mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/mm_inline.h>
  16 #include <linux/kthread.h>
  17 #include <linux/khugepaged.h>
  18 #include <linux/freezer.h>
  19 #include <linux/mman.h>
  20 #include <asm/tlb.h>
  21 #include <asm/pgalloc.h>
  22 #include "internal.h"
  23
  24 /*
  25  * By default transparent hugepage support is enabled for all mappings
  26  * and khugepaged scans all mappings. Defrag is only invoked by
  27  * khugepaged hugepage allocations and by page faults inside
  28  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  29  * allocations.
  30  */
  31 unsigned long transparent_hugepage_flags __read_mostly =
  32 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  33         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  34 #endif
  35 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  36         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  37 #endif
  38         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  39         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
  40
  41 /* default scan 8*512 pte (or vmas) every 30 second */
  42 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  43 static unsigned int khugepaged_pages_collapsed;
  44 static unsigned int khugepaged_full_scans;
  45 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  46 /* during fragmentation poll the hugepage allocator once every minute */
  47 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  48 static struct task_struct *khugepaged_thread __read_mostly;
  49 static DEFINE_MUTEX(khugepaged_mutex);
  50 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  51 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  52 /*
  53  * default collapse hugepages if there is at least one pte mapped like
  54  * it would have happened if the vma was large enough during page
  55  * fault.
  56  */
  57 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  58
  59 static int khugepaged(void *none);
  60 static int mm_slots_hash_init(void);
  61 static int khugepaged_slab_init(void);
  62 static void khugepaged_slab_free(void);
  63
  64 #define MM_SLOTS_HASH_HEADS 1024
  65 static struct hlist_head *mm_slots_hash __read_mostly;
  66 static struct kmem_cache *mm_slot_cache __read_mostly;
  67
  68 /**
  69  * struct mm_slot - hash lookup from mm to mm_slot
  70  * @hash: hash collision list
  71  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  72  * @mm: the mm that this information is valid for
  73  */
  74 struct mm_slot {
  75         struct hlist_node hash;
  76         struct list_head mm_node;
  77         struct mm_struct *mm;
  78 };
  79
  80 /**
  81  * struct khugepaged_scan - cursor for scanning
  82  * @mm_head: the head of the mm list to scan
  83  * @mm_slot: the current mm_slot we are scanning
  84  * @address: the next address inside that to be scanned
  85  *
  86  * There is only the one khugepaged_scan instance of this cursor structure.
  87  */
  88 struct khugepaged_scan {
  89         struct list_head mm_head;
  90         struct mm_slot *mm_slot;
  91         unsigned long address;
  92 };
  93 static struct khugepaged_scan khugepaged_scan = {
  94         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
  95 };
  96
  97
  98 static int set_recommended_min_free_kbytes(void)
  99 {
 100         struct zone *zone;
 101         int nr_zones = 0;
 102         unsigned long recommended_min;
 103         extern int min_free_kbytes;
 104
 105         if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG,
 106                       &transparent_hugepage_flags) &&
 107             !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 108                       &transparent_hugepage_flags))
 109                 return 0;
 110
 111         for_each_populated_zone(zone)
 112                 nr_zones++;
 113
 114         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 115         recommended_min = pageblock_nr_pages * nr_zones * 2;
 116
 117         /*
 118          * Make sure that on average at least two pageblocks are almost free
 119          * of another type, one for a migratetype to fall back to and a
 120          * second to avoid subsequent fallbacks of other types There are 3
 121          * MIGRATE_TYPES we care about.
 122          */
 123         recommended_min += pageblock_nr_pages * nr_zones *
 124                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 125
 126         /* don't ever allow to reserve more than 5% of the lowmem */
 127         recommended_min = min(recommended_min,
 128                               (unsigned long) nr_free_buffer_pages() / 20);
 129         recommended_min <<= (PAGE_SHIFT-10);
 130
 131         if (recommended_min > min_free_kbytes)
 132                 min_free_kbytes = recommended_min;
 133         setup_per_zone_wmarks();
 134         return 0;
 135 }
 136 late_initcall(set_recommended_min_free_kbytes);
 137
 138 static int start_khugepaged(void)
 139 {
 140         int err = 0;
 141         if (khugepaged_enabled()) {
 142                 int wakeup;
 143                 if (unlikely(!mm_slot_cache || !mm_slots_hash)) {
 144                         err = -ENOMEM;
 145                         goto out;
 146                 }
 147                 mutex_lock(&khugepaged_mutex);
 148                 if (!khugepaged_thread)
 149                         khugepaged_thread = kthread_run(khugepaged, NULL,
 150                                                         "khugepaged");
 151                 if (unlikely(IS_ERR(khugepaged_thread))) {
 152                         printk(KERN_ERR
 153                                "khugepaged: kthread_run(khugepaged) failed\n");
 154                         err = PTR_ERR(khugepaged_thread);
 155                         khugepaged_thread = NULL;
 156                 }
 157                 wakeup = !list_empty(&khugepaged_scan.mm_head);
 158                 mutex_unlock(&khugepaged_mutex);
 159                 if (wakeup)
 160                         wake_up_interruptible(&khugepaged_wait);
 161
 162                 set_recommended_min_free_kbytes();
 163         } else
 164                 /* wakeup to exit */
 165                 wake_up_interruptible(&khugepaged_wait);
 166 out:
 167         return err;
 168 }
 169
 170 #ifdef CONFIG_SYSFS
 171
 172 static ssize_t double_flag_show(struct kobject *kobj,
 173                                 struct kobj_attribute *attr, char *buf,
 174                                 enum transparent_hugepage_flag enabled,
 175                                 enum transparent_hugepage_flag req_madv)
 176 {
 177         if (test_bit(enabled, &transparent_hugepage_flags)) {
 178                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 179                 return sprintf(buf, "[always] madvise never\n");
 180         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 181                 return sprintf(buf, "always [madvise] never\n");
 182         else
 183                 return sprintf(buf, "always madvise [never]\n");
 184 }
 185 static ssize_t double_flag_store(struct kobject *kobj,
 186                                  struct kobj_attribute *attr,
 187                                  const char *buf, size_t count,
 188                                  enum transparent_hugepage_flag enabled,
 189                                  enum transparent_hugepage_flag req_madv)
 190 {
 191         if (!memcmp("always", buf,
 192                     min(sizeof("always")-1, count))) {
 193                 set_bit(enabled, &transparent_hugepage_flags);
 194                 clear_bit(req_madv, &transparent_hugepage_flags);
 195         } else if (!memcmp("madvise", buf,
 196                            min(sizeof("madvise")-1, count))) {
 197                 clear_bit(enabled, &transparent_hugepage_flags);
 198                 set_bit(req_madv, &transparent_hugepage_flags);
 199         } else if (!memcmp("never", buf,
 200                            min(sizeof("never")-1, count))) {
 201                 clear_bit(enabled, &transparent_hugepage_flags);
 202                 clear_bit(req_madv, &transparent_hugepage_flags);
 203         } else
 204                 return -EINVAL;
 205
 206         return count;
 207 }
 208
 209 static ssize_t enabled_show(struct kobject *kobj,
 210                             struct kobj_attribute *attr, char *buf)
 211 {
 212         return double_flag_show(kobj, attr, buf,
 213                                 TRANSPARENT_HUGEPAGE_FLAG,
 214                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 215 }
 216 static ssize_t enabled_store(struct kobject *kobj,
 217                              struct kobj_attribute *attr,
 218                              const char *buf, size_t count)
 219 {
 220         ssize_t ret;
 221
 222         ret = double_flag_store(kobj, attr, buf, count,
 223                                 TRANSPARENT_HUGEPAGE_FLAG,
 224                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 225
 226         if (ret > 0) {
 227                 int err = start_khugepaged();
 228                 if (err)
 229                         ret = err;
 230         }
 231
 232         if (ret > 0 &&
 233             (test_bit(TRANSPARENT_HUGEPAGE_FLAG,
 234                       &transparent_hugepage_flags) ||
 235              test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 236                       &transparent_hugepage_flags)))
 237                 set_recommended_min_free_kbytes();
 238
 239         return ret;
 240 }
 241 static struct kobj_attribute enabled_attr =
 242         __ATTR(enabled, 0644, enabled_show, enabled_store);
 243
 244 static ssize_t single_flag_show(struct kobject *kobj,
 245                                 struct kobj_attribute *attr, char *buf,
 246                                 enum transparent_hugepage_flag flag)
 247 {
 248         return sprintf(buf, "%d\n",
 249                        !!test_bit(flag, &transparent_hugepage_flags));
 250 }
 251
 252 static ssize_t single_flag_store(struct kobject *kobj,
 253                                  struct kobj_attribute *attr,
 254                                  const char *buf, size_t count,
 255                                  enum transparent_hugepage_flag flag)
 256 {
 257         unsigned long value;
 258         int ret;
 259
 260         ret = kstrtoul(buf, 10, &value);
 261         if (ret < 0)
 262                 return ret;
 263         if (value > 1)
 264                 return -EINVAL;
 265
 266         if (value)
 267                 set_bit(flag, &transparent_hugepage_flags);
 268         else
 269                 clear_bit(flag, &transparent_hugepage_flags);
 270
 271         return count;
 272 }
 273
 274 /*
 275  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 276  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 277  * memory just to allocate one more hugepage.
 278  */
 279 static ssize_t defrag_show(struct kobject *kobj,
 280                            struct kobj_attribute *attr, char *buf)
 281 {
 282         return double_flag_show(kobj, attr, buf,
 283                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 284                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 285 }
 286 static ssize_t defrag_store(struct kobject *kobj,
 287                             struct kobj_attribute *attr,
 288                             const char *buf, size_t count)
 289 {
 290         return double_flag_store(kobj, attr, buf, count,
 291                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 292                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 293 }
 294 static struct kobj_attribute defrag_attr =
 295         __ATTR(defrag, 0644, defrag_show, defrag_store);
 296
 297 #ifdef CONFIG_DEBUG_VM
 298 static ssize_t debug_cow_show(struct kobject *kobj,
 299                                 struct kobj_attribute *attr, char *buf)
 300 {
 301         return single_flag_show(kobj, attr, buf,
 302                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 303 }
 304 static ssize_t debug_cow_store(struct kobject *kobj,
 305                                struct kobj_attribute *attr,
 306                                const char *buf, size_t count)
 307 {
 308         return single_flag_store(kobj, attr, buf, count,
 309                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 310 }
 311 static struct kobj_attribute debug_cow_attr =
 312         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 313 #endif /* CONFIG_DEBUG_VM */
 314
 315 static struct attribute *hugepage_attr[] = {
 316         &enabled_attr.attr,
 317         &defrag_attr.attr,
 318 #ifdef CONFIG_DEBUG_VM
 319         &debug_cow_attr.attr,
 320 #endif
 321         NULL,
 322 };
 323
 324 static struct attribute_group hugepage_attr_group = {
 325         .attrs = hugepage_attr,
 326 };
 327
 328 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 329                                          struct kobj_attribute *attr,
 330                                          char *buf)
 331 {
 332         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 333 }
 334
 335 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 336                                           struct kobj_attribute *attr,
 337                                           const char *buf, size_t count)
 338 {
 339         unsigned long msecs;
 340         int err;
 341
 342         err = strict_strtoul(buf, 10, &msecs);
 343         if (err || msecs > UINT_MAX)
 344                 return -EINVAL;
 345
 346         khugepaged_scan_sleep_millisecs = msecs;
 347         wake_up_interruptible(&khugepaged_wait);
 348
 349         return count;
 350 }
 351 static struct kobj_attribute scan_sleep_millisecs_attr =
 352         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 353                scan_sleep_millisecs_store);
 354
 355 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 356                                           struct kobj_attribute *attr,
 357                                           char *buf)
 358 {
 359         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 360 }
 361
 362 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 363                                            struct kobj_attribute *attr,
 364                                            const char *buf, size_t count)
 365 {
 366         unsigned long msecs;
 367         int err;
 368
 369         err = strict_strtoul(buf, 10, &msecs);
 370         if (err || msecs > UINT_MAX)
 371                 return -EINVAL;
 372
 373         khugepaged_alloc_sleep_millisecs = msecs;
 374         wake_up_interruptible(&khugepaged_wait);
 375
 376         return count;
 377 }
 378 static struct kobj_attribute alloc_sleep_millisecs_attr =
 379         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 380                alloc_sleep_millisecs_store);
 381
 382 static ssize_t pages_to_scan_show(struct kobject *kobj,
 383                                   struct kobj_attribute *attr,
 384                                   char *buf)
 385 {
 386         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 387 }
 388 static ssize_t pages_to_scan_store(struct kobject *kobj,
 389                                    struct kobj_attribute *attr,
 390                                    const char *buf, size_t count)
 391 {
 392         int err;
 393         unsigned long pages;
 394
 395         err = strict_strtoul(buf, 10, &pages);
 396         if (err || !pages || pages > UINT_MAX)
 397                 return -EINVAL;
 398
 399         khugepaged_pages_to_scan = pages;
 400
 401         return count;
 402 }
 403 static struct kobj_attribute pages_to_scan_attr =
 404         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 405                pages_to_scan_store);
 406
 407 static ssize_t pages_collapsed_show(struct kobject *kobj,
 408                                     struct kobj_attribute *attr,
 409                                     char *buf)
 410 {
 411         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 412 }
 413 static struct kobj_attribute pages_collapsed_attr =
 414         __ATTR_RO(pages_collapsed);
 415
 416 static ssize_t full_scans_show(struct kobject *kobj,
 417                                struct kobj_attribute *attr,
 418                                char *buf)
 419 {
 420         return sprintf(buf, "%u\n", khugepaged_full_scans);
 421 }
 422 static struct kobj_attribute full_scans_attr =
 423         __ATTR_RO(full_scans);
 424
 425 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 426                                       struct kobj_attribute *attr, char *buf)
 427 {
 428         return single_flag_show(kobj, attr, buf,
 429                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 430 }
 431 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 432                                        struct kobj_attribute *attr,
 433                                        const char *buf, size_t count)
 434 {
 435         return single_flag_store(kobj, attr, buf, count,
 436                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 437 }
 438 static struct kobj_attribute khugepaged_defrag_attr =
 439         __ATTR(defrag, 0644, khugepaged_defrag_show,
 440                khugepaged_defrag_store);
 441
 442 /*
 443  * max_ptes_none controls if khugepaged should collapse hugepages over
 444  * any unmapped ptes in turn potentially increasing the memory
 445  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 446  * reduce the available free memory in the system as it
 447  * runs. Increasing max_ptes_none will instead potentially reduce the
 448  * free memory in the system during the khugepaged scan.
 449  */
 450 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 451                                              struct kobj_attribute *attr,
 452                                              char *buf)
 453 {
 454         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 455 }
 456 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 457                                               struct kobj_attribute *attr,
 458                                               const char *buf, size_t count)
 459 {
 460         int err;
 461         unsigned long max_ptes_none;
 462
 463         err = strict_strtoul(buf, 10, &max_ptes_none);
 464         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 465                 return -EINVAL;
 466
 467         khugepaged_max_ptes_none = max_ptes_none;
 468
 469         return count;
 470 }
 471 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 472         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 473                khugepaged_max_ptes_none_store);
 474
 475 static struct attribute *khugepaged_attr[] = {
 476         &khugepaged_defrag_attr.attr,
 477         &khugepaged_max_ptes_none_attr.attr,
 478         &pages_to_scan_attr.attr,
 479         &pages_collapsed_attr.attr,
 480         &full_scans_attr.attr,
 481         &scan_sleep_millisecs_attr.attr,
 482         &alloc_sleep_millisecs_attr.attr,
 483         NULL,
 484 };
 485
 486 static struct attribute_group khugepaged_attr_group = {
 487         .attrs = khugepaged_attr,
 488         .name = "khugepaged",
 489 };
 490 #endif /* CONFIG_SYSFS */
 491
 492 static int __init hugepage_init(void)
 493 {
 494         int err;
 495 #ifdef CONFIG_SYSFS
 496         static struct kobject *hugepage_kobj;
 497 #endif
 498
 499         err = -EINVAL;
 500         if (!has_transparent_hugepage()) {
 501                 transparent_hugepage_flags = 0;
 502                 goto out;
 503         }
 504
 505 #ifdef CONFIG_SYSFS
 506         err = -ENOMEM;
 507         hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 508         if (unlikely(!hugepage_kobj)) {
 509                 printk(KERN_ERR "hugepage: failed kobject create\n");
 510                 goto out;
 511         }
 512
 513         err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group);
 514         if (err) {
 515                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 516                 goto out;
 517         }
 518
 519         err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group);
 520         if (err) {
 521                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 522                 goto out;
 523         }
 524 #endif
 525
 526         err = khugepaged_slab_init();
 527         if (err)
 528                 goto out;
 529
 530         err = mm_slots_hash_init();
 531         if (err) {
 532                 khugepaged_slab_free();
 533                 goto out;
 534         }
 535
 536         /*
 537          * By default disable transparent hugepages on smaller systems,
 538          * where the extra memory used could hurt more than TLB overhead
 539          * is likely to save.  The admin can still enable it through /sys.
 540          */
 541         if (totalram_pages < (200 << (20 - PAGE_SHIFT)))
 542                 transparent_hugepage_flags = 0;
 543
 544         start_khugepaged();
 545
 546         set_recommended_min_free_kbytes();
 547
 548 out:
 549         return err;
 550 }
 551 module_init(hugepage_init)
 552
 553 static int __init setup_transparent_hugepage(char *str)
 554 {
 555         int ret = 0;
 556         if (!str)
 557                 goto out;
 558         if (!strcmp(str, "always")) {
 559                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 560                         &transparent_hugepage_flags);
 561                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 562                           &transparent_hugepage_flags);
 563                 ret = 1;
 564         } else if (!strcmp(str, "madvise")) {
 565                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 566                           &transparent_hugepage_flags);
 567                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 568                         &transparent_hugepage_flags);
 569                 ret = 1;
 570         } else if (!strcmp(str, "never")) {
 571                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 572                           &transparent_hugepage_flags);
 573                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 574                           &transparent_hugepage_flags);
 575                 ret = 1;
 576         }
 577 out:
 578         if (!ret)
 579                 printk(KERN_WARNING
 580                        "transparent_hugepage= cannot parse, ignored\n");
 581         return ret;
 582 }
 583 __setup("transparent_hugepage=", setup_transparent_hugepage);
 584
 585 static void prepare_pmd_huge_pte(pgtable_t pgtable,
 586                                  struct mm_struct *mm)
 587 {
 588         assert_spin_locked(&mm->page_table_lock);
 589
 590         /* FIFO */
 591         if (!mm->pmd_huge_pte)
 592                 INIT_LIST_HEAD(&pgtable->lru);
 593         else
 594                 list_add(&pgtable->lru, &mm->pmd_huge_pte->lru);
 595         mm->pmd_huge_pte = pgtable;
 596 }
 597
 598 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 599 {
 600         if (likely(vma->vm_flags & VM_WRITE))
 601                 pmd = pmd_mkwrite(pmd);
 602         return pmd;
 603 }
 604
 605 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 606                                         struct vm_area_struct *vma,
 607                                         unsigned long haddr, pmd_t *pmd,
 608                                         struct page *page)
 609 {
 610         int ret = 0;
 611         pgtable_t pgtable;
 612
 613         VM_BUG_ON(!PageCompound(page));
 614         pgtable = pte_alloc_one(mm, haddr);
 615         if (unlikely(!pgtable)) {
 616                 mem_cgroup_uncharge_page(page);
 617                 put_page(page);
 618                 return VM_FAULT_OOM;
 619         }
 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                 prepare_pmd_huge_pte(pgtable, mm);
 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 ret;
 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
 699                 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page);
 700         }
 701 out:
 702         /*
 703          * Use __pte_alloc instead of pte_alloc_map, because we can't
 704          * run pte_offset_map on the pmd, if an huge pmd could
 705          * materialize from under us from a different thread.
 706          */
 707         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
 708                 return VM_FAULT_OOM;
 709         /* if an huge pmd materialized from under us just retry later */
 710         if (unlikely(pmd_trans_huge(*pmd)))
 711                 return 0;
 712         /*
 713          * A regular pmd is established and it can't morph into a huge pmd
 714          * from under us anymore at this point because we hold the mmap_sem
 715          * read mode and khugepaged takes it in write mode. So now it's
 716          * safe to run pte_offset_map().
 717          */
 718         pte = pte_offset_map(pmd, address);
 719         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 720 }
 721
 722 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 723                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 724                   struct vm_area_struct *vma)
 725 {
 726         struct page *src_page;
 727         pmd_t pmd;
 728         pgtable_t pgtable;
 729         int ret;
 730
 731         ret = -ENOMEM;
 732         pgtable = pte_alloc_one(dst_mm, addr);
 733         if (unlikely(!pgtable))
 734                 goto out;
 735
 736         spin_lock(&dst_mm->page_table_lock);
 737         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 738
 739         ret = -EAGAIN;
 740         pmd = *src_pmd;
 741         if (unlikely(!pmd_trans_huge(pmd))) {
 742                 pte_free(dst_mm, pgtable);
 743                 goto out_unlock;
 744         }
 745         if (unlikely(pmd_trans_splitting(pmd))) {
 746                 /* split huge page running from under us */
 747                 spin_unlock(&src_mm->page_table_lock);
 748                 spin_unlock(&dst_mm->page_table_lock);
 749                 pte_free(dst_mm, pgtable);
 750
 751                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 752                 goto out;
 753         }
 754         src_page = pmd_page(pmd);
 755         VM_BUG_ON(!PageHead(src_page));
 756         get_page(src_page);
 757         page_dup_rmap(src_page);
 758         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 759
 760         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 761         pmd = pmd_mkold(pmd_wrprotect(pmd));
 762         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 763         prepare_pmd_huge_pte(pgtable, dst_mm);
 764         dst_mm->nr_ptes++;
 765
 766         ret = 0;
 767 out_unlock:
 768         spin_unlock(&src_mm->page_table_lock);
 769         spin_unlock(&dst_mm->page_table_lock);
 770 out:
 771         return ret;
 772 }
 773
 774 /* no "address" argument so destroys page coloring of some arch */
 775 pgtable_t get_pmd_huge_pte(struct mm_struct *mm)
 776 {
 777         pgtable_t pgtable;
 778
 779         assert_spin_locked(&mm->page_table_lock);
 780
 781         /* FIFO */
 782         pgtable = mm->pmd_huge_pte;
 783         if (list_empty(&pgtable->lru))
 784                 mm->pmd_huge_pte = NULL;
 785         else {
 786                 mm->pmd_huge_pte = list_entry(pgtable->lru.next,
 787                                               struct page, lru);
 788                 list_del(&pgtable->lru);
 789         }
 790         return pgtable;
 791 }
 792
 793 void huge_pmd_set_accessed(struct mm_struct *mm,
 794                            struct vm_area_struct *vma,
 795                            unsigned long address,
 796                            pmd_t *pmd, pmd_t orig_pmd,
 797                            int dirty)
 798 {
 799         pmd_t entry;
 800         unsigned long haddr;
 801
 802         spin_lock(&mm->page_table_lock);
 803         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 804                 goto unlock;
 805
 806         entry = pmd_mkyoung(orig_pmd);
 807         haddr = address & HPAGE_PMD_MASK;
 808         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 809                 update_mmu_cache_pmd(vma, address, pmd);
 810
 811 unlock:
 812         spin_unlock(&mm->page_table_lock);
 813 }
 814
 815 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 816                                         struct vm_area_struct *vma,
 817                                         unsigned long address,
 818                                         pmd_t *pmd, pmd_t orig_pmd,
 819                                         struct page *page,
 820                                         unsigned long haddr)
 821 {
 822         pgtable_t pgtable;
 823         pmd_t _pmd;
 824         int ret = 0, i;
 825         struct page **pages;
 826
 827         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 828                         GFP_KERNEL);
 829         if (unlikely(!pages)) {
 830                 ret |= VM_FAULT_OOM;
 831                 goto out;
 832         }
 833
 834         for (i = 0; i < HPAGE_PMD_NR; i++) {
 835                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 836                                                __GFP_OTHER_NODE,
 837                                                vma, address, page_to_nid(page));
 838                 if (unlikely(!pages[i] ||
 839                              mem_cgroup_newpage_charge(pages[i], mm,
 840                                                        GFP_KERNEL))) {
 841                         if (pages[i])
 842                                 put_page(pages[i]);
 843                         mem_cgroup_uncharge_start();
 844                         while (--i >= 0) {
 845                                 mem_cgroup_uncharge_page(pages[i]);
 846                                 put_page(pages[i]);
 847                         }
 848                         mem_cgroup_uncharge_end();
 849                         kfree(pages);
 850                         ret |= VM_FAULT_OOM;
 851                         goto out;
 852                 }
 853         }
 854
 855         for (i = 0; i < HPAGE_PMD_NR; i++) {
 856                 copy_user_highpage(pages[i], page + i,
 857                                    haddr + PAGE_SIZE * i, vma);
 858                 __SetPageUptodate(pages[i]);
 859                 cond_resched();
 860         }
 861
 862         spin_lock(&mm->page_table_lock);
 863         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 864                 goto out_free_pages;
 865         VM_BUG_ON(!PageHead(page));
 866
 867         pmdp_clear_flush_notify(vma, haddr, pmd);
 868         /* leave pmd empty until pte is filled */
 869
 870         pgtable = get_pmd_huge_pte(mm);
 871         pmd_populate(mm, &_pmd, pgtable);
 872
 873         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 874                 pte_t *pte, entry;
 875                 entry = mk_pte(pages[i], vma->vm_page_prot);
 876                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 877                 page_add_new_anon_rmap(pages[i], vma, haddr);
 878                 pte = pte_offset_map(&_pmd, haddr);
 879                 VM_BUG_ON(!pte_none(*pte));
 880                 set_pte_at(mm, haddr, pte, entry);
 881                 pte_unmap(pte);
 882         }
 883         kfree(pages);
 884
 885         smp_wmb(); /* make pte visible before pmd */
 886         pmd_populate(mm, pmd, pgtable);
 887         page_remove_rmap(page);
 888         spin_unlock(&mm->page_table_lock);
 889
 890         ret |= VM_FAULT_WRITE;
 891         put_page(page);
 892
 893 out:
 894         return ret;
 895
 896 out_free_pages:
 897         spin_unlock(&mm->page_table_lock);
 898         mem_cgroup_uncharge_start();
 899         for (i = 0; i < HPAGE_PMD_NR; i++) {
 900                 mem_cgroup_uncharge_page(pages[i]);
 901                 put_page(pages[i]);
 902         }
 903         mem_cgroup_uncharge_end();
 904         kfree(pages);
 905         goto out;
 906 }
 907
 908 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
 909                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
 910 {
 911         int ret = 0;
 912         struct page *page, *new_page;
 913         unsigned long haddr;
 914
 915         VM_BUG_ON(!vma->anon_vma);
 916         spin_lock(&mm->page_table_lock);
 917         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 918                 goto out_unlock;
 919
 920         page = pmd_page(orig_pmd);
 921         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
 922         haddr = address & HPAGE_PMD_MASK;
 923         if (page_mapcount(page) == 1) {
 924                 pmd_t entry;
 925                 entry = pmd_mkyoung(orig_pmd);
 926                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 927                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
 928                         update_mmu_cache(vma, address, pmd);
 929                 ret |= VM_FAULT_WRITE;
 930                 goto out_unlock;
 931         }
 932         get_page(page);
 933         spin_unlock(&mm->page_table_lock);
 934
 935         if (transparent_hugepage_enabled(vma) &&
 936             !transparent_hugepage_debug_cow())
 937                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 938                                               vma, haddr, numa_node_id(), 0);
 939         else
 940                 new_page = NULL;
 941
 942         if (unlikely(!new_page)) {
 943                 count_vm_event(THP_FAULT_FALLBACK);
 944                 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
 945                                                    pmd, orig_pmd, page, haddr);
 946                 if (ret & VM_FAULT_OOM)
 947                         split_huge_page(page);
 948                 put_page(page);
 949                 goto out;
 950         }
 951         count_vm_event(THP_FAULT_ALLOC);
 952
 953         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
 954                 put_page(new_page);
 955                 split_huge_page(page);
 956                 put_page(page);
 957                 ret |= VM_FAULT_OOM;
 958                 goto out;
 959         }
 960
 961         copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 962         __SetPageUptodate(new_page);
 963
 964         spin_lock(&mm->page_table_lock);
 965         put_page(page);
 966         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
 967                 mem_cgroup_uncharge_page(new_page);
 968                 put_page(new_page);
 969         } else {
 970                 pmd_t entry;
 971                 VM_BUG_ON(!PageHead(page));
 972                 entry = mk_pmd(new_page, vma->vm_page_prot);
 973                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 974                 entry = pmd_mkhuge(entry);
 975                 pmdp_clear_flush_notify(vma, haddr, pmd);
 976                 page_add_new_anon_rmap(new_page, vma, haddr);
 977                 set_pmd_at(mm, haddr, pmd, entry);
 978                 update_mmu_cache(vma, address, pmd);
 979                 page_remove_rmap(page);
 980                 put_page(page);
 981                 ret |= VM_FAULT_WRITE;
 982         }
 983 out_unlock:
 984         spin_unlock(&mm->page_table_lock);
 985 out:
 986         return ret;
 987 }
 988
 989 struct page *follow_trans_huge_pmd(struct mm_struct *mm,
 990                                    unsigned long addr,
 991                                    pmd_t *pmd,
 992                                    unsigned int flags)
 993 {
 994         struct page *page = NULL;
 995
 996         assert_spin_locked(&mm->page_table_lock);
 997
 998         if (flags & FOLL_WRITE && !pmd_write(*pmd))
 999                 goto out;
1000
1001         page = pmd_page(*pmd);
1002         VM_BUG_ON(!PageHead(page));
1003         if (flags & FOLL_TOUCH) {
1004                 pmd_t _pmd;
1005                 /*
1006                  * We should set the dirty bit only for FOLL_WRITE but
1007                  * for now the dirty bit in the pmd is meaningless.
1008                  * And if the dirty bit will become meaningful and
1009                  * we'll only set it with FOLL_WRITE, an atomic
1010                  * set_bit will be required on the pmd to set the
1011                  * young bit, instead of the current set_pmd_at.
1012                  */
1013                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1014                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1015         }
1016         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1017         VM_BUG_ON(!PageCompound(page));
1018         if (flags & FOLL_GET)
1019                 get_page_foll(page);
1020
1021 out:
1022         return page;
1023 }
1024
1025 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1026                  pmd_t *pmd)
1027 {
1028         int ret = 0;
1029
1030         spin_lock(&tlb->mm->page_table_lock);
1031         if (likely(pmd_trans_huge(*pmd))) {
1032                 if (unlikely(pmd_trans_splitting(*pmd))) {
1033                         spin_unlock(&tlb->mm->page_table_lock);
1034                         wait_split_huge_page(vma->anon_vma,
1035                                              pmd);
1036                 } else {
1037                         struct page *page;
1038                         pgtable_t pgtable;
1039                         pgtable = get_pmd_huge_pte(tlb->mm);
1040                         page = pmd_page(*pmd);
1041                         pmd_clear(pmd);
1042                         page_remove_rmap(page);
1043                         VM_BUG_ON(page_mapcount(page) < 0);
1044                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1045                         VM_BUG_ON(!PageHead(page));
1046                         tlb->mm->nr_ptes--;
1047                         spin_unlock(&tlb->mm->page_table_lock);
1048                         tlb_remove_page(tlb, page);
1049                         pte_free(tlb->mm, pgtable);
1050                         ret = 1;
1051                 }
1052         } else
1053                 spin_unlock(&tlb->mm->page_table_lock);
1054
1055         return ret;
1056 }
1057
1058 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1059                 unsigned long addr, unsigned long end,
1060                 unsigned char *vec)
1061 {
1062         int ret = 0;
1063
1064         spin_lock(&vma->vm_mm->page_table_lock);
1065         if (likely(pmd_trans_huge(*pmd))) {
1066                 ret = !pmd_trans_splitting(*pmd);
1067                 spin_unlock(&vma->vm_mm->page_table_lock);
1068                 if (unlikely(!ret))
1069                         wait_split_huge_page(vma->anon_vma, pmd);
1070                 else {
1071                         /*
1072                          * All logical pages in the range are present
1073                          * if backed by a huge page.
1074                          */
1075                         memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1076                 }
1077         } else
1078                 spin_unlock(&vma->vm_mm->page_table_lock);
1079
1080         return ret;
1081 }
1082
1083 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1084                   unsigned long old_addr,
1085                   unsigned long new_addr, unsigned long old_end,
1086                   pmd_t *old_pmd, pmd_t *new_pmd)
1087 {
1088         int ret = 0;
1089         pmd_t pmd;
1090
1091         struct mm_struct *mm = vma->vm_mm;
1092
1093         if ((old_addr & ~HPAGE_PMD_MASK) ||
1094             (new_addr & ~HPAGE_PMD_MASK) ||
1095             old_end - old_addr < HPAGE_PMD_SIZE ||
1096             (new_vma->vm_flags & VM_NOHUGEPAGE))
1097                 goto out;
1098
1099         /*
1100          * The destination pmd shouldn't be established, free_pgtables()
1101          * should have release it.
1102          */
1103         if (WARN_ON(!pmd_none(*new_pmd))) {
1104                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1105                 goto out;
1106         }
1107
1108         spin_lock(&mm->page_table_lock);
1109         if (likely(pmd_trans_huge(*old_pmd))) {
1110                 if (pmd_trans_splitting(*old_pmd)) {
1111                         spin_unlock(&mm->page_table_lock);
1112                         wait_split_huge_page(vma->anon_vma, old_pmd);
1113                         ret = -1;
1114                 } else {
1115                         pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1116                         VM_BUG_ON(!pmd_none(*new_pmd));
1117                         set_pmd_at(mm, new_addr, new_pmd, pmd);
1118                         spin_unlock(&mm->page_table_lock);
1119                         ret = 1;
1120                 }
1121         } else {
1122                 spin_unlock(&mm->page_table_lock);
1123         }
1124 out:
1125         return ret;
1126 }
1127
1128 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1129                 unsigned long addr, pgprot_t newprot)
1130 {
1131         struct mm_struct *mm = vma->vm_mm;
1132         int ret = 0;
1133
1134         spin_lock(&mm->page_table_lock);
1135         if (likely(pmd_trans_huge(*pmd))) {
1136                 if (unlikely(pmd_trans_splitting(*pmd))) {
1137                         spin_unlock(&mm->page_table_lock);
1138                         wait_split_huge_page(vma->anon_vma, pmd);
1139                 } else {
1140                         pmd_t entry;
1141
1142                         entry = pmdp_get_and_clear(mm, addr, pmd);
1143                         entry = pmd_modify(entry, newprot);
1144                         set_pmd_at(mm, addr, pmd, entry);
1145                         spin_unlock(&vma->vm_mm->page_table_lock);
1146                         flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
1147                         ret = 1;
1148                 }
1149         } else
1150                 spin_unlock(&vma->vm_mm->page_table_lock);
1151
1152         return ret;
1153 }
1154
1155 pmd_t *page_check_address_pmd(struct page *page,
1156                               struct mm_struct *mm,
1157                               unsigned long address,
1158                               enum page_check_address_pmd_flag flag)
1159 {
1160         pgd_t *pgd;
1161         pud_t *pud;
1162         pmd_t *pmd, *ret = NULL;
1163
1164         if (address & ~HPAGE_PMD_MASK)
1165                 goto out;
1166
1167         pgd = pgd_offset(mm, address);
1168         if (!pgd_present(*pgd))
1169                 goto out;
1170
1171         pud = pud_offset(pgd, address);
1172         if (!pud_present(*pud))
1173                 goto out;
1174
1175         pmd = pmd_offset(pud, address);
1176         if (pmd_none(*pmd))
1177                 goto out;
1178         if (pmd_page(*pmd) != page)
1179                 goto out;
1180         /*
1181          * split_vma() may create temporary aliased mappings. There is
1182          * no risk as long as all huge pmd are found and have their
1183          * splitting bit set before __split_huge_page_refcount
1184          * runs. Finding the same huge pmd more than once during the
1185          * same rmap walk is not a problem.
1186          */
1187         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1188             pmd_trans_splitting(*pmd))
1189                 goto out;
1190         if (pmd_trans_huge(*pmd)) {
1191                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1192                           !pmd_trans_splitting(*pmd));
1193                 ret = pmd;
1194         }
1195 out:
1196         return ret;
1197 }
1198
1199 static int __split_huge_page_splitting(struct page *page,
1200                                        struct vm_area_struct *vma,
1201                                        unsigned long address)
1202 {
1203         struct mm_struct *mm = vma->vm_mm;
1204         pmd_t *pmd;
1205         int ret = 0;
1206
1207         spin_lock(&mm->page_table_lock);
1208         pmd = page_check_address_pmd(page, mm, address,
1209                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1210         if (pmd) {
1211                 /*
1212                  * We can't temporarily set the pmd to null in order
1213                  * to split it, the pmd must remain marked huge at all
1214                  * times or the VM won't take the pmd_trans_huge paths
1215                  * and it won't wait on the anon_vma->root->mutex to
1216                  * serialize against split_huge_page*.
1217                  */
1218                 pmdp_splitting_flush_notify(vma, address, pmd);
1219                 ret = 1;
1220         }
1221         spin_unlock(&mm->page_table_lock);
1222
1223         return ret;
1224 }
1225
1226 static void __split_huge_page_refcount(struct page *page)
1227 {
1228         int i;
1229         unsigned long head_index = page->index;
1230         struct zone *zone = page_zone(page);
1231         int zonestat;
1232         int tail_count = 0;
1233
1234         /* prevent PageLRU to go away from under us, and freeze lru stats */
1235         spin_lock_irq(&zone->lru_lock);
1236         compound_lock(page);
1237
1238         for (i = 1; i < HPAGE_PMD_NR; i++) {
1239                 struct page *page_tail = page + i;
1240
1241                 /* tail_page->_mapcount cannot change */
1242                 BUG_ON(page_mapcount(page_tail) < 0);
1243                 tail_count += page_mapcount(page_tail);
1244                 /* check for overflow */
1245                 BUG_ON(tail_count < 0);
1246                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1247                 /*
1248                  * tail_page->_count is zero and not changing from
1249                  * under us. But get_page_unless_zero() may be running
1250                  * from under us on the tail_page. If we used
1251                  * atomic_set() below instead of atomic_add(), we
1252                  * would then run atomic_set() concurrently with
1253                  * get_page_unless_zero(), and atomic_set() is
1254                  * implemented in C not using locked ops. spin_unlock
1255                  * on x86 sometime uses locked ops because of PPro
1256                  * errata 66, 92, so unless somebody can guarantee
1257                  * atomic_set() here would be safe on all archs (and
1258                  * not only on x86), it's safer to use atomic_add().
1259                  */
1260                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1261                            &page_tail->_count);
1262
1263                 /* after clearing PageTail the gup refcount can be released */
1264                 smp_mb();
1265
1266                 /*
1267                  * retain hwpoison flag of the poisoned tail page:
1268                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1269                  *   by the memory-failure.
1270                  */
1271                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1272                 page_tail->flags |= (page->flags &
1273                                      ((1L << PG_referenced) |
1274                                       (1L << PG_swapbacked) |
1275                                       (1L << PG_mlocked) |
1276                                       (1L << PG_uptodate)));
1277                 page_tail->flags |= (1L << PG_dirty);
1278
1279                 /* clear PageTail before overwriting first_page */
1280                 smp_wmb();
1281
1282                 /*
1283                  * __split_huge_page_splitting() already set the
1284                  * splitting bit in all pmd that could map this
1285                  * hugepage, that will ensure no CPU can alter the
1286                  * mapcount on the head page. The mapcount is only
1287                  * accounted in the head page and it has to be
1288                  * transferred to all tail pages in the below code. So
1289                  * for this code to be safe, the split the mapcount
1290                  * can't change. But that doesn't mean userland can't
1291                  * keep changing and reading the page contents while
1292                  * we transfer the mapcount, so the pmd splitting
1293                  * status is achieved setting a reserved bit in the
1294                  * pmd, not by clearing the present bit.
1295                 */
1296                 page_tail->_mapcount = page->_mapcount;
1297
1298                 BUG_ON(page_tail->mapping);
1299                 page_tail->mapping = page->mapping;
1300
1301                 page_tail->index = ++head_index;
1302
1303                 BUG_ON(!PageAnon(page_tail));
1304                 BUG_ON(!PageUptodate(page_tail));
1305                 BUG_ON(!PageDirty(page_tail));
1306                 BUG_ON(!PageSwapBacked(page_tail));
1307
1308                 mem_cgroup_split_huge_fixup(page, page_tail);
1309
1310                 lru_add_page_tail(zone, page, page_tail);
1311         }
1312         atomic_sub(tail_count, &page->_count);
1313         BUG_ON(atomic_read(&page->_count) <= 0);
1314
1315         __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1316         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1317
1318         /*
1319          * A hugepage counts for HPAGE_PMD_NR pages on the LRU statistics,
1320          * so adjust those appropriately if this page is on the LRU.
1321          */
1322         if (PageLRU(page)) {
1323                 zonestat = NR_LRU_BASE + page_lru(page);
1324                 __mod_zone_page_state(zone, zonestat, -(HPAGE_PMD_NR-1));
1325         }
1326
1327         ClearPageCompound(page);
1328         compound_unlock(page);
1329         spin_unlock_irq(&zone->lru_lock);
1330
1331         for (i = 1; i < HPAGE_PMD_NR; i++) {
1332                 struct page *page_tail = page + i;
1333                 BUG_ON(page_count(page_tail) <= 0);
1334                 /*
1335                  * Tail pages may be freed if there wasn't any mapping
1336                  * like if add_to_swap() is running on a lru page that
1337                  * had its mapping zapped. And freeing these pages
1338                  * requires taking the lru_lock so we do the put_page
1339                  * of the tail pages after the split is complete.
1340                  */
1341                 put_page(page_tail);
1342         }
1343
1344         /*
1345          * Only the head page (now become a regular page) is required
1346          * to be pinned by the caller.
1347          */
1348         BUG_ON(page_count(page) <= 0);
1349 }
1350
1351 static int __split_huge_page_map(struct page *page,
1352                                  struct vm_area_struct *vma,
1353                                  unsigned long address)
1354 {
1355         struct mm_struct *mm = vma->vm_mm;
1356         pmd_t *pmd, _pmd;
1357         int ret = 0, i;
1358         pgtable_t pgtable;
1359         unsigned long haddr;
1360
1361         spin_lock(&mm->page_table_lock);
1362         pmd = page_check_address_pmd(page, mm, address,
1363                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1364         if (pmd) {
1365                 pgtable = get_pmd_huge_pte(mm);
1366                 pmd_populate(mm, &_pmd, pgtable);
1367
1368                 for (i = 0, haddr = address; i < HPAGE_PMD_NR;
1369                      i++, haddr += PAGE_SIZE) {
1370                         pte_t *pte, entry;
1371                         BUG_ON(PageCompound(page+i));
1372                         entry = mk_pte(page + i, vma->vm_page_prot);
1373                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1374                         if (!pmd_write(*pmd))
1375                                 entry = pte_wrprotect(entry);
1376                         else
1377                                 BUG_ON(page_mapcount(page) != 1);
1378                         if (!pmd_young(*pmd))
1379                                 entry = pte_mkold(entry);
1380                         pte = pte_offset_map(&_pmd, haddr);
1381                         BUG_ON(!pte_none(*pte));
1382                         set_pte_at(mm, haddr, pte, entry);
1383                         pte_unmap(pte);
1384                 }
1385
1386                 smp_wmb(); /* make pte visible before pmd */
1387                 /*
1388                  * Up to this point the pmd is present and huge and
1389                  * userland has the whole access to the hugepage
1390                  * during the split (which happens in place). If we
1391                  * overwrite the pmd with the not-huge version
1392                  * pointing to the pte here (which of course we could
1393                  * if all CPUs were bug free), userland could trigger
1394                  * a small page size TLB miss on the small sized TLB
1395                  * while the hugepage TLB entry is still established
1396                  * in the huge TLB. Some CPU doesn't like that. See
1397                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1398                  * Erratum 383 on page 93. Intel should be safe but is
1399                  * also warns that it's only safe if the permission
1400                  * and cache attributes of the two entries loaded in
1401                  * the two TLB is identical (which should be the case
1402                  * here). But it is generally safer to never allow
1403                  * small and huge TLB entries for the same virtual
1404                  * address to be loaded simultaneously. So instead of
1405                  * doing "pmd_populate(); flush_tlb_range();" we first
1406                  * mark the current pmd notpresent (atomically because
1407                  * here the pmd_trans_huge and pmd_trans_splitting
1408                  * must remain set at all times on the pmd until the
1409                  * split is complete for this pmd), then we flush the
1410                  * SMP TLB and finally we write the non-huge version
1411                  * of the pmd entry with pmd_populate.
1412                  */
1413                 set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd));
1414                 flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
1415                 pmd_populate(mm, pmd, pgtable);
1416                 ret = 1;
1417         }
1418         spin_unlock(&mm->page_table_lock);
1419
1420         return ret;
1421 }
1422
1423 /* must be called with anon_vma->root->mutex hold */
1424 static void __split_huge_page(struct page *page,
1425                               struct anon_vma *anon_vma)
1426 {
1427         int mapcount, mapcount2;
1428         struct anon_vma_chain *avc;
1429
1430         BUG_ON(!PageHead(page));
1431         BUG_ON(PageTail(page));
1432
1433         mapcount = 0;
1434         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1435                 struct vm_area_struct *vma = avc->vma;
1436                 unsigned long addr = vma_address(page, vma);
1437                 BUG_ON(is_vma_temporary_stack(vma));
1438                 if (addr == -EFAULT)
1439                         continue;
1440                 mapcount += __split_huge_page_splitting(page, vma, addr);
1441         }
1442         /*
1443          * It is critical that new vmas are added to the tail of the
1444          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1445          * and establishes a child pmd before
1446          * __split_huge_page_splitting() freezes the parent pmd (so if
1447          * we fail to prevent copy_huge_pmd() from running until the
1448          * whole __split_huge_page() is complete), we will still see
1449          * the newly established pmd of the child later during the
1450          * walk, to be able to set it as pmd_trans_splitting too.
1451          */
1452         if (mapcount != page_mapcount(page))
1453                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1454                        mapcount, page_mapcount(page));
1455         BUG_ON(mapcount != page_mapcount(page));
1456
1457         __split_huge_page_refcount(page);
1458
1459         mapcount2 = 0;
1460         list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1461                 struct vm_area_struct *vma = avc->vma;
1462                 unsigned long addr = vma_address(page, vma);
1463                 BUG_ON(is_vma_temporary_stack(vma));
1464                 if (addr == -EFAULT)
1465                         continue;
1466                 mapcount2 += __split_huge_page_map(page, vma, addr);
1467         }
1468         if (mapcount != mapcount2)
1469                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1470                        mapcount, mapcount2, page_mapcount(page));
1471         BUG_ON(mapcount != mapcount2);
1472 }
1473
1474 int split_huge_page(struct page *page)
1475 {
1476         struct anon_vma *anon_vma;
1477         int ret = 1;
1478
1479         BUG_ON(!PageAnon(page));
1480         anon_vma = page_lock_anon_vma(page);
1481         if (!anon_vma)
1482                 goto out;
1483         ret = 0;
1484         if (!PageCompound(page))
1485                 goto out_unlock;
1486
1487         BUG_ON(!PageSwapBacked(page));
1488         __split_huge_page(page, anon_vma);
1489         count_vm_event(THP_SPLIT);
1490
1491         BUG_ON(PageCompound(page));
1492 out_unlock:
1493         page_unlock_anon_vma(anon_vma);
1494 out:
1495         return ret;
1496 }
1497
1498 #define VM_NO_THP (VM_SPECIAL|VM_INSERTPAGE|VM_MIXEDMAP|VM_SAO| \
1499                    VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1500
1501 int hugepage_madvise(struct vm_area_struct *vma,
1502                      unsigned long *vm_flags, int advice)
1503 {
1504         switch (advice) {
1505         case MADV_HUGEPAGE:
1506                 /*
1507                  * Be somewhat over-protective like KSM for now!
1508                  */
1509                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1510                         return -EINVAL;
1511                 *vm_flags &= ~VM_NOHUGEPAGE;
1512                 *vm_flags |= VM_HUGEPAGE;
1513                 /*
1514                  * If the vma become good for khugepaged to scan,
1515                  * register it here without waiting a page fault that
1516                  * may not happen any time soon.
1517                  */
1518                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1519                         return -ENOMEM;
1520                 break;
1521         case MADV_NOHUGEPAGE:
1522                 /*
1523                  * Be somewhat over-protective like KSM for now!
1524                  */
1525                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1526                         return -EINVAL;
1527                 *vm_flags &= ~VM_HUGEPAGE;
1528                 *vm_flags |= VM_NOHUGEPAGE;
1529                 /*
1530                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1531                  * this vma even if we leave the mm registered in khugepaged if
1532                  * it got registered before VM_NOHUGEPAGE was set.
1533                  */
1534                 break;
1535         }
1536
1537         return 0;
1538 }
1539
1540 static int __init khugepaged_slab_init(void)
1541 {
1542         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1543                                           sizeof(struct mm_slot),
1544                                           __alignof__(struct mm_slot), 0, NULL);
1545         if (!mm_slot_cache)
1546                 return -ENOMEM;
1547
1548         return 0;
1549 }
1550
1551 static void __init khugepaged_slab_free(void)
1552 {
1553         kmem_cache_destroy(mm_slot_cache);
1554         mm_slot_cache = NULL;
1555 }
1556
1557 static inline struct mm_slot *alloc_mm_slot(void)
1558 {
1559         if (!mm_slot_cache)     /* initialization failed */
1560                 return NULL;
1561         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1562 }
1563
1564 static inline void free_mm_slot(struct mm_slot *mm_slot)
1565 {
1566         kmem_cache_free(mm_slot_cache, mm_slot);
1567 }
1568
1569 static int __init mm_slots_hash_init(void)
1570 {
1571         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1572                                 GFP_KERNEL);
1573         if (!mm_slots_hash)
1574                 return -ENOMEM;
1575         return 0;
1576 }
1577
1578 #if 0
1579 static void __init mm_slots_hash_free(void)
1580 {
1581         kfree(mm_slots_hash);
1582         mm_slots_hash = NULL;
1583 }
1584 #endif
1585
1586 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1587 {
1588         struct mm_slot *mm_slot;
1589         struct hlist_head *bucket;
1590         struct hlist_node *node;
1591
1592         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1593                                 % MM_SLOTS_HASH_HEADS];
1594         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1595                 if (mm == mm_slot->mm)
1596                         return mm_slot;
1597         }
1598         return NULL;
1599 }
1600
1601 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1602                                     struct mm_slot *mm_slot)
1603 {
1604         struct hlist_head *bucket;
1605
1606         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1607                                 % MM_SLOTS_HASH_HEADS];
1608         mm_slot->mm = mm;
1609         hlist_add_head(&mm_slot->hash, bucket);
1610 }
1611
1612 static inline int khugepaged_test_exit(struct mm_struct *mm)
1613 {
1614         return atomic_read(&mm->mm_users) == 0;
1615 }
1616
1617 int __khugepaged_enter(struct mm_struct *mm)
1618 {
1619         struct mm_slot *mm_slot;
1620         int wakeup;
1621
1622         mm_slot = alloc_mm_slot();
1623         if (!mm_slot)
1624                 return -ENOMEM;
1625
1626         /* __khugepaged_exit() must not run from under us */
1627         VM_BUG_ON(khugepaged_test_exit(mm));
1628         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1629                 free_mm_slot(mm_slot);
1630                 return 0;
1631         }
1632
1633         spin_lock(&khugepaged_mm_lock);
1634         insert_to_mm_slots_hash(mm, mm_slot);
1635         /*
1636          * Insert just behind the scanning cursor, to let the area settle
1637          * down a little.
1638          */
1639         wakeup = list_empty(&khugepaged_scan.mm_head);
1640         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1641         spin_unlock(&khugepaged_mm_lock);
1642
1643         atomic_inc(&mm->mm_count);
1644         if (wakeup)
1645                 wake_up_interruptible(&khugepaged_wait);
1646
1647         return 0;
1648 }
1649
1650 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1651 {
1652         unsigned long hstart, hend;
1653         if (!vma->anon_vma)
1654                 /*
1655                  * Not yet faulted in so we will register later in the
1656                  * page fault if needed.
1657                  */
1658                 return 0;
1659         if (vma->vm_ops)
1660                 /* khugepaged not yet working on file or special mappings */
1661                 return 0;
1662         /*
1663          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1664          * true too, verify it here.
1665          */
1666         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1667         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1668         hend = vma->vm_end & HPAGE_PMD_MASK;
1669         if (hstart < hend)
1670                 return khugepaged_enter(vma);
1671         return 0;
1672 }
1673
1674 void __khugepaged_exit(struct mm_struct *mm)
1675 {
1676         struct mm_slot *mm_slot;
1677         int free = 0;
1678
1679         spin_lock(&khugepaged_mm_lock);
1680         mm_slot = get_mm_slot(mm);
1681         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1682                 hlist_del(&mm_slot->hash);
1683                 list_del(&mm_slot->mm_node);
1684                 free = 1;
1685         }
1686         spin_unlock(&khugepaged_mm_lock);
1687
1688         if (free) {
1689                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1690                 free_mm_slot(mm_slot);
1691                 mmdrop(mm);
1692         } else if (mm_slot) {
1693                 /*
1694                  * This is required to serialize against
1695                  * khugepaged_test_exit() (which is guaranteed to run
1696                  * under mmap sem read mode). Stop here (after we
1697                  * return all pagetables will be destroyed) until
1698                  * khugepaged has finished working on the pagetables
1699                  * under the mmap_sem.
1700                  */
1701                 down_write(&mm->mmap_sem);
1702                 up_write(&mm->mmap_sem);
1703         }
1704 }
1705
1706 static void release_pte_page(struct page *page)
1707 {
1708         /* 0 stands for page_is_file_cache(page) == false */
1709         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1710         unlock_page(page);
1711         putback_lru_page(page);
1712 }
1713
1714 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1715 {
1716         while (--_pte >= pte) {
1717                 pte_t pteval = *_pte;
1718                 if (!pte_none(pteval))
1719                         release_pte_page(pte_page(pteval));
1720         }
1721 }
1722
1723 static void release_all_pte_pages(pte_t *pte)
1724 {
1725         release_pte_pages(pte, pte + HPAGE_PMD_NR);
1726 }
1727
1728 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1729                                         unsigned long address,
1730                                         pte_t *pte)
1731 {
1732         struct page *page;
1733         pte_t *_pte;
1734         int referenced = 0, isolated = 0, none = 0;
1735         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1736              _pte++, address += PAGE_SIZE) {
1737                 pte_t pteval = *_pte;
1738                 if (pte_none(pteval)) {
1739                         if (++none <= khugepaged_max_ptes_none)
1740                                 continue;
1741                         else {
1742                                 release_pte_pages(pte, _pte);
1743                                 goto out;
1744                         }
1745                 }
1746                 if (!pte_present(pteval) || !pte_write(pteval)) {
1747                         release_pte_pages(pte, _pte);
1748                         goto out;
1749                 }
1750                 page = vm_normal_page(vma, address, pteval);
1751                 if (unlikely(!page)) {
1752                         release_pte_pages(pte, _pte);
1753                         goto out;
1754                 }
1755                 VM_BUG_ON(PageCompound(page));
1756                 BUG_ON(!PageAnon(page));
1757                 VM_BUG_ON(!PageSwapBacked(page));
1758
1759                 /* cannot use mapcount: can't collapse if there's a gup pin */
1760                 if (page_count(page) != 1) {
1761                         release_pte_pages(pte, _pte);
1762                         goto out;
1763                 }
1764                 /*
1765                  * We can do it before isolate_lru_page because the
1766                  * page can't be freed from under us. NOTE: PG_lock
1767                  * is needed to serialize against split_huge_page
1768                  * when invoked from the VM.
1769                  */
1770                 if (!trylock_page(page)) {
1771                         release_pte_pages(pte, _pte);
1772                         goto out;
1773                 }
1774                 /*
1775                  * Isolate the page to avoid collapsing an hugepage
1776                  * currently in use by the VM.
1777                  */
1778                 if (isolate_lru_page(page)) {
1779                         unlock_page(page);
1780                         release_pte_pages(pte, _pte);
1781                         goto out;
1782                 }
1783                 /* 0 stands for page_is_file_cache(page) == false */
1784                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1785                 VM_BUG_ON(!PageLocked(page));
1786                 VM_BUG_ON(PageLRU(page));
1787
1788                 /* If there is no mapped pte young don't collapse the page */
1789                 if (pte_young(pteval) || PageReferenced(page) ||
1790                     mmu_notifier_test_young(vma->vm_mm, address))
1791                         referenced = 1;
1792         }
1793         if (unlikely(!referenced))
1794                 release_all_pte_pages(pte);
1795         else
1796                 isolated = 1;
1797 out:
1798         return isolated;
1799 }
1800
1801 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1802                                       struct vm_area_struct *vma,
1803                                       unsigned long address,
1804                                       spinlock_t *ptl)
1805 {
1806         pte_t *_pte;
1807         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1808                 pte_t pteval = *_pte;
1809                 struct page *src_page;
1810
1811                 if (pte_none(pteval)) {
1812                         clear_user_highpage(page, address);
1813                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1814                 } else {
1815                         src_page = pte_page(pteval);
1816                         copy_user_highpage(page, src_page, address, vma);
1817                         VM_BUG_ON(page_mapcount(src_page) != 1);
1818                         VM_BUG_ON(page_count(src_page) != 2);
1819                         release_pte_page(src_page);
1820                         /*
1821                          * ptl mostly unnecessary, but preempt has to
1822                          * be disabled to update the per-cpu stats
1823                          * inside page_remove_rmap().
1824                          */
1825                         spin_lock(ptl);
1826                         /*
1827                          * paravirt calls inside pte_clear here are
1828                          * superfluous.
1829                          */
1830                         pte_clear(vma->vm_mm, address, _pte);
1831                         page_remove_rmap(src_page);
1832                         spin_unlock(ptl);
1833                         free_page_and_swap_cache(src_page);
1834                 }
1835
1836                 address += PAGE_SIZE;
1837                 page++;
1838         }
1839 }
1840
1841 static void collapse_huge_page(struct mm_struct *mm,
1842                                unsigned long address,
1843                                struct page **hpage,
1844                                struct vm_area_struct *vma,
1845                                int node)
1846 {
1847         pgd_t *pgd;
1848         pud_t *pud;
1849         pmd_t *pmd, _pmd;
1850         pte_t *pte;
1851         pgtable_t pgtable;
1852         struct page *new_page;
1853         spinlock_t *ptl;
1854         int isolated;
1855         unsigned long hstart, hend;
1856
1857         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1858 #ifndef CONFIG_NUMA
1859         up_read(&mm->mmap_sem);
1860         VM_BUG_ON(!*hpage);
1861         new_page = *hpage;
1862 #else
1863         VM_BUG_ON(*hpage);
1864         /*
1865          * Allocate the page while the vma is still valid and under
1866          * the mmap_sem read mode so there is no memory allocation
1867          * later when we take the mmap_sem in write mode. This is more
1868          * friendly behavior (OTOH it may actually hide bugs) to
1869          * filesystems in userland with daemons allocating memory in
1870          * the userland I/O paths.  Allocating memory with the
1871          * mmap_sem in read mode is good idea also to allow greater
1872          * scalability.
1873          */
1874         new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
1875                                       node, __GFP_OTHER_NODE);
1876
1877         /*
1878          * After allocating the hugepage, release the mmap_sem read lock in
1879          * preparation for taking it in write mode.
1880          */
1881         up_read(&mm->mmap_sem);
1882         if (unlikely(!new_page)) {
1883                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1884                 *hpage = ERR_PTR(-ENOMEM);
1885                 return;
1886         }
1887 #endif
1888
1889         count_vm_event(THP_COLLAPSE_ALLOC);
1890         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1891 #ifdef CONFIG_NUMA
1892                 put_page(new_page);
1893 #endif
1894                 return;
1895         }
1896
1897         /*
1898          * Prevent all access to pagetables with the exception of
1899          * gup_fast later hanlded by the ptep_clear_flush and the VM
1900          * handled by the anon_vma lock + PG_lock.
1901          */
1902         down_write(&mm->mmap_sem);
1903         if (unlikely(khugepaged_test_exit(mm)))
1904                 goto out;
1905
1906         vma = find_vma(mm, address);
1907         if (!vma)
1908                 goto out;
1909         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1910         hend = vma->vm_end & HPAGE_PMD_MASK;
1911         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
1912                 goto out;
1913
1914         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
1915             (vma->vm_flags & VM_NOHUGEPAGE))
1916                 goto out;
1917
1918         if (!vma->anon_vma || vma->vm_ops)
1919                 goto out;
1920         if (is_vma_temporary_stack(vma))
1921                 goto out;
1922         /*
1923          * If is_pfn_mapping() is true is_learn_pfn_mapping() must be
1924          * true too, verify it here.
1925          */
1926         VM_BUG_ON(is_linear_pfn_mapping(vma) || vma->vm_flags & VM_NO_THP);
1927
1928         pgd = pgd_offset(mm, address);
1929         if (!pgd_present(*pgd))
1930                 goto out;
1931
1932         pud = pud_offset(pgd, address);
1933         if (!pud_present(*pud))
1934                 goto out;
1935
1936         pmd = pmd_offset(pud, address);
1937         /* pmd can't go away or become huge under us */
1938         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
1939                 goto out;
1940
1941         anon_vma_lock(vma->anon_vma);
1942
1943         pte = pte_offset_map(pmd, address);
1944         ptl = pte_lockptr(mm, pmd);
1945
1946         spin_lock(&mm->page_table_lock); /* probably unnecessary */
1947         /*
1948          * After this gup_fast can't run anymore. This also removes
1949          * any huge TLB entry from the CPU so we won't allow
1950          * huge and small TLB entries for the same virtual address
1951          * to avoid the risk of CPU bugs in that area.
1952          */
1953         _pmd = pmdp_clear_flush_notify(vma, address, pmd);
1954         spin_unlock(&mm->page_table_lock);
1955
1956         spin_lock(ptl);
1957         isolated = __collapse_huge_page_isolate(vma, address, pte);
1958         spin_unlock(ptl);
1959
1960         if (unlikely(!isolated)) {
1961                 pte_unmap(pte);
1962                 spin_lock(&mm->page_table_lock);
1963                 BUG_ON(!pmd_none(*pmd));
1964                 /*
1965                  * We can only use set_pmd_at when establishing
1966                  * hugepmds and never for establishing regular pmds that
1967                  * points to regular pagetables. Use pmd_populate for that
1968                  */
1969                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1970                 spin_unlock(&mm->page_table_lock);
1971                 anon_vma_unlock(vma->anon_vma);
1972                 goto out;
1973         }
1974
1975         /*
1976          * All pages are isolated and locked so anon_vma rmap
1977          * can't run anymore.
1978          */
1979         anon_vma_unlock(vma->anon_vma);
1980
1981         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
1982         pte_unmap(pte);
1983         __SetPageUptodate(new_page);
1984         pgtable = pmd_pgtable(_pmd);
1985         VM_BUG_ON(page_count(pgtable) != 1);
1986         VM_BUG_ON(page_mapcount(pgtable) != 0);
1987
1988         _pmd = mk_pmd(new_page, vma->vm_page_prot);
1989         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1990         _pmd = pmd_mkhuge(_pmd);
1991
1992         /*
1993          * spin_lock() below is not the equivalent of smp_wmb(), so
1994          * this is needed to avoid the copy_huge_page writes to become
1995          * visible after the set_pmd_at() write.
1996          */
1997         smp_wmb();
1998
1999         spin_lock(&mm->page_table_lock);
2000         BUG_ON(!pmd_none(*pmd));
2001         page_add_new_anon_rmap(new_page, vma, address);
2002         set_pmd_at(mm, address, pmd, _pmd);
2003         update_mmu_cache(vma, address, pmd);
2004         prepare_pmd_huge_pte(pgtable, mm);
2005         spin_unlock(&mm->page_table_lock);
2006
2007 #ifndef CONFIG_NUMA
2008         *hpage = NULL;
2009 #endif
2010         khugepaged_pages_collapsed++;
2011 out_up_write:
2012         up_write(&mm->mmap_sem);
2013         return;
2014
2015 out:
2016         mem_cgroup_uncharge_page(new_page);
2017 #ifdef CONFIG_NUMA
2018         put_page(new_page);
2019 #endif
2020         goto out_up_write;
2021 }
2022
2023 static int khugepaged_scan_pmd(struct mm_struct *mm,
2024                                struct vm_area_struct *vma,
2025                                unsigned long address,
2026                                struct page **hpage)
2027 {
2028         pgd_t *pgd;
2029         pud_t *pud;
2030         pmd_t *pmd;
2031         pte_t *pte, *_pte;
2032         int ret = 0, referenced = 0, none = 0;
2033         struct page *page;
2034         unsigned long _address;
2035         spinlock_t *ptl;
2036         int node = -1;
2037
2038         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2039
2040         pgd = pgd_offset(mm, address);
2041         if (!pgd_present(*pgd))
2042                 goto out;
2043
2044         pud = pud_offset(pgd, address);
2045         if (!pud_present(*pud))
2046                 goto out;
2047
2048         pmd = pmd_offset(pud, address);
2049         if (!pmd_present(*pmd) || pmd_trans_huge(*pmd))
2050                 goto out;
2051
2052         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2053         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2054              _pte++, _address += PAGE_SIZE) {
2055                 pte_t pteval = *_pte;
2056                 if (pte_none(pteval)) {
2057                         if (++none <= khugepaged_max_ptes_none)
2058                                 continue;
2059                         else
2060                                 goto out_unmap;
2061                 }
2062                 if (!pte_present(pteval) || !pte_write(pteval))
2063                         goto out_unmap;
2064                 page = vm_normal_page(vma, _address, pteval);
2065                 if (unlikely(!page))
2066                         goto out_unmap;
2067                 /*
2068                  * Chose the node of the first page. This could
2069                  * be more sophisticated and look at more pages,
2070                  * but isn't for now.
2071                  */
2072                 if (node == -1)
2073                         node = page_to_nid(page);
2074                 VM_BUG_ON(PageCompound(page));
2075                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2076                         goto out_unmap;
2077                 /* cannot use mapcount: can't collapse if there's a gup pin */
2078                 if (page_count(page) != 1)
2079                         goto out_unmap;
2080                 if (pte_young(pteval) || PageReferenced(page) ||
2081                     mmu_notifier_test_young(vma->vm_mm, address))
2082                         referenced = 1;
2083         }
2084         if (referenced)
2085                 ret = 1;
2086 out_unmap:
2087         pte_unmap_unlock(pte, ptl);
2088         if (ret)
2089                 /* collapse_huge_page will return with the mmap_sem released */
2090                 collapse_huge_page(mm, address, hpage, vma, node);
2091 out:
2092         return ret;
2093 }
2094
2095 static void collect_mm_slot(struct mm_slot *mm_slot)
2096 {
2097         struct mm_struct *mm = mm_slot->mm;
2098
2099         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2100
2101         if (khugepaged_test_exit(mm)) {
2102                 /* free mm_slot */
2103                 hlist_del(&mm_slot->hash);
2104                 list_del(&mm_slot->mm_node);
2105
2106                 /*
2107                  * Not strictly needed because the mm exited already.
2108                  *
2109                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2110                  */
2111
2112                 /* khugepaged_mm_lock actually not necessary for the below */
2113                 free_mm_slot(mm_slot);
2114                 mmdrop(mm);
2115         }
2116 }
2117
2118 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2119                                             struct page **hpage)
2120         __releases(&khugepaged_mm_lock)
2121         __acquires(&khugepaged_mm_lock)
2122 {
2123         struct mm_slot *mm_slot;
2124         struct mm_struct *mm;
2125         struct vm_area_struct *vma;
2126         int progress = 0;
2127
2128         VM_BUG_ON(!pages);
2129         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2130
2131         if (khugepaged_scan.mm_slot)
2132                 mm_slot = khugepaged_scan.mm_slot;
2133         else {
2134                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2135                                      struct mm_slot, mm_node);
2136                 khugepaged_scan.address = 0;
2137                 khugepaged_scan.mm_slot = mm_slot;
2138         }
2139         spin_unlock(&khugepaged_mm_lock);
2140
2141         mm = mm_slot->mm;
2142         down_read(&mm->mmap_sem);
2143         if (unlikely(khugepaged_test_exit(mm)))
2144                 vma = NULL;
2145         else
2146                 vma = find_vma(mm, khugepaged_scan.address);
2147
2148         progress++;
2149         for (; vma; vma = vma->vm_next) {
2150                 unsigned long hstart, hend;
2151
2152                 cond_resched();
2153                 if (unlikely(khugepaged_test_exit(mm))) {
2154                         progress++;
2155                         break;
2156                 }
2157
2158                 if ((!(vma->vm_flags & VM_HUGEPAGE) &&
2159                      !khugepaged_always()) ||
2160                     (vma->vm_flags & VM_NOHUGEPAGE)) {
2161                 skip:
2162                         progress++;
2163                         continue;
2164                 }
2165                 if (!vma->anon_vma || vma->vm_ops)
2166                         goto skip;
2167                 if (is_vma_temporary_stack(vma))
2168                         goto skip;
2169                 /*
2170                  * If is_pfn_mapping() is true is_learn_pfn_mapping()
2171                  * must be true too, verify it here.
2172                  */
2173                 VM_BUG_ON(is_linear_pfn_mapping(vma) ||
2174                           vma->vm_flags & VM_NO_THP);
2175
2176                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2177                 hend = vma->vm_end & HPAGE_PMD_MASK;
2178                 if (hstart >= hend)
2179                         goto skip;
2180                 if (khugepaged_scan.address > hend)
2181                         goto skip;
2182                 if (khugepaged_scan.address < hstart)
2183                         khugepaged_scan.address = hstart;
2184                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2185
2186                 while (khugepaged_scan.address < hend) {
2187                         int ret;
2188                         cond_resched();
2189                         if (unlikely(khugepaged_test_exit(mm)))
2190                                 goto breakouterloop;
2191
2192                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2193                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2194                                   hend);
2195                         ret = khugepaged_scan_pmd(mm, vma,
2196                                                   khugepaged_scan.address,
2197                                                   hpage);
2198                         /* move to next address */
2199                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2200                         progress += HPAGE_PMD_NR;
2201                         if (ret)
2202                                 /* we released mmap_sem so break loop */
2203                                 goto breakouterloop_mmap_sem;
2204                         if (progress >= pages)
2205                                 goto breakouterloop;
2206                 }
2207         }
2208 breakouterloop:
2209         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2210 breakouterloop_mmap_sem:
2211
2212         spin_lock(&khugepaged_mm_lock);
2213         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2214         /*
2215          * Release the current mm_slot if this mm is about to die, or
2216          * if we scanned all vmas of this mm.
2217          */
2218         if (khugepaged_test_exit(mm) || !vma) {
2219                 /*
2220                  * Make sure that if mm_users is reaching zero while
2221                  * khugepaged runs here, khugepaged_exit will find
2222                  * mm_slot not pointing to the exiting mm.
2223                  */
2224                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2225                         khugepaged_scan.mm_slot = list_entry(
2226                                 mm_slot->mm_node.next,
2227                                 struct mm_slot, mm_node);
2228                         khugepaged_scan.address = 0;
2229                 } else {
2230                         khugepaged_scan.mm_slot = NULL;
2231                         khugepaged_full_scans++;
2232                 }
2233
2234                 collect_mm_slot(mm_slot);
2235         }
2236
2237         return progress;
2238 }
2239
2240 static int khugepaged_has_work(void)
2241 {
2242         return !list_empty(&khugepaged_scan.mm_head) &&
2243                 khugepaged_enabled();
2244 }
2245
2246 static int khugepaged_wait_event(void)
2247 {
2248         return !list_empty(&khugepaged_scan.mm_head) ||
2249                 !khugepaged_enabled();
2250 }
2251
2252 static void khugepaged_do_scan(struct page **hpage)
2253 {
2254         unsigned int progress = 0, pass_through_head = 0;
2255         unsigned int pages = khugepaged_pages_to_scan;
2256
2257         barrier(); /* write khugepaged_pages_to_scan to local stack */
2258
2259         while (progress < pages) {
2260                 cond_resched();
2261
2262 #ifndef CONFIG_NUMA
2263                 if (!*hpage) {
2264                         *hpage = alloc_hugepage(khugepaged_defrag());
2265                         if (unlikely(!*hpage)) {
2266                                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2267                                 break;
2268                         }
2269                         count_vm_event(THP_COLLAPSE_ALLOC);
2270                 }
2271 #else
2272                 if (IS_ERR(*hpage))
2273                         break;
2274 #endif
2275
2276                 if (unlikely(kthread_should_stop() || freezing(current)))
2277                         break;
2278
2279                 spin_lock(&khugepaged_mm_lock);
2280                 if (!khugepaged_scan.mm_slot)
2281                         pass_through_head++;
2282                 if (khugepaged_has_work() &&
2283                     pass_through_head < 2)
2284                         progress += khugepaged_scan_mm_slot(pages - progress,
2285                                                             hpage);
2286                 else
2287                         progress = pages;
2288                 spin_unlock(&khugepaged_mm_lock);
2289         }
2290 }
2291
2292 static void khugepaged_alloc_sleep(void)
2293 {
2294         wait_event_freezable_timeout(khugepaged_wait, false,
2295                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2296 }
2297
2298 #ifndef CONFIG_NUMA
2299 static struct page *khugepaged_alloc_hugepage(void)
2300 {
2301         struct page *hpage;
2302
2303         do {
2304                 hpage = alloc_hugepage(khugepaged_defrag());
2305                 if (!hpage) {
2306                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2307                         khugepaged_alloc_sleep();
2308                 } else
2309                         count_vm_event(THP_COLLAPSE_ALLOC);
2310         } while (unlikely(!hpage) &&
2311                  likely(khugepaged_enabled()));
2312         return hpage;
2313 }
2314 #endif
2315
2316 static void khugepaged_loop(void)
2317 {
2318         struct page *hpage;
2319
2320 #ifdef CONFIG_NUMA
2321         hpage = NULL;
2322 #endif
2323         while (likely(khugepaged_enabled())) {
2324 #ifndef CONFIG_NUMA
2325                 hpage = khugepaged_alloc_hugepage();
2326                 if (unlikely(!hpage))
2327                         break;
2328 #else
2329                 if (IS_ERR(hpage)) {
2330                         khugepaged_alloc_sleep();
2331                         hpage = NULL;
2332                 }
2333 #endif
2334
2335                 khugepaged_do_scan(&hpage);
2336 #ifndef CONFIG_NUMA
2337                 if (hpage)
2338                         put_page(hpage);
2339 #endif
2340                 try_to_freeze();
2341                 if (unlikely(kthread_should_stop()))
2342                         break;
2343                 if (khugepaged_has_work()) {
2344                         if (!khugepaged_scan_sleep_millisecs)
2345                                 continue;
2346                         wait_event_freezable_timeout(khugepaged_wait, false,
2347                             msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2348                 } else if (khugepaged_enabled())
2349                         wait_event_freezable(khugepaged_wait,
2350                                              khugepaged_wait_event());
2351         }
2352 }
2353
2354 static int khugepaged(void *none)
2355 {
2356         struct mm_slot *mm_slot;
2357
2358         set_freezable();
2359         set_user_nice(current, 19);
2360
2361         /* serialize with start_khugepaged() */
2362         mutex_lock(&khugepaged_mutex);
2363
2364         for (;;) {
2365                 mutex_unlock(&khugepaged_mutex);
2366                 VM_BUG_ON(khugepaged_thread != current);
2367                 khugepaged_loop();
2368                 VM_BUG_ON(khugepaged_thread != current);
2369
2370                 mutex_lock(&khugepaged_mutex);
2371                 if (!khugepaged_enabled())
2372                         break;
2373                 if (unlikely(kthread_should_stop()))
2374                         break;
2375         }
2376
2377         spin_lock(&khugepaged_mm_lock);
2378         mm_slot = khugepaged_scan.mm_slot;
2379         khugepaged_scan.mm_slot = NULL;
2380         if (mm_slot)
2381                 collect_mm_slot(mm_slot);
2382         spin_unlock(&khugepaged_mm_lock);
2383
2384         khugepaged_thread = NULL;
2385         mutex_unlock(&khugepaged_mutex);
2386
2387         return 0;
2388 }
2389
2390 void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd)
2391 {
2392         struct page *page;
2393
2394         spin_lock(&mm->page_table_lock);
2395         if (unlikely(!pmd_trans_huge(*pmd))) {
2396                 spin_unlock(&mm->page_table_lock);
2397                 return;
2398         }
2399         page = pmd_page(*pmd);
2400         VM_BUG_ON(!page_count(page));
2401         get_page(page);
2402         spin_unlock(&mm->page_table_lock);
2403
2404         split_huge_page(page);
2405
2406         put_page(page);
2407         BUG_ON(pmd_trans_huge(*pmd));
2408 }
2409
2410 static void split_huge_page_address(struct mm_struct *mm,
2411                                     unsigned long address)
2412 {
2413         pgd_t *pgd;
2414         pud_t *pud;
2415         pmd_t *pmd;
2416
2417         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2418
2419         pgd = pgd_offset(mm, address);
2420         if (!pgd_present(*pgd))
2421                 return;
2422
2423         pud = pud_offset(pgd, address);
2424         if (!pud_present(*pud))
2425                 return;
2426
2427         pmd = pmd_offset(pud, address);
2428         if (!pmd_present(*pmd))
2429                 return;
2430         /*
2431          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2432          * materialize from under us.
2433          */
2434         split_huge_page_pmd(mm, pmd);
2435 }
2436
2437 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2438                              unsigned long start,
2439                              unsigned long end,
2440                              long adjust_next)
2441 {
2442         /*
2443          * If the new start address isn't hpage aligned and it could
2444          * previously contain an hugepage: check if we need to split
2445          * an huge pmd.
2446          */
2447         if (start & ~HPAGE_PMD_MASK &&
2448             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2449             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2450                 split_huge_page_address(vma->vm_mm, start);
2451
2452         /*
2453          * If the new end address isn't hpage aligned and it could
2454          * previously contain an hugepage: check if we need to split
2455          * an huge pmd.
2456          */
2457         if (end & ~HPAGE_PMD_MASK &&
2458             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2459             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2460                 split_huge_page_address(vma->vm_mm, end);
2461
2462         /*
2463          * If we're also updating the vma->vm_next->vm_start, if the new
2464          * vm_next->vm_start isn't page aligned and it could previously
2465          * contain an hugepage: check if we need to split an huge pmd.
2466          */
2467         if (adjust_next > 0) {
2468                 struct vm_area_struct *next = vma->vm_next;
2469                 unsigned long nstart = next->vm_start;
2470                 nstart += adjust_next << PAGE_SHIFT;
2471                 if (nstart & ~HPAGE_PMD_MASK &&
2472                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2473                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2474                         split_huge_page_address(next->vm_mm, nstart);
2475         }
2476 }