pgtable_t pgtable;
 789         pmd_t _pmd;
 790         int ret = 0, i;
 791         struct page **pages;
 792         unsigned long mmun_start;       /* For mmu_notifiers */
 793         unsigned long mmun_end;         /* For mmu_notifiers */
 794
 795         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 796                         GFP_KERNEL);
 797         if (unlikely(!pages)) {
 798                 ret |= VM_FAULT_OOM;
 799                 goto out;
 800         }
 801
 802         for (i = 0; i < HPAGE_PMD_NR; i++) {
 803                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 804                                                __GFP_OTHER_NODE,
 805                                                vma, address, page_to_nid(page));
 806                 if (unlikely(!pages[i] ||
 807                              mem_cgroup_newpage_charge(pages[i], mm,
 808                                                        GFP_KERNEL))) {
 809                         if (pages[i])
 810                                 put_page(pages[i]);
 811                         mem_cgroup_uncharge_start();
 812                         while (--i >= 0) {
 813                                 mem_cgroup_uncharge_page(pages[i]);
 814                                 put_page(pages[i]);
 815                         }
 816                         mem_cgroup_uncharge_end();
 817                         kfree(pages);
 818                         ret |= VM_FAULT_OOM;
 819                         goto out;
 820                 }
 821         }
 822
 823         for (i = 0; i < HPAGE_PMD_NR; i++) {
 824                 copy_user_highpage(pages[i], page + i,
 825                                    haddr + PAGE_SIZE * i, vma);
 826                 __SetPageUptodate(pages[i]);
 827                 cond_resched();
 828         }
 829
 830         mmun_start = haddr;
 831         mmun_end   = haddr + HPAGE_PMD_SIZE;
 832         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 833
 834         spin_lock(&mm->page_table_lock);
 835         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 836                 goto out_free_pages;
 837         VM_BUG_ON(!PageHead(page));
 838
 839         pmdp_clear_flush(vma, haddr, pmd);
 840         /* leave pmd empty until pte is filled */
 841
 842         pgtable = pgtable_trans_huge_withdraw(mm);
 843         pmd_populate(mm, &_pmd, pgtable);
 844
 845         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 846                 pte_t *pte, entry;
 847                 entry = mk_pte(pages[i], vma->vm_page_prot);
 848                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 849                 page_add_new_anon_rmap(pages[i], vma, haddr);
 850                 pte = pte_offset_map(&_pmd, haddr);
 851                 VM_BUG_ON(!pte_none(*pte));
 852                 set_pte_at(mm, haddr, pte, entry);
 853                 pte_unmap(pte);
 854         }
 855         kfree(pages);
 856
 857         smp_wmb(); /* make pte visible before pmd */
 858         pmd_populate(mm, pmd, pgtable);
 859         page_remove_rmap(page);
 860         spin_unlock(&mm->page_table_lock);
 861
 862         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 863
 864         ret |= VM_FAULT_WRITE;
 865         put_page(page);
 866
 867 out:
 868         return ret;
 869
 870 out_free_pages:
 871         spin_unlock(&mm->page_table_lock);
 872         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 873         mem_cgroup_uncharge_start();
 874         for (i = 0; i < HPAGE_PMD_NR; i++) {
 875                 mem_cgroup_uncharge_page(pages[i]);
 876                 put_page(pages[i]);
 877         }
 878         mem_cgroup_uncharge_end();
 879         kfree(pages);
 880         goto out;
 881 }
 882
 883 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 884                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 885 {
 886         int ret = 0;
 887         struct page *page, *new_page;
 888         unsigned long haddr;
 889         unsigned long mmun_start;       /* For mmu_notifiers */
 890         unsigned long mmun_end;         /* For mmu_notifiers */
 891
 892         VM_BUG_ON(!vma->anon_vma);
 893         spin_lock(&mm->page_table_lock);
 894         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 895                 goto out_unlock;
 896
 897         page = pmd_page(orig_pmd);
 898         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 899         haddr = address & HPAGE_PMD_MASK;
 900         if (page_mapcount(page) == 1) {
 901                 pmd_t entry;
 902                 entry = pmd_mkyoung(orig_pmd);
 903                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 904                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 905                         update_mmu_cache_pmd(vma, address, pmd);
 906                 ret |= VM_FAULT_WRITE;
 907                 goto out_unlock;
 908         }
 909         get_page(page);
 910         spin_unlock(&mm->page_table_lock);
 911
 912         if (transparent_hugepage_enabled(vma) &&
 913             !transparent_hugepage_debug_cow())
 914                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 915                                               vma, haddr, numa_node_id(), 0);
 916         else
 917                 new_page = NULL;
 918
 919         if (unlikely(!new_page)) {
 920                 count_vm_event(THP_FAULT_FALLBACK);
 921                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 922                                                    pmd, orig_pmd, page, haddr);
 923                 if (ret & VM_FAULT_OOM)
 924                         split_huge_page(page);
 925                 put_page(page);
 926                 goto out;
 927         }
 928         count_vm_event(THP_FAULT_ALLOC);
 929
 930         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 931                 put_page(new_page);
 932                 split_huge_page(page);
 933                 put_page(page);
 934                 ret |= VM_FAULT_OOM;
 935                 goto out;
 936         }
 937
 938         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 939         __SetPageUptodate(new_page);
 940
 941         mmun_start = haddr;
 942         mmun_end   = haddr + HPAGE_PMD_SIZE;
 943         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 944
 945         spin_lock(&mm->page_table_lock);
 946         put_page(page);
 947         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 948                 spin_unlock(&mm->page_table_lock);
 949                 mem_cgroup_uncharge_page(new_page);
 950                 put_page(new_page);
 951                 goto out_mn;
 952         } else {
 953                 pmd_t entry;
 954                 VM_BUG_ON(!PageHead(page));
 955                 entry = mk_pmd(new_page, vma->vm_page_prot);
 956                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 957                 entry = pmd_mkhuge(entry);
 958                 pmdp_clear_flush(vma, haddr, pmd);
 959                 page_add_new_anon_rmap(new_page, vma, haddr);
 960                 set_pmd_at(mm, haddr, pmd, entry);
 961                 update_mmu_cache_pmd(vma, address, pmd);
 962                 page_remove_rmap(page);
 963                 put_page(page);
 964                 ret |= VM_FAULT_WRITE;
 965         }
 966         spin_unlock(&mm->page_table_lock);
 967 out_mn:
 968         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 969 out:
 970         return ret;
 971 out_unlock:
 972         spin_unlock(&mm->page_table_lock);
 973         return ret;
 974 }
 975
 976 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
 977                                    unsigned long addr,
 978                                    pmd_t *pmd,
 979                                    unsigned int flags)
 980 {
 981         struct mm_struct *mm = vma->vm_mm;
 982         struct page *page = NULL;
 983
 984         assert_spin_locked(&mm->page_table_lock);
 985
 986         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 987                 goto out;
 988
 989         page = pmd_page(*pmd);
 990         VM_BUG_ON(!PageHead(page));
 991         if (flags & FOLL_TOUCH) {
 992                 pmd_t _pmd;
 993                 /*
 994                  * We should set the dirty bit only for FOLL_WRITE but
 995                  * for now the dirty bit in the pmd is meaningless.
 996                  * And if the dirty bit will become meaningful and
 997                  * we'll only set it with FOLL_WRITE, an atomic
 998                  * set_bit will be required on the pmd to set the
 999                  * young bit, instead of the current set_pmd_at.
1000                  */
1001                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1002                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1003         }
1004         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1005                 if (page->mapping && trylock_page(page)) {
1006                         lru_add_drain();
1007                         if (page->mapping)
1008                                 mlock_vma_page(page);
1009                         unlock_page(page);
1010                 }
1011         }
1012         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1013         VM_BUG_ON(!PageCompound(page));
1014         if (flags & FOLL_GET)
1015                 get_page_foll(page);
1016
1017 out:
1018         return page;
1019 }
1020
1021 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1022                  pmd_t *pmd, unsigned long addr)
1023 {
1024         int ret = 0;
1025
1026         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1027                 struct page *page;
1028                 pgtable_t pgtable;
1029                 pmd_t orig_pmd;
1030                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1031                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1032                 page = pmd_page(orig_pmd);
1033                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1034                 page_remove_rmap(page);
1035                 VM_BUG_ON(page_mapcount(page) < 0);
1036                 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1037                 VM_BUG_ON(!PageHead(page));
1038                 tlb->mm->nr_ptes--;
1039                 spin_unlock(&tlb->mm->page_table_lock);
1040                 tlb_remove_page(tlb, page);
1041                 pte_free(tlb->mm, pgtable);
1042                 ret = 1;
1043         }
1044         return ret;
1045 }
1046
1047 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1048                 unsigned long addr, unsigned long end,
1049                 unsigned char *vec)
1050 {
1051         int ret = 0;
1052
1053         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1054                 /*
1055                  * All logical pages in the range are present
1056                  * if backed by a huge page.
1057                  */
1058                 spin_unlock(&vma->vm_mm->page_table_lock);
1059                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1060                 ret = 1;
1061         }
1062
1063         return ret;
1064 }
1065
1066 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1067                   unsigned long old_addr,
1068                   unsigned long new_addr, unsigned long old_end,
1069                   pmd_t *old_pmd, pmd_t *new_pmd)
1070 {
1071         int ret = 0;
1072         pmd_t pmd;
1073
1074         struct mm_struct *mm = vma->vm_mm;
1075
1076         if ((old_addr & ~HPAGE_PMD_MASK) ||
1077             (new_addr & ~HPAGE_PMD_MASK) ||
1078             old_end - old_addr < HPAGE_PMD_SIZE ||
1079             (new_vma->vm_flags & VM_NOHUGEPAGE))
1080                 goto out;
1081
1082         /*
1083          * The destination pmd shouldn't be established, free_pgtables()
1084          * should have release it.
1085          */
1086         if (WARN_ON(!pmd_none(*new_pmd))) {
1087                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1088                 goto out;
1089         }
1090
1091         ret = __pmd_trans_huge_lock(old_pmd, vma);
1092         if (ret == 1) {
1093                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1094                 VM_BUG_ON(!pmd_none(*new_pmd));
1095                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1096                 spin_unlock(&mm->page_table_lock);
1097         }
1098 out:
1099         return ret;
1100 }
1101
1102 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1103                 unsigned long addr, pgprot_t newprot)
1104 {
1105         struct mm_struct *mm = vma->vm_mm;
1106         int ret = 0;
1107
1108         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1109                 pmd_t entry;
1110                 entry = pmdp_get_and_clear(mm, addr, pmd);
1111                 entry = pmd_modify(entry, newprot);
1112                 set_pmd_at(mm, addr, pmd, entry);
1113                 spin_unlock(&vma->vm_mm->page_table_lock);
1114                 ret = 1;
1115         }
1116
1117         return ret;
1118 }
1119
1120 /*
1121  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1122  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1123  *
1124  * Note that if it returns 1, this routine returns without unlocking page
1125  * table locks. So callers must unlock them.
1126  */
1127 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1128 {
1129         spin_lock(&vma->vm_mm->page_table_lock);
1130         if (likely(pmd_trans_huge(*pmd))) {
1131                 if (unlikely(pmd_trans_splitting(*pmd))) {
1132                         spin_unlock(&vma->vm_mm->page_table_lock);
1133                         wait_split_huge_page(vma->anon_vma, pmd);
1134                         return -1;
1135                 } else {
1136                         /* Thp mapped by 'pmd' is stable, so we can
1137                          * handle it as it is. */
1138                         return 1;
1139                 }
1140         }
1141         spin_unlock(&vma->vm_mm->page_table_lock);
1142         return 0;
1143 }
1144
1145 pmd_t *page_check_address_pmd(struct page *page,
1146                               struct mm_struct *mm,
1147                               unsigned long address,
1148                               enum page_check_address_pmd_flag flag)
1149 {
1150         pgd_t *pgd;
1151         pud_t *pud;
1152         pmd_t *pmd, *ret = NULL;
1153
1154         if (address & ~HPAGE_PMD_MASK)
1155                 goto out;
1156
1157         pgd = pgd_offset(mm, address);
1158         if (!pgd_present(*pgd))
1159                 goto out;
1160
1161         pud = pud_offset(pgd, address);
1162         if (!pud_present(*pud))
1163                 goto out;
1164
1165         pmd = pmd_offset(pud, address);
1166         if (pmd_none(*pmd))
1167                 goto out;
1168         if (pmd_page(*pmd) != page)
1169                 goto out;
1170         /*
1171          * split_vma() may create temporary aliased mappings. There is
1172          * no risk as long as all huge pmd are found and have their
1173          * splitting bit set before __split_huge_page_refcount
1174          * runs. Finding the same huge pmd more than once during the
1175          * same rmap walk is not a problem.
1176          */
1177         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1178             pmd_trans_splitting(*pmd))
1179                 goto out;
1180         if (pmd_trans_huge(*pmd)) {
1181                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1182                           !pmd_trans_splitting(*pmd));
1183                 ret = pmd;
1184         }
1185 out:
1186         return ret;
1187 }
1188
1189 static int __split_huge_page_splitting(struct page *page,
1190                                        struct vm_area_struct *vma,
1191                                        unsigned long address)
1192 {
1193         struct mm_struct *mm = vma->vm_mm;
1194         pmd_t *pmd;
1195         int ret = 0;
1196         /* For mmu_notifiers */
1197         const unsigned long mmun_start = address;
1198         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1199
1200         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1201         spin_lock(&mm->page_table_lock);
1202         pmd = page_check_address_pmd(page, mm, address,
1203                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1204         if (pmd) {
1205                 /*
1206                  * We can't temporarily set the pmd to null in order
1207                  * to split it, the pmd must remain marked huge at all
1208                  * times or the VM won't take the pmd_trans_huge paths
1209                  * and it won't wait on the anon_vma->root->mutex to
1210                  * serialize against split_huge_page*.
1211                  */
1212                 pmdp_splitting_flush(vma, address, pmd);
1213                 ret = 1;
1214         }
1215         spin_unlock(&mm->page_table_lock);
1216         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1217
1218         return ret;
1219 }
1220
1221 static void __split_huge_page_refcount(struct page *page)
1222 {
1223         int i;
1224         struct zone *zone = page_zone(page);
1225         struct lruvec *lruvec;
1226         int tail_count = 0;
1227
1228         /* prevent PageLRU to go away from under us, and freeze lru stats */
1229         spin_lock_irq(&zone->lru_lock);
1230         lruvec = mem_cgroup_page_lruvec(page, zone);
1231
1232         compound_lock(page);
1233         /* complete memcg works before add pages to LRU */
1234         mem_cgroup_split_huge_fixup(page);
1235
1236         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1237                 struct page *page_tail = page + i;
1238
1239                 /* tail_page->_mapcount cannot change */
1240                 BUG_ON(page_mapcount(page_tail) < 0);
1241                 tail_count += page_mapcount(page_tail);
1242                 /* check for overflow */
1243                 BUG_ON(tail_count < 0);
1244                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1245                 /*
1246                  * tail_page->_count is zero and not changing from
1247                  * under us. But get_page_unless_zero() may be running
1248                  * from under us on the tail_page. If we used
1249                  * atomic_set() below instead of atomic_add(), we
1250                  * would then run atomic_set() concurrently with
1251                  * get_page_unless_zero(), and atomic_set() is
1252                  * implemented in C not using locked ops. spin_unlock
1253                  * on x86 sometime uses locked ops because of PPro
1254                  * errata 66, 92, so unless somebody can guarantee
1255                  * atomic_set() here would be safe on all archs (and
1256                  * not only on x86), it's safer to use atomic_add().
1257                  */
1258                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1259                            &page_tail->_count);
1260
1261                 /* after clearing PageTail the gup refcount can be released */
1262                 smp_mb();
1263
1264                 /*
1265                  * retain hwpoison flag of the poisoned tail page:
1266                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1267                  *   by the memory-failure.
1268                  */
1269                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1270                 page_tail->flags |= (page->flags &
1271                                      ((1L << PG_referenced) |
1272                                       (1L << PG_swapbacked) |
1273                                       (1L << PG_mlocked) |
1274                                       (1L << PG_uptodate)));
1275                 page_tail->flags |= (1L << PG_dirty);
1276
1277                 /* clear PageTail before overwriting first_page */
1278                 smp_wmb();
1279
1280                 /*
1281                  * __split_huge_page_splitting() already set the
1282                  * splitting bit in all pmd that could map this
1283                  * hugepage, that will ensure no CPU can alter the
1284                  * mapcount on the head page. The mapcount is only
1285                  * accounted in the head page and it has to be
1286                  * transferred to all tail pages in the below code. So
1287                  * for this code to be safe, the split the mapcount
1288                  * can't change. But that doesn't mean userland can't
1289                  * keep changing and reading the page contents while
1290                  * we transfer the mapcount, so the pmd splitting
1291                  * status is achieved setting a reserved bit in the
1292                  * pmd, not by clearing the present bit.
1293                 */
1294                 page_tail->_mapcount = page->_mapcount;
1295
1296                 BUG_ON(page_tail->mapping);
1297                 page_tail->mapping = page->mapping;
1298
1299                 page_tail->index = page->index + i;
1300
1301                 BUG_ON(!PageAnon(page_tail));
1302                 BUG_ON(!PageUptodate(page_tail));
1303                 BUG_ON(!PageDirty(page_tail));
1304                 BUG_ON(!PageSwapBacked(page_tail));
1305
1306                 lru_add_page_tail(page, page_tail, lruvec);
1307         }
1308         atomic_sub(tail_count, &page->_count);
1309         BUG_ON(atomic_read(&page->_count) <= 0);
1310
1311         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1312         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1313
1314         ClearPageCompound(page);
1315         compound_unlock(page);
1316         spin_unlock_irq(&zone->lru_lock);
1317
1318         for (i = 1; i < HPAGE_PMD_NR; i++) {
1319                 struct page *page_tail = page + i;
1320                 BUG_ON(page_count(page_tail) <= 0);
1321                 /*
1322                  * Tail pages may be freed if there wasn't any mapping
1323                  * like if add_to_swap() is running on a lru page that
1324                  * had its mapping zapped. And freeing these pages
1325                  * requires taking the lru_lock so we do the put_page
1326                  * of the tail pages after the split is complete.
1327                  */
1328                 put_page(page_tail);
1329         }
1330
1331         /*
1332          * Only the head page (now become a regular page) is required
1333          * to be pinned by the caller.
1334          */
1335         BUG_ON(page_count(page) <= 0);
1336 }
1337
1338 static int __split_huge_page_map(struct page *page,
1339                                  struct vm_area_struct *vma,
1340                                  unsigned long address)
1341 {
1342         struct mm_struct *mm = vma->vm_mm;
1343         pmd_t *pmd, _pmd;
1344         int ret = 0, i;
1345         pgtable_t pgtable;
1346         unsigned long haddr;
1347
1348         spin_lock(&mm->page_table_lock);
1349         pmd = page_check_address_pmd(page, mm, address,
1350                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1351         if (pmd) {
1352                 pgtable = pgtable_trans_huge_withdraw(mm);
1353                 pmd_populate(mm, &_pmd, pgtable);
1354
1355                 haddr = address;
1356                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1357                         pte_t *pte, entry;
1358                         BUG_ON(PageCompound(page+i));
1359                         entry = mk_pte(page + i, vma->vm_page_prot);
1360                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1361                         if (!pmd_write(*pmd))
1362                                 entry = pte_wrprotect(entry);
1363                         else
1364                                 BUG_ON(page_mapcount(page) != 1);
1365                         if (!pmd_young(*pmd))
1366                                 entry = pte_mkold(entry);
1367                         pte = pte_offset_map(&_pmd, haddr);
1368                         BUG_ON(!pte_none(*pte));
1369                         set_pte_at(mm, haddr, pte, entry);
1370                         pte_unmap(pte);
1371                 }
1372
1373                 smp_wmb(); /* make pte visible before pmd */
1374                 /*
1375                  * Up to this point the pmd is present and huge and
1376                  * userland has the whole access to the hugepage
1377                  * during the split (which happens in place). If we
1378                  * overwrite the pmd with the not-huge version
1379                  * pointing to the pte here (which of course we could
1380                  * if all CPUs were bug free), userland could trigger
1381                  * a small page size TLB miss on the small sized TLB
1382                  * while the hugepage TLB entry is still established
1383                  * in the huge TLB. Some CPU doesn't like that. See
1384                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1385                  * Erratum 383 on page 93. Intel should be safe but is
1386                  * also warns that it's only safe if the permission
1387                  * and cache attributes of the two entries loaded in
1388                  * the two TLB is identical (which should be the case
1389                  * here). But it is generally safer to never allow
1390                  * small and huge TLB entries for the same virtual
1391                  * address to be loaded simultaneously. So instead of
1392                  * doing "pmd_populate(); flush_tlb_range();" we first
1393                  * mark the current pmd notpresent (atomically because
1394                  * here the pmd_trans_huge and pmd_trans_splitting
1395                  * must remain set at all times on the pmd until the
1396                  * split is complete for this pmd), then we flush the
1397                  * SMP TLB and finally we write the non-huge version
1398                  * of the pmd entry with pmd_populate.
1399                  */
1400                 pmdp_invalidate(vma, address, pmd);
1401                 pmd_populate(mm, pmd, pgtable);
1402                 ret = 1;
1403         }
1404         spin_unlock(&mm->page_table_lock);
1405
1406         return ret;
1407 }
1408
1409 /* must be called with anon_vma->root->mutex hold */
1410 static void __split_huge_page(struct page *page,
1411                               struct anon_vma *anon_vma)
1412 {
1413         int mapcount, mapcount2;
1414         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1415         struct anon_vma_chain *avc;
1416
1417         BUG_ON(!PageHead(page));
1418         BUG_ON(PageTail(page));
1419
1420         mapcount = 0;
1421         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1422                 struct vm_area_struct *vma = avc->vma;
1423                 unsigned long addr = vma_address(page, vma);
1424                 BUG_ON(is_vma_temporary_stack(vma));
1425                 mapcount += __split_huge_page_splitting(page, vma, addr);
1426         }
1427         /*
1428          * It is critical that new vmas are added to the tail of the
1429          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1430          * and establishes a child pmd before
1431          * __split_huge_page_splitting() freezes the parent pmd (so if
1432          * we fail to prevent copy_huge_pmd() from running until the
1433          * whole __split_huge_page() is complete), we will still see
1434          * the newly established pmd of the child later during the
1435          * walk, to be able to set it as pmd_trans_splitting too.
1436          */
1437         if (mapcount != page_mapcount(page))
1438                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1439                        mapcount, page_mapcount(page));
1440         BUG_ON(mapcount != page_mapcount(page));
1441
1442         __split_huge_page_refcount(page);
1443
1444         mapcount2 = 0;
1445         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1446                 struct vm_area_struct *vma = avc->vma;
1447                 unsigned long addr = vma_address(page, vma);
1448                 BUG_ON(is_vma_temporary_stack(vma));
1449                 mapcount2 += __split_huge_page_map(page, vma, addr);
1450         }
1451         if (mapcount != mapcount2)
1452                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1453                        mapcount, mapcount2, page_mapcount(page));
1454         BUG_ON(mapcount != mapcount2);
1455 }
1456
1457 int split_huge_page(struct page *page)
1458 {
1459         struct anon_vma *anon_vma;
1460         int ret = 1;
1461
1462         BUG_ON(!PageAnon(page));
1463         anon_vma = page_lock_anon_vma(page);
1464         if (!anon_vma)
1465                 goto out;
1466         ret = 0;
1467         if (!PageCompound(page))
1468                 goto out_unlock;
1469
1470         BUG_ON(!PageSwapBacked(page));
1471         __split_huge_page(page, anon_vma);
1472         count_vm_event(THP_SPLIT);
1473
1474         BUG_ON(PageCompound(page));
1475 out_unlock:
1476         page_unlock_anon_vma(anon_vma);
1477 out:
1478         return ret;
1479 }
1480
1481 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1482
1483 int hugepage_madvise(struct vm_area_struct *vma,
1484                      unsigned long *vm_flags, int advice)
1485 {
1486         struct mm_struct *mm = vma->vm_mm;
1487
1488         switch (advice) {
1489         case MADV_HUGEPAGE:
1490                 /*
1491                  * Be somewhat over-protective like KSM for now!
1492                  */
1493                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1494                         return -EINVAL;
1495                 if (mm->def_flags & VM_NOHUGEPAGE)
1496                         return -EINVAL;
1497                 *vm_flags &= ~VM_NOHUGEPAGE;
1498                 *vm_flags |= VM_HUGEPAGE;
1499                 /*
1500                  * If the vma become good for khugepaged to scan,
1501                  * register it here without waiting a page fault that
1502                  * may not happen any time soon.
1503                  */
1504                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1505                         return -ENOMEM;
1506                 break;
1507         case MADV_NOHUGEPAGE:
1508                 /*
1509                  * Be somewhat over-protective like KSM for now!
1510                  */
1511                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1512                         return -EINVAL;
1513                 *vm_flags &= ~VM_HUGEPAGE;
1514                 *vm_flags |= VM_NOHUGEPAGE;
1515                 /*
1516                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1517                  * this vma even if we leave the mm registered in khugepaged if
1518                  * it got registered before VM_NOHUGEPAGE was set.
1519                  */
1520                 break;
1521         }
1522
1523         return 0;
1524 }
1525
1526 static int __init khugepaged_slab_init(void)
1527 {
1528         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1529                                           sizeof(struct mm_slot),
1530                                           __alignof__(struct mm_slot), 0, NULL);
1531         if (!mm_slot_cache)
1532                 return -ENOMEM;
1533
1534         return 0;
1535 }
1536
1537 static void __init khugepaged_slab_free(void)
1538 {
1539         kmem_cache_destroy(mm_slot_cache);
1540         mm_slot_cache = NULL;
1541 }
1542
1543 static inline struct mm_slot *alloc_mm_slot(void)
1544 {
1545         if (!mm_slot_cache)     /* initialization failed */
1546                 return NULL;
1547         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1548 }
1549
1550 static inline void free_mm_slot(struct mm_slot *mm_slot)
1551 {
1552         kmem_cache_free(mm_slot_cache, mm_slot);
1553 }
1554
1555 static int __init mm_slots_hash_init(void)
1556 {
1557         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1558                                 GFP_KERNEL);
1559         if (!mm_slots_hash)
1560                 return -ENOMEM;
1561         return 0;
1562 }
1563
1564 #if 0
1565 static void __init mm_slots_hash_free(void)
1566 {
1567         kfree(mm_slots_hash);
1568         mm_slots_hash = NULL;
1569 }
1570 #endif
1571
1572 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1573 {
1574         struct mm_slot *mm_slot;
1575         struct hlist_head *bucket;
1576         struct hlist_node *node;
1577
1578         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1579                                 % MM_SLOTS_HASH_HEADS];
1580         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1581                 if (mm == mm_slot->mm)
1582                         return mm_slot;
1583         }
1584         return NULL;
1585 }
1586
1587 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1588                                     struct mm_slot *mm_slot)
1589 {
1590         struct hlist_head *bucket;
1591
1592         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1593                                 % MM_SLOTS_HASH_HEADS];
1594         mm_slot->mm = mm;
1595         hlist_add_head(&mm_slot->hash, bucket);
1596 }
1597
1598 static inline int khugepaged_test_exit(struct mm_struct *mm)
1599 {
1600         return atomic_read(&mm->mm_users) == 0;
1601 }
1602
1603 int __khugepaged_enter(struct mm_struct *mm)
1604 {
1605         struct mm_slot *mm_slot;
1606         int wakeup;
1607
1608         mm_slot = alloc_mm_slot();
1609         if (!mm_slot)
1610                 return -ENOMEM;
1611
1612         /* __khugepaged_exit() must not run from under us */
1613         VM_BUG_ON(khugepaged_test_exit(mm));
1614         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1615                 free_mm_slot(mm_slot);
1616                 return 0;
1617         }
1618
1619         spin_lock(&khugepaged_mm_lock);
1620         insert_to_mm_slots_hash(mm, mm_slot);
1621         /*
1622          * Insert just behind the scanning cursor, to let the area settle
1623          * down a little.
1624          */
1625         wakeup = list_empty(&khugepaged_scan.mm_head);
1626         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1627         spin_unlock(&khugepaged_mm_lock);
1628
1629         atomic_inc(&mm->mm_count);
1630         if (wakeup)
1631                 wake_up_interruptible(&khugepaged_wait);
1632
1633         return 0;
1634 }
1635
1636 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1637 {
1638         unsigned long hstart, hend;
1639         if (!vma->anon_vma)
1640                 /*
1641                  * Not yet faulted in so we will register later in the
1642                  * page fault if needed.
1643                  */
1644                 return 0;
1645         if (vma->vm_ops)
1646                 /* khugepaged not yet working on file or special mappings */
1647                 return 0;
1648         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1649         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1650         hend = vma->vm_end & HPAGE_PMD_MASK;
1651         if (hstart < hend)
1652                 return khugepaged_enter(vma);
1653         return 0;
1654 }
1655
1656 void __khugepaged_exit(struct mm_struct *mm)
1657 {
1658         struct mm_slot *mm_slot;
1659         int free = 0;
1660
1661         spin_lock(&khugepaged_mm_lock);
1662         mm_slot = get_mm_slot(mm);
1663         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1664                 hlist_del(&mm_slot->hash);
1665                 list_del(&mm_slot->mm_node);
1666                 free = 1;
1667         }
1668         spin_unlock(&khugepaged_mm_lock);
1669
1670         if (free) {
1671                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1672                 free_mm_slot(mm_slot);
1673                 mmdrop(mm);
1674         } else if (mm_slot) {
1675                 /*
1676                  * This is required to serialize against
1677                  * khugepaged_test_exit() (which is guaranteed to run
1678                  * under mmap sem read mode). Stop here (after we
1679                  * return all pagetables will be destroyed) until
1680                  * khugepaged has finished working on the pagetables
1681                  * under the mmap_sem.
1682                  */
1683                 down_write(&mm->mmap_sem);
1684                 up_write(&mm->mmap_sem);
1685         }
1686 }
1687
1688 static void release_pte_page(struct page *page)
1689 {
1690         /* 0 stands for page_is_file_cache(page) == false */
1691         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1692         unlock_page(page);
1693         putback_lru_page(page);
1694 }
1695
1696 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1697 {
1698         while (--_pte >= pte) {
1699                 pte_t pteval = *_pte;
1700                 if (!pte_none(pteval))
1701                         release_pte_page(pte_page(pteval));
1702         }
1703 }
1704
1705 static void release_all_pte_pages(pte_t *pte)
1706 {
1707         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1708 }
1709
1710 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1711                                         unsigned long address,
1712                                         pte_t *pte)
1713 {
1714         struct page *page;
1715         pte_t *_pte;
1716         int referenced = 0, isolated = 0, none = 0;
1717         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1718              _pte++, address += PAGE_SIZE) {
1719                 pte_t pteval = *_pte;
1720                 if (pte_none(pteval)) {
1721                         if (++none <= khugepaged_max_ptes_none)
1722                                 continue;
1723                         else {
1724                                 release_pte_pages(pte, _pte);
1725                                 goto out;
1726                         }
1727                 }
1728                 if (!pte_present(pteval) || !pte_write(pteval)) {
1729                         release_pte_pages(pte, _pte);
1730                         goto out;
1731                 }
1732                 page = vm_normal_page(vma, address, pteval);
1733                 if (unlikely(!page)) {
1734                         release_pte_pages(pte, _pte);
1735                         goto out;
1736                 }
1737                 VM_BUG_ON(PageCompound(page));
1738                 BUG_ON(!PageAnon(page));
1739                 VM_BUG_ON(!PageSwapBacked(page));
1740
1741                 /* cannot use mapcount: can't collapse if there's a gup pin */
1742                 if (page_count(page) != 1) {
1743                         release_pte_pages(pte, _pte);
1744                         goto out;
1745                 }
1746                 /*
1747                  * We can do it before isolate_lru_page because the
1748                  * page can't be freed from under us. NOTE: PG_lock
1749                  * is needed to serialize against split_huge_page
1750                  * when invoked from the VM.
1751                  */
1752                 if (!trylock_page(page)) {
1753                         release_pte_pages(pte, _pte);
1754                         goto out;
1755                 }
1756                 /*
1757                  * Isolate the page to avoid collapsing an hugepage
1758                  * currently in use by the VM.
1759                  */
1760                 if (isolate_lru_page(page)) {
1761                         unlock_page(page);
1762                         release_pte_pages(pte, _pte);
1763                         goto out;
1764                 }
1765                 /* 0 stands for page_is_file_cache(page) == false */
1766                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1767                 VM_BUG_ON(!PageLocked(page));
1768                 VM_BUG_ON(PageLRU(page));
1769
1770                 /* If there is no mapped pte young don't collapse the page */
1771                 if (pte_young(pteval) || PageReferenced(page) ||
1772                     mmu_notifier_test_young(vma->vm_mm, address))
1773                         referenced = 1;
1774         }
1775         if (unlikely(!referenced))
1776                 release_all_pte_pages(pte);
1777         else
1778                 isolated = 1;
1779 out:
1780         return isolated;
1781 }
1782
1783 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1784                                       struct vm_area_struct *vma,
1785                                       unsigned long address,
1786                                       spinlock_t *ptl)
1787 {
1788         pte_t *_pte;
1789         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1790                 pte_t pteval = *_pte;
1791                 struct page *src_page;
1792
1793                 if (pte_none(pteval)) {
1794                         clear_user_highpage(page, address);
1795                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1796                 } else {
1797                         src_page = pte_page(pteval);
1798                         copy_user_highpage(page, src_page, address, vma);
1799                         VM_BUG_ON(page_mapcount(src_page) != 1);
1800                         release_pte_page(src_page);
1801                         /*
1802                          * ptl mostly unnecessary, but preempt has to
1803                          * be disabled to update the per-cpu stats
1804                          * inside page_remove_rmap().
1805                          */
1806                         spin_lock(ptl);
1807                         /*
1808                          * paravirt calls inside pte_clear here are
1809                          * superfluous.
1810                          */
1811                         pte_clear(vma->vm_mm, address, _pte);
1812                         page_remove_rmap(src_page);
1813                         spin_unlock(ptl);
1814                         free_page_and_swap_cache(src_page);
1815                 }
1816
1817                 address += PAGE_SIZE;
1818                 page++;
1819         }
1820 }
1821
1822 static void khugepaged_alloc_sleep(void)
1823 {
1824         wait_event_freezable_timeout(khugepaged_wait, false,
1825                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1826 }
1827
1828 #ifdef CONFIG_NUMA
1829 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1830 {
1831         if (IS_ERR(*hpage)) {
1832                 if (!*wait)
1833                         return false;
1834
1835                 *wait = false;
1836                 *hpage = NULL;
1837                 khugepaged_alloc_sleep();
1838         } else if (*hpage) {
1839                 put_page(*hpage);
1840                 *hpage = NULL;
1841         }
1842
1843         return true;
1844 }
1845
1846 static struct page
1847 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1848                        struct vm_area_struct *vma, unsigned long address,
1849                        int node)
1850 {
1851         VM_BUG_ON(*hpage);
1852         /*
1853          * Allocate the page while the vma is still valid and under
1854          * the mmap_sem read mode so there is no memory allocation
1855          * later when we take the mmap_sem in write mode. This is more
1856          * friendly behavior (OTOH it may actually hide bugs) to
1857          * filesystems in userland with daemons allocating memory in
1858          * the userland I/O paths.  Allocating memory with the
1859          * mmap_sem in read mode is good idea also to allow greater
1860          * scalability.
1861          */
1862         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1863                                       node, __GFP_OTHER_NODE);
1864
1865         /*
1866          * After allocating the hugepage, release the mmap_sem read lock in
1867          * preparation for taking it in write mode.
1868          */
1869         up_read(&mm->mmap_sem);
1870         if (unlikely(!*hpage)) {
1871                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1872                 *hpage = ERR_PTR(-ENOMEM);
1873                 return NULL;
1874         }
1875
1876         count_vm_event(THP_COLLAPSE_ALLOC);
1877         return *hpage;
1878 }
1879 #else
1880 static struct page *khugepaged_alloc_hugepage(bool *wait)
1881 {
1882         struct page *hpage;
1883
1884         do {
1885                 hpage = alloc_hugepage(khugepaged_defrag());
1886                 if (!hpage) {
1887                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1888                         if (!*wait)
1889                                 return NULL;
1890
1891                         *wait = false;
1892                         khugepaged_alloc_sleep();
1893                 } else
1894                         count_vm_event(THP_COLLAPSE_ALLOC);
1895         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
1896
1897         return hpage;
1898 }
1899
1900 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1901 {
1902         if (!*hpage)
1903                 *hpage = khugepaged_alloc_hugepage(wait);
1904
1905         if (unlikely(!*hpage))
1906                 return false;
1907
1908         return true;
1909 }
1910
1911 static struct page
1912 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1913                        struct vm_area_struct *vma, unsigned long address,
1914                        int node)
1915 {
1916         up_read(&mm->mmap_sem);
1917         VM_BUG_ON(!*hpage);
1918         return  *hpage;
1919 }
1920 #endif
1921
1922 static void collapse_huge_page(struct mm_struct *mm,
1923                                    unsigned long address,
1924                                    struct page **hpage,
1925                                    struct vm_area_struct *vma,
1926                                    int node)
1927 {
1928         pgd_t *pgd;
1929         pud_t *pud;
1930         pmd_t *pmd, _pmd;
1931         pte_t *pte;
1932         pgtable_t pgtable;
1933         struct page *new_page;
1934         spinlock_t *ptl;
1935         int isolated;
1936         unsigned long hstart, hend;
1937         unsigned long mmun_start;       /* For mmu_notifiers */
1938         unsigned long mmun_end;         /* For mmu_notifiers */
1939
1940         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1941
1942         /* release the mmap_sem read lock. */
1943         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
1944         if (!new_page)
1945                 return;
1946
1947         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
1948                 return;
1949
1950         /*
1951          * Prevent all access to pagetables with the exception of
1952          * gup_fast later hanlded by the ptep_clear_flush and the VM
1953          * handled by the anon_vma lock + PG_lock.
1954          */
1955         down_write(&mm->mmap_sem);
1956         if (unlikely(khugepaged_test_exit(mm)))
1957                 goto out;
1958
1959         vma = find_vma(mm, address);
1960         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1961         hend = vma->vm_end & HPAGE_PMD_MASK;
1962         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1963                 goto out;
1964
1965         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1966             (vma->vm_flags & VM_NOHUGEPAGE))
1967                 goto out;
1968
1969         if (!vma->anon_vma || vma->vm_ops)
1970                 goto out;
1971         if (is_vma_temporary_stack(vma))
1972                 goto out;
1973         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1974
1975         pgd = pgd_offset(mm, address);
1976         if (!pgd_present(*pgd))
1977                 goto out;
1978
1979         pud = pud_offset(pgd, address);
1980         if (!pud_present(*pud))
1981                 goto out;
1982
1983         pmd = pmd_offset(pud, address);
1984         /* pmd can't go away or become huge under us */
1985         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1986                 goto out;
1987
1988         anon_vma_lock(vma->anon_vma);
1989
1990         pte = pte_offset_map(pmd, address);
1991         ptl = pte_lockptr(mm, pmd);
1992
1993         mmun_start = address;
1994         mmun_end   = address + HPAGE_PMD_SIZE;
1995         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1996         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1997         /*
1998          * After this gup_fast can't run anymore. This also removes
1999          * any huge TLB entry from the CPU so we won't allow
2000          * huge and small TLB entries for the same virtual address
2001          * to avoid the risk of CPU bugs in that area.
2002          */
2003         _pmd = pmdp_clear_flush(vma, address, pmd);
2004         spin_unlock(&mm->page_table_lock);
2005         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2006
2007         spin_lock(ptl);
2008         isolated = __collapse_huge_page_isolate(vma, address, pte);
2009         spin_unlock(ptl);
2010
2011         if (unlikely(!isolated)) {
2012                 pte_unmap(pte);
2013                 spin_lock(&mm->page_table_lock);
2014                 BUG_ON(!pmd_none(*pmd));
2015                 set_pmd_at(mm, address, pmd, _pmd);
2016                 spin_unlock(&mm->page_table_lock);
2017                 anon_vma_unlock(vma->anon_vma);
2018                 goto out;
2019         }
2020
2021         /*
2022          * All pages are isolated and locked so anon_vma rmap
2023          * can't run anymore.
2024          */
2025         anon_vma_unlock(vma->anon_vma);
2026
2027         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2028         pte_unmap(pte);
2029         __SetPageUptodate(new_page);
2030         pgtable = pmd_pgtable(_pmd);
2031
2032         _pmd = mk_pmd(new_page, vma->vm_page_prot);
2033         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2034         _pmd = pmd_mkhuge(_pmd);
2035
2036         /*
2037          * spin_lock() below is not the equivalent of smp_wmb(), so
2038          * this is needed to avoid the copy_huge_page writes to become
2039          * visible after the set_pmd_at() write.
2040          */
2041         smp_wmb();
2042
2043         spin_lock(&mm->page_table_lock);
2044         BUG_ON(!pmd_none(*pmd));
2045         page_add_new_anon_rmap(new_page, vma, address);
2046         set_pmd_at(mm, address, pmd, _pmd);
2047         update_mmu_cache_pmd(vma, address, pmd);
2048         pgtable_trans_huge_deposit(mm, pgtable);
2049         spin_unlock(&mm->page_table_lock);
2050
2051         *hpage = NULL;
2052
2053         khugepaged_pages_collapsed++;
2054 out_up_write:
2055         up_write(&mm->mmap_sem);
2056         return;
2057
2058 out:
2059         mem_cgroup_uncharge_page(new_page);
2060         goto out_up_write;
2061 }
2062
2063 static int khugepaged_scan_pmd(struct mm_struct *mm,
2064                                struct vm_area_struct *vma,
2065                                unsigned long address,
2066                                struct page **hpage)
2067 {
2068         pgd_t *pgd;
2069         pud_t *pud;
2070         pmd_t *pmd;
2071         pte_t *pte, *_pte;
2072         int ret = 0, referenced = 0, none = 0;
2073         struct page *page;
2074         unsigned long _address;
2075         spinlock_t *ptl;
2076         int node = -1;
2077
2078         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2079
2080         pgd = pgd_offset(mm, address);
2081         if (!pgd_present(*pgd))
2082                 goto out;
2083
2084         pud = pud_offset(pgd, address);
2085         if (!pud_present(*pud))
2086                 goto out;
2087
2088         pmd = pmd_offset(pud, address);
2089         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2090                 goto out;
2091
2092         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2093         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2094              _pte++, _address += PAGE_SIZE) {
2095                 pte_t pteval = *_pte;
2096                 if (pte_none(pteval)) {
2097                         if (++none <= khugepaged_max_ptes_none)
2098                                 continue;
2099                         else
2100                                 goto out_unmap;
2101                 }
2102                 if (!pte_present(pteval) || !pte_write(pteval))
2103                         goto out_unmap;
2104                 page = vm_normal_page(vma, _address, pteval);
2105                 if (unlikely(!page))
2106                         goto out_unmap;
2107                 /*
2108                  * Chose the node of the first page. This could
2109                  * be more sophisticated and look at more pages,
2110                  * but isn't for now.
2111                  */
2112                 if (node == -1)
2113                         node = page_to_nid(page);
2114                 VM_BUG_ON(PageCompound(page));
2115                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2116                         goto out_unmap;
2117                 /* cannot use mapcount: can't collapse if there's a gup pin */
2118                 if (page_count(page) != 1)
2119                         goto out_unmap;
2120                 if (pte_young(pteval) || PageReferenced(page) ||
2121                     mmu_notifier_test_young(vma->vm_mm, address))
2122                         referenced = 1;
2123         }
2124         if (referenced)
2125                 ret = 1;
2126 out_unmap:
2127         pte_unmap_unlock(pte, ptl);
2128         if (ret)
2129                 /* collapse_huge_page will return with the mmap_sem released */
2130                 collapse_huge_page(mm, address, hpage, vma, node);
2131 out:
2132         return ret;
2133 }
2134
2135 static void collect_mm_slot(struct mm_slot *mm_slot)
2136 {
2137         struct mm_struct *mm = mm_slot->mm;
2138
2139         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2140
2141         if (khugepaged_test_exit(mm)) {
2142                 /* free mm_slot */
2143                 hlist_del(&mm_slot->hash);
2144                 list_del(&mm_slot->mm_node);
2145
2146                 /*
2147                  * Not strictly needed because the mm exited already.
2148                  *
2149                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2150                  */
2151
2152                 /* khugepaged_mm_lock actually not necessary for the below */
2153                 free_mm_slot(mm_slot);
2154                 mmdrop(mm);
2155         }
2156 }
2157
2158 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2159                                             struct page **hpage)
2160         __releases(&khugepaged_mm_lock)
2161         __acquires(&khugepaged_mm_lock)
2162 {
2163         struct mm_slot *mm_slot;
2164         struct mm_struct *mm;
2165         struct vm_area_struct *vma;
2166         int progress = 0;
2167
2168         VM_BUG_ON(!pages);
2169         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2170
2171         if (khugepaged_scan.mm_slot)
2172                 mm_slot = khugepaged_scan.mm_slot;
2173         else {
2174                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2175                                      struct mm_slot, mm_node);
2176                 khugepaged_scan.address = 0;
2177                 khugepaged_scan.mm_slot = mm_slot;
2178         }
2179         spin_unlock(&khugepaged_mm_lock);
2180
2181         mm = mm_slot->mm;
2182         down_read(&mm->mmap_sem);
2183         if (unlikely(khugepaged_test_exit(mm)))
2184                 vma = NULL;
2185         else
2186                 vma = find_vma(mm, khugepaged_scan.address);
2187
2188         progress++;
2189         for (; vma; vma = vma->vm_next) {
2190                 unsigned long hstart, hend;
2191
2192                 cond_resched();
2193                 if (unlikely(khugepaged_test_exit(mm))) {
2194                         progress++;
2195                         break;
2196                 }
2197
2198                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2199                      !khugepaged_always()) ||
2200                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2201                 skip:
2202                         progress++;
2203                         continue;
2204                 }
2205                 if (!vma->anon_vma || vma->vm_ops)
2206                         goto skip;
2207                 if (is_vma_temporary_stack(vma))
2208                         goto skip;
2209                 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2210
2211                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2212                 hend = vma->vm_end & HPAGE_PMD_MASK;
2213                 if (hstart >= hend)
2214                         goto skip;
2215                 if (khugepaged_scan.address > hend)
2216                         goto skip;
2217                 if (khugepaged_scan.address < hstart)
2218                         khugepaged_scan.address = hstart;
2219                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2220
2221                 while (khugepaged_scan.address < hend) {
2222                         int ret;
2223                         cond_resched();
2224                         if (unlikely(khugepaged_test_exit(mm)))
2225                                 goto breakouterloop;
2226
2227                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2228                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2229                                   hend);
2230                         ret = khugepaged_scan_pmd(mm, vma,
2231                                                   khugepaged_scan.address,
2232                                                   hpage);
2233                         /* move to next address */
2234                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2235                         progress += HPAGE_PMD_NR;
2236                         if (ret)
2237                                 /* we released mmap_sem so break loop */
2238                                 goto breakouterloop_mmap_sem;
2239                         if (progress >= pages)
2240                                 goto breakouterloop;
2241                 }
2242         }
2243 breakouterloop:
2244         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2245 breakouterloop_mmap_sem:
2246
2247         spin_lock(&khugepaged_mm_lock);
2248         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2249         /*
2250          * Release the current mm_slot if this mm is about to die, or
2251          * if we scanned all vmas of this mm.
2252          */
2253         if (khugepaged_test_exit(mm) || !vma) {
2254                 /*
2255                  * Make sure that if mm_users is reaching zero while
2256                  * khugepaged runs here, khugepaged_exit will find
2257                  * mm_slot not pointing to the exiting mm.
2258                  */
2259                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2260                         khugepaged_scan.mm_slot = list_entry(
2261                                 mm_slot->mm_node.next,
2262                                 struct mm_slot, mm_node);
2263                         khugepaged_scan.address = 0;
2264                 } else {
2265                         khugepaged_scan.mm_slot = NULL;
2266                         khugepaged_full_scans++;
2267                 }
2268
2269                 collect_mm_slot(mm_slot);
2270         }
2271
2272         return progress;
2273 }
2274
2275 static int khugepaged_has_work(void)
2276 {
2277         return !list_empty(&khugepaged_scan.mm_head) &&
2278                 khugepaged_enabled();
2279 }
2280
2281 static int khugepaged_wait_event(void)
2282 {
2283         return !list_empty(&khugepaged_scan.mm_head) ||
2284                 kthread_should_stop();
2285 }
2286
2287 static void khugepaged_do_scan(void)
2288 {
2289         struct page *hpage = NULL;
2290         unsigned int progress = 0, pass_through_head = 0;
2291         unsigned int pages = khugepaged_pages_to_scan;
2292         bool wait = true;
2293
2294         barrier(); /* write khugepaged_pages_to_scan to local stack */
2295
2296         while (progress < pages) {
2297                 if (!khugepaged_prealloc_page(&hpage, &wait))
2298                         break;
2299
2300                 cond_resched();
2301
2302                 if (unlikely(kthread_should_stop() || freezing(current)))
2303                         break;
2304
2305                 spin_lock(&khugepaged_mm_lock);
2306                 if (!khugepaged_scan.mm_slot)
2307                         pass_through_head++;
2308                 if (khugepaged_has_work() &&
2309                     pass_through_head < 2)
2310                         progress += khugepaged_scan_mm_slot(pages - progress,
2311                                                             &hpage);
2312                 else
2313                         progress = pages;
2314                 spin_unlock(&khugepaged_mm_lock);
2315         }
2316
2317         if (!IS_ERR_OR_NULL(hpage))
2318                 put_page(hpage);
2319 }
2320
2321 static void khugepaged_wait_work(void)
2322 {
2323         try_to_freeze();
2324
2325         if (khugepaged_has_work()) {
2326                 if (!khugepaged_scan_sleep_millisecs)
2327                         return;
2328
2329                 wait_event_freezable_timeout(khugepaged_wait,
2330                                              kthread_should_stop(),
2331                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2332                 return;
2333         }
2334
2335         if (khugepaged_enabled())
2336                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2337 }
2338
2339 static int khugepaged(void *none)
2340 {
2341         struct mm_slot *mm_slot;
2342
2343         set_freezable();
2344         set_user_nice(current, 19);
2345
2346         while (!kthread_should_stop()) {
2347                 khugepaged_do_scan();
2348                 khugepaged_wait_work();
2349         }
2350
2351         spin_lock(&khugepaged_mm_lock);
2352         mm_slot = khugepaged_scan.mm_slot;
2353         khugepaged_scan.mm_slot = NULL;
2354         if (mm_slot)
2355                 collect_mm_slot(mm_slot);
2356         spin_unlock(&khugepaged_mm_lock);
2357         return 0;
2358 }
2359
2360 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2361 {
2362         struct page *page;
2363
2364         spin_lock(&mm->page_table_lock);
2365         if (unlikely(!pmd_trans_huge(*pmd))) {
2366                 spin_unlock(&mm->page_table_lock);
2367                 return;
2368         }
2369         page = pmd_page(*pmd);
2370         VM_BUG_ON(!page_count(page));
2371         get_page(page);
2372         spin_unlock(&mm->page_table_lock);
2373
2374         split_huge_page(page);
2375
2376         put_page(page);
2377         BUG_ON(pmd_trans_huge(*pmd));
2378 }
2379
2380 static void split_huge_page_address(struct mm_struct *mm,
2381                                     unsigned long address)
2382 {
2383         pgd_t *pgd;
2384         pud_t *pud;
2385         pmd_t *pmd;
2386
2387         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2388
2389         pgd = pgd_offset(mm, address);
2390         if (!pgd_present(*pgd))
2391                 return;
2392
2393         pud = pud_offset(pgd, address);
2394         if (!pud_present(*pud))
2395                 return;
2396
2397         pmd = pmd_offset(pud, address);
2398         if (!pmd_present(*pmd))
2399                 return;
2400         /*
2401          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2402          * materialize from under us.
2403          */
2404         split_huge_page_pmd(mm, pmd);
2405 }
2406
2407 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2408                              unsigned long start,
2409                              unsigned long end,
2410                              long adjust_next)
2411 {
2412         /*
2413          * If the new start address isn't hpage aligned and it could
2414          * previously contain an hugepage: check if we need to split
2415          * an huge pmd.
2416          */
2417         if (start & ~HPAGE_PMD_MASK &&
2418             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2419             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2420                 split_huge_page_address(vma->vm_mm, start);
2421
2422         /*
2423          * If the new end address isn't hpage aligned and it could
2424          * previously contain an hugepage: check if we need to split
2425          * an huge pmd.
2426          */
2427         if (end & ~HPAGE_PMD_MASK &&
2428             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2429             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2430                 split_huge_page_address(vma->vm_mm, end);
2431
2432         /*
2433          * If we're also updating the vma->vm_next->vm_start, if the new
2434          * vm_next->vm_start isn't page aligned and it could previously
2435          * contain an hugepage: check if we need to split an huge pmd.
2436          */
2437         if (adjust_next > 0) {
2438                 struct vm_area_struct *next = vma->vm_next;
2439                 unsigned long nstart = next->vm_start;
2440                 nstart += adjust_next << PAGE_SHIFT;
2441                 if (nstart & ~HPAGE_PMD_MASK &&
2442                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2443                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2444                         split_huge_page_address(next->vm_mm, nstart);
2445         }
2446 }