fs/btrfs/extent_map.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/gfp.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/page-flags.h>
   8 #include <linux/module.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/swap.h>
  12 #include <linux/version.h>
  13 #include <linux/writeback.h>
  14 #include <linux/pagevec.h>
  15 #include "extent_map.h"
  16
  17 /* temporary define until extent_map moves out of btrfs */
  18 struct kmem_cache *btrfs_cache_create(const char *name, size_t size,
  19                                        unsigned long extra_flags,
  20                                        void (*ctor)(void *, struct kmem_cache *,
  21                                                     unsigned long));
  22
  23 static struct kmem_cache *extent_map_cache;
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26
  27 static LIST_HEAD(buffers);
  28 static LIST_HEAD(states);
  29
  30 static spinlock_t state_lock = SPIN_LOCK_UNLOCKED;
  31 #define BUFFER_LRU_MAX 64
  32
  33 struct tree_entry {
  34         u64 start;
  35         u64 end;
  36         int in_tree;
  37         struct rb_node rb_node;
  38 };
  39
  40 struct extent_page_data {
  41         struct bio *bio;
  42         struct extent_map_tree *tree;
  43         get_extent_t *get_extent;
  44 };
  45 int __init extent_map_init(void)
  46 {
  47         extent_map_cache = btrfs_cache_create("extent_map",
  48                                             sizeof(struct extent_map), 0,
  49                                             NULL);
  50         if (!extent_map_cache)
  51                 return -ENOMEM;
  52         extent_state_cache = btrfs_cache_create("extent_state",
  53                                             sizeof(struct extent_state), 0,
  54                                             NULL);
  55         if (!extent_state_cache)
  56                 goto free_map_cache;
  57         extent_buffer_cache = btrfs_cache_create("extent_buffers",
  58                                             sizeof(struct extent_buffer), 0,
  59                                             NULL);
  60         if (!extent_buffer_cache)
  61                 goto free_state_cache;
  62         return 0;
  63
  64 free_state_cache:
  65         kmem_cache_destroy(extent_state_cache);
  66 free_map_cache:
  67         kmem_cache_destroy(extent_map_cache);
  68         return -ENOMEM;
  69 }
  70
  71 void __exit extent_map_exit(void)
  72 {
  73         struct extent_state *state;
  74
  75         while (!list_empty(&states)) {
  76                 state = list_entry(states.next, struct extent_state, list);
  77                 printk("state leak: start %Lu end %Lu state %lu in tree %d refs %d\n", state->start, state->end, state->state, state->in_tree, atomic_read(&state->refs));
  78                 list_del(&state->list);
  79                 kmem_cache_free(extent_state_cache, state);
  80
  81         }
  82
  83         if (extent_map_cache)
  84                 kmem_cache_destroy(extent_map_cache);
  85         if (extent_state_cache)
  86                 kmem_cache_destroy(extent_state_cache);
  87         if (extent_buffer_cache)
  88                 kmem_cache_destroy(extent_buffer_cache);
  89 }
  90
  91 void extent_map_tree_init(struct extent_map_tree *tree,
  92                           struct address_space *mapping, gfp_t mask)
  93 {
  94         tree->map.rb_node = NULL;
  95         tree->state.rb_node = NULL;
  96         tree->ops = NULL;
  97         rwlock_init(&tree->lock);
  98         spin_lock_init(&tree->lru_lock);
  99         tree->mapping = mapping;
 100         INIT_LIST_HEAD(&tree->buffer_lru);
 101         tree->lru_size = 0;
 102 }
 103 EXPORT_SYMBOL(extent_map_tree_init);
 104
 105 void extent_map_tree_empty_lru(struct extent_map_tree *tree)
 106 {
 107         struct extent_buffer *eb;
 108         while(!list_empty(&tree->buffer_lru)) {
 109                 eb = list_entry(tree->buffer_lru.next, struct extent_buffer,
 110                                 lru);
 111                 list_del_init(&eb->lru);
 112                 free_extent_buffer(eb);
 113         }
 114 }
 115 EXPORT_SYMBOL(extent_map_tree_empty_lru);
 116
 117 struct extent_map *alloc_extent_map(gfp_t mask)
 118 {
 119         struct extent_map *em;
 120         em = kmem_cache_alloc(extent_map_cache, mask);
 121         if (!em || IS_ERR(em))
 122                 return em;
 123         em->in_tree = 0;
 124         atomic_set(&em->refs, 1);
 125         return em;
 126 }
 127 EXPORT_SYMBOL(alloc_extent_map);
 128
 129 void free_extent_map(struct extent_map *em)
 130 {
 131         if (!em)
 132                 return;
 133         if (atomic_dec_and_test(&em->refs)) {
 134                 WARN_ON(em->in_tree);
 135                 kmem_cache_free(extent_map_cache, em);
 136         }
 137 }
 138 EXPORT_SYMBOL(free_extent_map);
 139
 140
 141 struct extent_state *alloc_extent_state(gfp_t mask)
 142 {
 143         struct extent_state *state;
 144         unsigned long flags;
 145
 146         state = kmem_cache_alloc(extent_state_cache, mask);
 147         if (!state || IS_ERR(state))
 148                 return state;
 149         state->state = 0;
 150         state->in_tree = 0;
 151         state->private = 0;
 152
 153         spin_lock_irqsave(&state_lock, flags);
 154         list_add(&state->list, &states);
 155         spin_unlock_irqrestore(&state_lock, flags);
 156
 157         atomic_set(&state->refs, 1);
 158         init_waitqueue_head(&state->wq);
 159         return state;
 160 }
 161 EXPORT_SYMBOL(alloc_extent_state);
 162
 163 void free_extent_state(struct extent_state *state)
 164 {
 165         unsigned long flags;
 166         if (!state)
 167                 return;
 168         if (atomic_dec_and_test(&state->refs)) {
 169                 WARN_ON(state->in_tree);
 170                 spin_lock_irqsave(&state_lock, flags);
 171                 list_del(&state->list);
 172                 spin_unlock_irqrestore(&state_lock, flags);
 173                 kmem_cache_free(extent_state_cache, state);
 174         }
 175 }
 176 EXPORT_SYMBOL(free_extent_state);
 177
 178 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 179                                    struct rb_node *node)
 180 {
 181         struct rb_node ** p = &root->rb_node;
 182         struct rb_node * parent = NULL;
 183         struct tree_entry *entry;
 184
 185         while(*p) {
 186                 parent = *p;
 187                 entry = rb_entry(parent, struct tree_entry, rb_node);
 188
 189                 if (offset < entry->start)
 190                         p = &(*p)->rb_left;
 191                 else if (offset > entry->end)
 192                         p = &(*p)->rb_right;
 193                 else
 194                         return parent;
 195         }
 196
 197         entry = rb_entry(node, struct tree_entry, rb_node);
 198         entry->in_tree = 1;
 199         rb_link_node(node, parent, p);
 200         rb_insert_color(node, root);
 201         return NULL;
 202 }
 203
 204 static struct rb_node *__tree_search(struct rb_root *root, u64 offset,
 205                                    struct rb_node **prev_ret)
 206 {
 207         struct rb_node * n = root->rb_node;
 208         struct rb_node *prev = NULL;
 209         struct tree_entry *entry;
 210         struct tree_entry *prev_entry = NULL;
 211
 212         while(n) {
 213                 entry = rb_entry(n, struct tree_entry, rb_node);
 214                 prev = n;
 215                 prev_entry = entry;
 216
 217                 if (offset < entry->start)
 218                         n = n->rb_left;
 219                 else if (offset > entry->end)
 220                         n = n->rb_right;
 221                 else
 222                         return n;
 223         }
 224         if (!prev_ret)
 225                 return NULL;
 226         while(prev && offset > prev_entry->end) {
 227                 prev = rb_next(prev);
 228                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 229         }
 230         *prev_ret = prev;
 231         return NULL;
 232 }
 233
 234 static inline struct rb_node *tree_search(struct rb_root *root, u64 offset)
 235 {
 236         struct rb_node *prev;
 237         struct rb_node *ret;
 238         ret = __tree_search(root, offset, &prev);
 239         if (!ret)
 240                 return prev;
 241         return ret;
 242 }
 243
 244 static int tree_delete(struct rb_root *root, u64 offset)
 245 {
 246         struct rb_node *node;
 247         struct tree_entry *entry;
 248
 249         node = __tree_search(root, offset, NULL);
 250         if (!node)
 251                 return -ENOENT;
 252         entry = rb_entry(node, struct tree_entry, rb_node);
 253         entry->in_tree = 0;
 254         rb_erase(node, root);
 255         return 0;
 256 }
 257
 258 /*
 259  * add_extent_mapping tries a simple backward merge with existing
 260  * mappings.  The extent_map struct passed in will be inserted into
 261  * the tree directly (no copies made, just a reference taken).
 262  */
 263 int add_extent_mapping(struct extent_map_tree *tree,
 264                        struct extent_map *em)
 265 {
 266         int ret = 0;
 267         struct extent_map *prev = NULL;
 268         struct rb_node *rb;
 269
 270         write_lock_irq(&tree->lock);
 271         rb = tree_insert(&tree->map, em->end, &em->rb_node);
 272         if (rb) {
 273                 prev = rb_entry(rb, struct extent_map, rb_node);
 274                 printk("found extent map %Lu %Lu on insert of %Lu %Lu\n", prev->start, prev->end, em->start, em->end);
 275                 ret = -EEXIST;
 276                 goto out;
 277         }
 278         atomic_inc(&em->refs);
 279         if (em->start != 0) {
 280                 rb = rb_prev(&em->rb_node);
 281                 if (rb)
 282                         prev = rb_entry(rb, struct extent_map, rb_node);
 283                 if (prev && prev->end + 1 == em->start &&
 284                     ((em->block_start == EXTENT_MAP_HOLE &&
 285                       prev->block_start == EXTENT_MAP_HOLE) ||
 286                      (em->block_start == EXTENT_MAP_INLINE &&
 287                       prev->block_start == EXTENT_MAP_INLINE) ||
 288                      (em->block_start == EXTENT_MAP_DELALLOC &&
 289                       prev->block_start == EXTENT_MAP_DELALLOC) ||
 290                      (em->block_start < EXTENT_MAP_DELALLOC - 1 &&
 291                       em->block_start == prev->block_end + 1))) {
 292                         em->start = prev->start;
 293                         em->block_start = prev->block_start;
 294                         rb_erase(&prev->rb_node, &tree->map);
 295                         prev->in_tree = 0;
 296                         free_extent_map(prev);
 297                 }
 298          }
 299 out:
 300         write_unlock_irq(&tree->lock);
 301         return ret;
 302 }
 303 EXPORT_SYMBOL(add_extent_mapping);
 304
 305 /*
 306  * lookup_extent_mapping returns the first extent_map struct in the
 307  * tree that intersects the [start, end] (inclusive) range.  There may
 308  * be additional objects in the tree that intersect, so check the object
 309  * returned carefully to make sure you don't need additional lookups.
 310  */
 311 struct extent_map *lookup_extent_mapping(struct extent_map_tree *tree,
 312                                          u64 start, u64 end)
 313 {
 314         struct extent_map *em;
 315         struct rb_node *rb_node;
 316
 317         read_lock_irq(&tree->lock);
 318         rb_node = tree_search(&tree->map, start);
 319         if (!rb_node) {
 320                 em = NULL;
 321                 goto out;
 322         }
 323         if (IS_ERR(rb_node)) {
 324                 em = ERR_PTR(PTR_ERR(rb_node));
 325                 goto out;
 326         }
 327         em = rb_entry(rb_node, struct extent_map, rb_node);
 328         if (em->end < start || em->start > end) {
 329                 em = NULL;
 330                 goto out;
 331         }
 332         atomic_inc(&em->refs);
 333 out:
 334         read_unlock_irq(&tree->lock);
 335         return em;
 336 }
 337 EXPORT_SYMBOL(lookup_extent_mapping);
 338
 339 /*
 340  * removes an extent_map struct from the tree.  No reference counts are
 341  * dropped, and no checks are done to  see if the range is in use
 342  */
 343 int remove_extent_mapping(struct extent_map_tree *tree, struct extent_map *em)
 344 {
 345         int ret;
 346
 347         write_lock_irq(&tree->lock);
 348         ret = tree_delete(&tree->map, em->end);
 349         write_unlock_irq(&tree->lock);
 350         return ret;
 351 }
 352 EXPORT_SYMBOL(remove_extent_mapping);
 353
 354 /*
 355  * utility function to look for merge candidates inside a given range.
 356  * Any extents with matching state are merged together into a single
 357  * extent in the tree.  Extents with EXTENT_IO in their state field
 358  * are not merged because the end_io handlers need to be able to do
 359  * operations on them without sleeping (or doing allocations/splits).
 360  *
 361  * This should be called with the tree lock held.
 362  */
 363 static int merge_state(struct extent_map_tree *tree,
 364                        struct extent_state *state)
 365 {
 366         struct extent_state *other;
 367         struct rb_node *other_node;
 368
 369         if (state->state & EXTENT_IOBITS)
 370                 return 0;
 371
 372         other_node = rb_prev(&state->rb_node);
 373         if (other_node) {
 374                 other = rb_entry(other_node, struct extent_state, rb_node);
 375                 if (other->end == state->start - 1 &&
 376                     other->state == state->state) {
 377                         state->start = other->start;
 378                         other->in_tree = 0;
 379                         rb_erase(&other->rb_node, &tree->state);
 380                         free_extent_state(other);
 381                 }
 382         }
 383         other_node = rb_next(&state->rb_node);
 384         if (other_node) {
 385                 other = rb_entry(other_node, struct extent_state, rb_node);
 386                 if (other->start == state->end + 1 &&
 387                     other->state == state->state) {
 388                         other->start = state->start;
 389                         state->in_tree = 0;
 390                         rb_erase(&state->rb_node, &tree->state);
 391                         free_extent_state(state);
 392                 }
 393         }
 394         return 0;
 395 }
 396
 397 /*
 398  * insert an extent_state struct into the tree.  'bits' are set on the
 399  * struct before it is inserted.
 400  *
 401  * This may return -EEXIST if the extent is already there, in which case the
 402  * state struct is freed.
 403  *
 404  * The tree lock is not taken internally.  This is a utility function and
 405  * probably isn't what you want to call (see set/clear_extent_bit).
 406  */
 407 static int insert_state(struct extent_map_tree *tree,
 408                         struct extent_state *state, u64 start, u64 end,
 409                         int bits)
 410 {
 411         struct rb_node *node;
 412
 413         if (end < start) {
 414                 printk("end < start %Lu %Lu\n", end, start);
 415                 WARN_ON(1);
 416         }
 417         state->state |= bits;
 418         state->start = start;
 419         state->end = end;
 420         node = tree_insert(&tree->state, end, &state->rb_node);
 421         if (node) {
 422                 struct extent_state *found;
 423                 found = rb_entry(node, struct extent_state, rb_node);
 424                 printk("found node %Lu %Lu on insert of %Lu %Lu\n", found->start, found->end, start, end);
 425                 free_extent_state(state);
 426                 return -EEXIST;
 427         }
 428         merge_state(tree, state);
 429         return 0;
 430 }
 431
 432 /*
 433  * split a given extent state struct in two, inserting the preallocated
 434  * struct 'prealloc' as the newly created second half.  'split' indicates an
 435  * offset inside 'orig' where it should be split.
 436  *
 437  * Before calling,
 438  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 439  * are two extent state structs in the tree:
 440  * prealloc: [orig->start, split - 1]
 441  * orig: [ split, orig->end ]
 442  *
 443  * The tree locks are not taken by this function. They need to be held
 444  * by the caller.
 445  */
 446 static int split_state(struct extent_map_tree *tree, struct extent_state *orig,
 447                        struct extent_state *prealloc, u64 split)
 448 {
 449         struct rb_node *node;
 450         prealloc->start = orig->start;
 451         prealloc->end = split - 1;
 452         prealloc->state = orig->state;
 453         orig->start = split;
 454
 455         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 456         if (node) {
 457                 struct extent_state *found;
 458                 found = rb_entry(node, struct extent_state, rb_node);
 459                 printk("found node %Lu %Lu on insert of %Lu %Lu\n", found->start, found->end, prealloc->start, prealloc->end);
 460                 free_extent_state(prealloc);
 461                 return -EEXIST;
 462         }
 463         return 0;
 464 }
 465
 466 /*
 467  * utility function to clear some bits in an extent state struct.
 468  * it will optionally wake up any one waiting on this state (wake == 1), or
 469  * forcibly remove the state from the tree (delete == 1).
 470  *
 471  * If no bits are set on the state struct after clearing things, the
 472  * struct is freed and removed from the tree
 473  */
 474 static int clear_state_bit(struct extent_map_tree *tree,
 475                             struct extent_state *state, int bits, int wake,
 476                             int delete)
 477 {
 478         int ret = state->state & bits;
 479         state->state &= ~bits;
 480         if (wake)
 481                 wake_up(&state->wq);
 482         if (delete || state->state == 0) {
 483                 if (state->in_tree) {
 484                         rb_erase(&state->rb_node, &tree->state);
 485                         state->in_tree = 0;
 486                         free_extent_state(state);
 487                 } else {
 488                         WARN_ON(1);
 489                 }
 490         } else {
 491                 merge_state(tree, state);
 492         }
 493         return ret;
 494 }
 495
 496 /*
 497  * clear some bits on a range in the tree.  This may require splitting
 498  * or inserting elements in the tree, so the gfp mask is used to
 499  * indicate which allocations or sleeping are allowed.
 500  *
 501  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 502  * the given range from the tree regardless of state (ie for truncate).
 503  *
 504  * the range [start, end] is inclusive.
 505  *
 506  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 507  * bits were already set, or zero if none of the bits were already set.
 508  */
 509 int clear_extent_bit(struct extent_map_tree *tree, u64 start, u64 end,
 510                      int bits, int wake, int delete, gfp_t mask)
 511 {
 512         struct extent_state *state;
 513         struct extent_state *prealloc = NULL;
 514         struct rb_node *node;
 515         unsigned long flags;
 516         int err;
 517         int set = 0;
 518
 519 again:
 520         if (!prealloc && (mask & __GFP_WAIT)) {
 521                 prealloc = alloc_extent_state(mask);
 522                 if (!prealloc)
 523                         return -ENOMEM;
 524         }
 525
 526         write_lock_irqsave(&tree->lock, flags);
 527         /*
 528          * this search will find the extents that end after
 529          * our range starts
 530          */
 531         node = tree_search(&tree->state, start);
 532         if (!node)
 533                 goto out;
 534         state = rb_entry(node, struct extent_state, rb_node);
 535         if (state->start > end)
 536                 goto out;
 537         WARN_ON(state->end < start);
 538
 539         /*
 540          *     | ---- desired range ---- |
 541          *  | state | or
 542          *  | ------------- state -------------- |
 543          *
 544          * We need to split the extent we found, and may flip
 545          * bits on second half.
 546          *
 547          * If the extent we found extends past our range, we
 548          * just split and search again.  It'll get split again
 549          * the next time though.
 550          *
 551          * If the extent we found is inside our range, we clear
 552          * the desired bit on it.
 553          */
 554
 555         if (state->start < start) {
 556                 err = split_state(tree, state, prealloc, start);
 557                 BUG_ON(err == -EEXIST);
 558                 prealloc = NULL;
 559                 if (err)
 560                         goto out;
 561                 if (state->end <= end) {
 562                         start = state->end + 1;
 563                         set |= clear_state_bit(tree, state, bits,
 564                                         wake, delete);
 565                 } else {
 566                         start = state->start;
 567                 }
 568                 goto search_again;
 569         }
 570         /*
 571          * | ---- desired range ---- |
 572          *                        | state |
 573          * We need to split the extent, and clear the bit
 574          * on the first half
 575          */
 576         if (state->start <= end && state->end > end) {
 577                 err = split_state(tree, state, prealloc, end + 1);
 578                 BUG_ON(err == -EEXIST);
 579
 580                 if (wake)
 581                         wake_up(&state->wq);
 582                 set |= clear_state_bit(tree, prealloc, bits,
 583                                        wake, delete);
 584                 prealloc = NULL;
 585                 goto out;
 586         }
 587
 588         start = state->end + 1;
 589         set |= clear_state_bit(tree, state, bits, wake, delete);
 590         goto search_again;
 591
 592 out:
 593         write_unlock_irqrestore(&tree->lock, flags);
 594         if (prealloc)
 595                 free_extent_state(prealloc);
 596
 597         return set;
 598
 599 search_again:
 600         if (start > end)
 601                 goto out;
 602         write_unlock_irqrestore(&tree->lock, flags);
 603         if (mask & __GFP_WAIT)
 604                 cond_resched();
 605         goto again;
 606 }
 607 EXPORT_SYMBOL(clear_extent_bit);
 608
 609 static int wait_on_state(struct extent_map_tree *tree,
 610                          struct extent_state *state)
 611 {
 612         DEFINE_WAIT(wait);
 613         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 614         read_unlock_irq(&tree->lock);
 615         schedule();
 616         read_lock_irq(&tree->lock);
 617         finish_wait(&state->wq, &wait);
 618         return 0;
 619 }
 620
 621 /*
 622  * waits for one or more bits to clear on a range in the state tree.
 623  * The range [start, end] is inclusive.
 624  * The tree lock is taken by this function
 625  */
 626 int wait_extent_bit(struct extent_map_tree *tree, u64 start, u64 end, int bits)
 627 {
 628         struct extent_state *state;
 629         struct rb_node *node;
 630
 631         read_lock_irq(&tree->lock);
 632 again:
 633         while (1) {
 634                 /*
 635                  * this search will find all the extents that end after
 636                  * our range starts
 637                  */
 638                 node = tree_search(&tree->state, start);
 639                 if (!node)
 640                         break;
 641
 642                 state = rb_entry(node, struct extent_state, rb_node);
 643
 644                 if (state->start > end)
 645                         goto out;
 646
 647                 if (state->state & bits) {
 648                         start = state->start;
 649                         atomic_inc(&state->refs);
 650                         wait_on_state(tree, state);
 651                         free_extent_state(state);
 652                         goto again;
 653                 }
 654                 start = state->end + 1;
 655
 656                 if (start > end)
 657                         break;
 658
 659                 if (need_resched()) {
 660                         read_unlock_irq(&tree->lock);
 661                         cond_resched();
 662                         read_lock_irq(&tree->lock);
 663                 }
 664         }
 665 out:
 666         read_unlock_irq(&tree->lock);
 667         return 0;
 668 }
 669 EXPORT_SYMBOL(wait_extent_bit);
 670
 671 /*
 672  * set some bits on a range in the tree.  This may require allocations
 673  * or sleeping, so the gfp mask is used to indicate what is allowed.
 674  *
 675  * If 'exclusive' == 1, this will fail with -EEXIST if some part of the
 676  * range already has the desired bits set.  The start of the existing
 677  * range is returned in failed_start in this case.
 678  *
 679  * [start, end] is inclusive
 680  * This takes the tree lock.
 681  */
 682 int set_extent_bit(struct extent_map_tree *tree, u64 start, u64 end, int bits,
 683                    int exclusive, u64 *failed_start, gfp_t mask)
 684 {
 685         struct extent_state *state;
 686         struct extent_state *prealloc = NULL;
 687         struct rb_node *node;
 688         unsigned long flags;
 689         int err = 0;
 690         int set;
 691         u64 last_start;
 692         u64 last_end;
 693 again:
 694         if (!prealloc && (mask & __GFP_WAIT)) {
 695                 prealloc = alloc_extent_state(mask);
 696                 if (!prealloc)
 697                         return -ENOMEM;
 698         }
 699
 700         write_lock_irqsave(&tree->lock, flags);
 701         /*
 702          * this search will find all the extents that end after
 703          * our range starts.
 704          */
 705         node = tree_search(&tree->state, start);
 706         if (!node) {
 707                 err = insert_state(tree, prealloc, start, end, bits);
 708                 prealloc = NULL;
 709                 BUG_ON(err == -EEXIST);
 710                 goto out;
 711         }
 712
 713         state = rb_entry(node, struct extent_state, rb_node);
 714         last_start = state->start;
 715         last_end = state->end;
 716
 717         /*
 718          * | ---- desired range ---- |
 719          * | state |
 720          *
 721          * Just lock what we found and keep going
 722          */
 723         if (state->start == start && state->end <= end) {
 724                 set = state->state & bits;
 725                 if (set && exclusive) {
 726                         *failed_start = state->start;
 727                         err = -EEXIST;
 728                         goto out;
 729                 }
 730                 state->state |= bits;
 731                 start = state->end + 1;
 732                 merge_state(tree, state);
 733                 goto search_again;
 734         }
 735
 736         /*
 737          *     | ---- desired range ---- |
 738          * | state |
 739          *   or
 740          * | ------------- state -------------- |
 741          *
 742          * We need to split the extent we found, and may flip bits on
 743          * second half.
 744          *
 745          * If the extent we found extends past our
 746          * range, we just split and search again.  It'll get split
 747          * again the next time though.
 748          *
 749          * If the extent we found is inside our range, we set the
 750          * desired bit on it.
 751          */
 752         if (state->start < start) {
 753                 set = state->state & bits;
 754                 if (exclusive && set) {
 755                         *failed_start = start;
 756                         err = -EEXIST;
 757                         goto out;
 758                 }
 759                 err = split_state(tree, state, prealloc, start);
 760                 BUG_ON(err == -EEXIST);
 761                 prealloc = NULL;
 762                 if (err)
 763                         goto out;
 764                 if (state->end <= end) {
 765                         state->state |= bits;
 766                         start = state->end + 1;
 767                         merge_state(tree, state);
 768                 } else {
 769                         start = state->start;
 770                 }
 771                 goto search_again;
 772         }
 773         /*
 774          * | ---- desired range ---- |
 775          *     | state | or               | state |
 776          *
 777          * There's a hole, we need to insert something in it and
 778          * ignore the extent we found.
 779          */
 780         if (state->start > start) {
 781                 u64 this_end;
 782                 if (end < last_start)
 783                         this_end = end;
 784                 else
 785                         this_end = last_start -1;
 786                 err = insert_state(tree, prealloc, start, this_end,
 787                                    bits);
 788                 prealloc = NULL;
 789                 BUG_ON(err == -EEXIST);
 790                 if (err)
 791                         goto out;
 792                 start = this_end + 1;
 793                 goto search_again;
 794         }
 795         /*
 796          * | ---- desired range ---- |
 797          *                        | state |
 798          * We need to split the extent, and set the bit
 799          * on the first half
 800          */
 801         if (state->start <= end && state->end > end) {
 802                 set = state->state & bits;
 803                 if (exclusive && set) {
 804                         *failed_start = start;
 805                         err = -EEXIST;
 806                         goto out;
 807                 }
 808                 err = split_state(tree, state, prealloc, end + 1);
 809                 BUG_ON(err == -EEXIST);
 810
 811                 prealloc->state |= bits;
 812                 merge_state(tree, prealloc);
 813                 prealloc = NULL;
 814                 goto out;
 815         }
 816
 817         goto search_again;
 818
 819 out:
 820         write_unlock_irqrestore(&tree->lock, flags);
 821         if (prealloc)
 822                 free_extent_state(prealloc);
 823
 824         return err;
 825
 826 search_again:
 827         if (start > end)
 828                 goto out;
 829         write_unlock_irqrestore(&tree->lock, flags);
 830         if (mask & __GFP_WAIT)
 831                 cond_resched();
 832         goto again;
 833 }
 834 EXPORT_SYMBOL(set_extent_bit);
 835
 836 /* wrappers around set/clear extent bit */
 837 int set_extent_dirty(struct extent_map_tree *tree, u64 start, u64 end,
 838                      gfp_t mask)
 839 {
 840         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 841                               mask);
 842 }
 843 EXPORT_SYMBOL(set_extent_dirty);
 844
 845 int set_extent_bits(struct extent_map_tree *tree, u64 start, u64 end,
 846                     int bits, gfp_t mask)
 847 {
 848         return set_extent_bit(tree, start, end, bits, 0, NULL,
 849                               mask);
 850 }
 851 EXPORT_SYMBOL(set_extent_bits);
 852
 853 int clear_extent_bits(struct extent_map_tree *tree, u64 start, u64 end,
 854                       int bits, gfp_t mask)
 855 {
 856         return clear_extent_bit(tree, start, end, bits, 0, 0, mask);
 857 }
 858 EXPORT_SYMBOL(clear_extent_bits);
 859
 860 int set_extent_delalloc(struct extent_map_tree *tree, u64 start, u64 end,
 861                      gfp_t mask)
 862 {
 863         return set_extent_bit(tree, start, end,
 864                               EXTENT_DELALLOC | EXTENT_DIRTY, 0, NULL,
 865                               mask);
 866 }
 867 EXPORT_SYMBOL(set_extent_delalloc);
 868
 869 int clear_extent_dirty(struct extent_map_tree *tree, u64 start, u64 end,
 870                        gfp_t mask)
 871 {
 872         return clear_extent_bit(tree, start, end,
 873                                 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, mask);
 874 }
 875 EXPORT_SYMBOL(clear_extent_dirty);
 876
 877 int set_extent_new(struct extent_map_tree *tree, u64 start, u64 end,
 878                      gfp_t mask)
 879 {
 880         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 881                               mask);
 882 }
 883 EXPORT_SYMBOL(set_extent_new);
 884
 885 int clear_extent_new(struct extent_map_tree *tree, u64 start, u64 end,
 886                        gfp_t mask)
 887 {
 888         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0, mask);
 889 }
 890 EXPORT_SYMBOL(clear_extent_new);
 891
 892 int set_extent_uptodate(struct extent_map_tree *tree, u64 start, u64 end,
 893                         gfp_t mask)
 894 {
 895         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, NULL,
 896                               mask);
 897 }
 898 EXPORT_SYMBOL(set_extent_uptodate);
 899
 900 int clear_extent_uptodate(struct extent_map_tree *tree, u64 start, u64 end,
 901                           gfp_t mask)
 902 {
 903         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0, mask);
 904 }
 905 EXPORT_SYMBOL(clear_extent_uptodate);
 906
 907 int set_extent_writeback(struct extent_map_tree *tree, u64 start, u64 end,
 908                          gfp_t mask)
 909 {
 910         return set_extent_bit(tree, start, end, EXTENT_WRITEBACK,
 911                               0, NULL, mask);
 912 }
 913 EXPORT_SYMBOL(set_extent_writeback);
 914
 915 int clear_extent_writeback(struct extent_map_tree *tree, u64 start, u64 end,
 916                            gfp_t mask)
 917 {
 918         return clear_extent_bit(tree, start, end, EXTENT_WRITEBACK, 1, 0, mask);
 919 }
 920 EXPORT_SYMBOL(clear_extent_writeback);
 921
 922 int wait_on_extent_writeback(struct extent_map_tree *tree, u64 start, u64 end)
 923 {
 924         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 925 }
 926 EXPORT_SYMBOL(wait_on_extent_writeback);
 927
 928 /*
 929  * locks a range in ascending order, waiting for any locked regions
 930  * it hits on the way.  [start,end] are inclusive, and this will sleep.
 931  */
 932 int lock_extent(struct extent_map_tree *tree, u64 start, u64 end, gfp_t mask)
 933 {
 934         int err;
 935         u64 failed_start;
 936         while (1) {
 937                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED, 1,
 938                                      &failed_start, mask);
 939                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
 940                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
 941                         start = failed_start;
 942                 } else {
 943                         break;
 944                 }
 945                 WARN_ON(start > end);
 946         }
 947         return err;
 948 }
 949 EXPORT_SYMBOL(lock_extent);
 950
 951 int unlock_extent(struct extent_map_tree *tree, u64 start, u64 end,
 952                   gfp_t mask)
 953 {
 954         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, mask);
 955 }
 956 EXPORT_SYMBOL(unlock_extent);
 957
 958 /*
 959  * helper function to set pages and extents in the tree dirty
 960  */
 961 int set_range_dirty(struct extent_map_tree *tree, u64 start, u64 end)
 962 {
 963         unsigned long index = start >> PAGE_CACHE_SHIFT;
 964         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 965         struct page *page;
 966
 967         while (index <= end_index) {
 968                 page = find_get_page(tree->mapping, index);
 969                 BUG_ON(!page);
 970                 __set_page_dirty_nobuffers(page);
 971                 page_cache_release(page);
 972                 index++;
 973         }
 974         set_extent_dirty(tree, start, end, GFP_NOFS);
 975         return 0;
 976 }
 977 EXPORT_SYMBOL(set_range_dirty);
 978
 979 /*
 980  * helper function to set both pages and extents in the tree writeback
 981  */
 982 int set_range_writeback(struct extent_map_tree *tree, u64 start, u64 end)
 983 {
 984         unsigned long index = start >> PAGE_CACHE_SHIFT;
 985         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 986         struct page *page;
 987
 988         while (index <= end_index) {
 989                 page = find_get_page(tree->mapping, index);
 990                 BUG_ON(!page);
 991                 set_page_writeback(page);
 992                 page_cache_release(page);
 993                 index++;
 994         }
 995         set_extent_writeback(tree, start, end, GFP_NOFS);
 996         return 0;
 997 }
 998 EXPORT_SYMBOL(set_range_writeback);
 999
1000 int find_first_extent_bit(struct extent_map_tree *tree, u64 start,
1001                           u64 *start_ret, u64 *end_ret, int bits)
1002 {
1003         struct rb_node *node;
1004         struct extent_state *state;
1005         int ret = 1;
1006
1007         read_lock_irq(&tree->lock);
1008         /*
1009          * this search will find all the extents that end after
1010          * our range starts.
1011          */
1012         node = tree_search(&tree->state, start);
1013         if (!node || IS_ERR(node)) {
1014                 goto out;
1015         }
1016
1017         while(1) {
1018                 state = rb_entry(node, struct extent_state, rb_node);
1019                 if (state->end >= start && (state->state & bits)) {
1020                         *start_ret = state->start;
1021                         *end_ret = state->end;
1022                         ret = 0;
1023                         break;
1024                 }
1025                 node = rb_next(node);
1026                 if (!node)
1027                         break;
1028         }
1029 out:
1030         read_unlock_irq(&tree->lock);
1031         return ret;
1032 }
1033 EXPORT_SYMBOL(find_first_extent_bit);
1034
1035 u64 find_lock_delalloc_range(struct extent_map_tree *tree,
1036                              u64 start, u64 lock_start, u64 *end, u64 max_bytes)
1037 {
1038         struct rb_node *node;
1039         struct extent_state *state;
1040         u64 cur_start = start;
1041         u64 found = 0;
1042         u64 total_bytes = 0;
1043
1044         write_lock_irq(&tree->lock);
1045         /*
1046          * this search will find all the extents that end after
1047          * our range starts.
1048          */
1049 search_again:
1050         node = tree_search(&tree->state, cur_start);
1051         if (!node || IS_ERR(node)) {
1052                 goto out;
1053         }
1054
1055         while(1) {
1056                 state = rb_entry(node, struct extent_state, rb_node);
1057                 if (state->start != cur_start) {
1058                         goto out;
1059                 }
1060                 if (!(state->state & EXTENT_DELALLOC)) {
1061                         goto out;
1062                 }
1063                 if (state->start >= lock_start) {
1064                         if (state->state & EXTENT_LOCKED) {
1065                                 DEFINE_WAIT(wait);
1066                                 atomic_inc(&state->refs);
1067                                 prepare_to_wait(&state->wq, &wait,
1068                                                 TASK_UNINTERRUPTIBLE);
1069                                 write_unlock_irq(&tree->lock);
1070                                 schedule();
1071                                 write_lock_irq(&tree->lock);
1072                                 finish_wait(&state->wq, &wait);
1073                                 free_extent_state(state);
1074                                 goto search_again;
1075                         }
1076                         state->state |= EXTENT_LOCKED;
1077                 }
1078                 found++;
1079                 *end = state->end;
1080                 cur_start = state->end + 1;
1081                 node = rb_next(node);
1082                 if (!node)
1083                         break;
1084                 total_bytes += state->end - state->start + 1;
1085                 if (total_bytes >= max_bytes)
1086                         break;
1087         }
1088 out:
1089         write_unlock_irq(&tree->lock);
1090         return found;
1091 }
1092
1093 /*
1094  * helper function to lock both pages and extents in the tree.
1095  * pages must be locked first.
1096  */
1097 int lock_range(struct extent_map_tree *tree, u64 start, u64 end)
1098 {
1099         unsigned long index = start >> PAGE_CACHE_SHIFT;
1100         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1101         struct page *page;
1102         int err;
1103
1104         while (index <= end_index) {
1105                 page = grab_cache_page(tree->mapping, index);
1106                 if (!page) {
1107                         err = -ENOMEM;
1108                         goto failed;
1109                 }
1110                 if (IS_ERR(page)) {
1111                         err = PTR_ERR(page);
1112                         goto failed;
1113                 }
1114                 index++;
1115         }
1116         lock_extent(tree, start, end, GFP_NOFS);
1117         return 0;
1118
1119 failed:
1120         /*
1121          * we failed above in getting the page at 'index', so we undo here
1122          * up to but not including the page at 'index'
1123          */
1124         end_index = index;
1125         index = start >> PAGE_CACHE_SHIFT;
1126         while (index < end_index) {
1127                 page = find_get_page(tree->mapping, index);
1128                 unlock_page(page);
1129                 page_cache_release(page);
1130                 index++;
1131         }
1132         return err;
1133 }
1134 EXPORT_SYMBOL(lock_range);
1135
1136 /*
1137  * helper function to unlock both pages and extents in the tree.
1138  */
1139 int unlock_range(struct extent_map_tree *tree, u64 start, u64 end)
1140 {
1141         unsigned long index = start >> PAGE_CACHE_SHIFT;
1142         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1143         struct page *page;
1144
1145         while (index <= end_index) {
1146                 page = find_get_page(tree->mapping, index);
1147                 unlock_page(page);
1148                 page_cache_release(page);
1149                 index++;
1150         }
1151         unlock_extent(tree, start, end, GFP_NOFS);
1152         return 0;
1153 }
1154 EXPORT_SYMBOL(unlock_range);
1155
1156 int set_state_private(struct extent_map_tree *tree, u64 start, u64 private)
1157 {
1158         struct rb_node *node;
1159         struct extent_state *state;
1160         int ret = 0;
1161
1162         write_lock_irq(&tree->lock);
1163         /*
1164          * this search will find all the extents that end after
1165          * our range starts.
1166          */
1167         node = tree_search(&tree->state, start);
1168         if (!node || IS_ERR(node)) {
1169                 ret = -ENOENT;
1170                 goto out;
1171         }
1172         state = rb_entry(node, struct extent_state, rb_node);
1173         if (state->start != start) {
1174                 ret = -ENOENT;
1175                 goto out;
1176         }
1177         state->private = private;
1178 out:
1179         write_unlock_irq(&tree->lock);
1180         return ret;
1181 }
1182
1183 int get_state_private(struct extent_map_tree *tree, u64 start, u64 *private)
1184 {
1185         struct rb_node *node;
1186         struct extent_state *state;
1187         int ret = 0;
1188
1189         read_lock_irq(&tree->lock);
1190         /*
1191          * this search will find all the extents that end after
1192          * our range starts.
1193          */
1194         node = tree_search(&tree->state, start);
1195         if (!node || IS_ERR(node)) {
1196                 ret = -ENOENT;
1197                 goto out;
1198         }
1199         state = rb_entry(node, struct extent_state, rb_node);
1200         if (state->start != start) {
1201                 ret = -ENOENT;
1202                 goto out;
1203         }
1204         *private = state->private;
1205 out:
1206         read_unlock_irq(&tree->lock);
1207         return ret;
1208 }
1209
1210 /*
1211  * searches a range in the state tree for a given mask.
1212  * If 'filled' == 1, this returns 1 only if ever extent in the tree
1213  * has the bits set.  Otherwise, 1 is returned if any bit in the
1214  * range is found set.
1215  */
1216 int test_range_bit(struct extent_map_tree *tree, u64 start, u64 end,
1217                    int bits, int filled)
1218 {
1219         struct extent_state *state = NULL;
1220         struct rb_node *node;
1221         int bitset = 0;
1222
1223         read_lock_irq(&tree->lock);
1224         node = tree_search(&tree->state, start);
1225         while (node && start <= end) {
1226                 state = rb_entry(node, struct extent_state, rb_node);
1227
1228                 if (filled && state->start > start) {
1229                         bitset = 0;
1230                         break;
1231                 }
1232
1233                 if (state->start > end)
1234                         break;
1235
1236                 if (state->state & bits) {
1237                         bitset = 1;
1238                         if (!filled)
1239                                 break;
1240                 } else if (filled) {
1241                         bitset = 0;
1242                         break;
1243                 }
1244                 start = state->end + 1;
1245                 if (start > end)
1246                         break;
1247                 node = rb_next(node);
1248         }
1249         read_unlock_irq(&tree->lock);
1250         return bitset;
1251 }
1252 EXPORT_SYMBOL(test_range_bit);
1253
1254 /*
1255  * helper function to set a given page up to date if all the
1256  * extents in the tree for that page are up to date
1257  */
1258 static int check_page_uptodate(struct extent_map_tree *tree,
1259                                struct page *page)
1260 {
1261         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1262         u64 end = start + PAGE_CACHE_SIZE - 1;
1263         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1))
1264                 SetPageUptodate(page);
1265         return 0;
1266 }
1267
1268 /*
1269  * helper function to unlock a page if all the extents in the tree
1270  * for that page are unlocked
1271  */
1272 static int check_page_locked(struct extent_map_tree *tree,
1273                              struct page *page)
1274 {
1275         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1276         u64 end = start + PAGE_CACHE_SIZE - 1;
1277         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0))
1278                 unlock_page(page);
1279         return 0;
1280 }
1281
1282 /*
1283  * helper function to end page writeback if all the extents
1284  * in the tree for that page are done with writeback
1285  */
1286 static int check_page_writeback(struct extent_map_tree *tree,
1287                              struct page *page)
1288 {
1289         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1290         u64 end = start + PAGE_CACHE_SIZE - 1;
1291         if (!test_range_bit(tree, start, end, EXTENT_WRITEBACK, 0))
1292                 end_page_writeback(page);
1293         return 0;
1294 }
1295
1296 /* lots and lots of room for performance fixes in the end_bio funcs */
1297
1298 /*
1299  * after a writepage IO is done, we need to:
1300  * clear the uptodate bits on error
1301  * clear the writeback bits in the extent tree for this IO
1302  * end_page_writeback if the page has no more pending IO
1303  *
1304  * Scheduling is not allowed, so the extent state tree is expected
1305  * to have one and only one object corresponding to this IO.
1306  */
1307 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1308 static void end_bio_extent_writepage(struct bio *bio, int err)
1309 #else
1310 static int end_bio_extent_writepage(struct bio *bio,
1311                                    unsigned int bytes_done, int err)
1312 #endif
1313 {
1314         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1315         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1316         struct extent_map_tree *tree = bio->bi_private;
1317         u64 start;
1318         u64 end;
1319         int whole_page;
1320
1321 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1322         if (bio->bi_size)
1323                 return 1;
1324 #endif
1325
1326         do {
1327                 struct page *page = bvec->bv_page;
1328                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1329                          bvec->bv_offset;
1330                 end = start + bvec->bv_len - 1;
1331
1332                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1333                         whole_page = 1;
1334                 else
1335                         whole_page = 0;
1336
1337                 if (--bvec >= bio->bi_io_vec)
1338                         prefetchw(&bvec->bv_page->flags);
1339
1340                 if (!uptodate) {
1341                         clear_extent_uptodate(tree, start, end, GFP_ATOMIC);
1342                         ClearPageUptodate(page);
1343                         SetPageError(page);
1344                 }
1345                 clear_extent_writeback(tree, start, end, GFP_ATOMIC);
1346
1347                 if (whole_page)
1348                         end_page_writeback(page);
1349                 else
1350                         check_page_writeback(tree, page);
1351                 if (tree->ops && tree->ops->writepage_end_io_hook)
1352                         tree->ops->writepage_end_io_hook(page, start, end);
1353         } while (bvec >= bio->bi_io_vec);
1354
1355         bio_put(bio);
1356 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1357         return 0;
1358 #endif
1359 }
1360
1361 /*
1362  * after a readpage IO is done, we need to:
1363  * clear the uptodate bits on error
1364  * set the uptodate bits if things worked
1365  * set the page up to date if all extents in the tree are uptodate
1366  * clear the lock bit in the extent tree
1367  * unlock the page if there are no other extents locked for it
1368  *
1369  * Scheduling is not allowed, so the extent state tree is expected
1370  * to have one and only one object corresponding to this IO.
1371  */
1372 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1373 static void end_bio_extent_readpage(struct bio *bio, int err)
1374 #else
1375 static int end_bio_extent_readpage(struct bio *bio,
1376                                    unsigned int bytes_done, int err)
1377 #endif
1378 {
1379         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1380         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1381         struct extent_map_tree *tree = bio->bi_private;
1382         u64 start;
1383         u64 end;
1384         int whole_page;
1385         int ret;
1386
1387 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1388         if (bio->bi_size)
1389                 return 1;
1390 #endif
1391
1392         do {
1393                 struct page *page = bvec->bv_page;
1394                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1395                         bvec->bv_offset;
1396                 end = start + bvec->bv_len - 1;
1397
1398                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1399                         whole_page = 1;
1400                 else
1401                         whole_page = 0;
1402
1403                 if (--bvec >= bio->bi_io_vec)
1404                         prefetchw(&bvec->bv_page->flags);
1405
1406                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1407                         ret = tree->ops->readpage_end_io_hook(page, start, end);
1408                         if (ret)
1409                                 uptodate = 0;
1410                 }
1411                 if (uptodate) {
1412                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1413                         if (whole_page)
1414                                 SetPageUptodate(page);
1415                         else
1416                                 check_page_uptodate(tree, page);
1417                 } else {
1418                         ClearPageUptodate(page);
1419                         SetPageError(page);
1420                 }
1421
1422                 unlock_extent(tree, start, end, GFP_ATOMIC);
1423
1424                 if (whole_page)
1425                         unlock_page(page);
1426                 else
1427                         check_page_locked(tree, page);
1428         } while (bvec >= bio->bi_io_vec);
1429
1430         bio_put(bio);
1431 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1432         return 0;
1433 #endif
1434 }
1435
1436 /*
1437  * IO done from prepare_write is pretty simple, we just unlock
1438  * the structs in the extent tree when done, and set the uptodate bits
1439  * as appropriate.
1440  */
1441 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1442 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1443 #else
1444 static int end_bio_extent_preparewrite(struct bio *bio,
1445                                        unsigned int bytes_done, int err)
1446 #endif
1447 {
1448         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1449         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1450         struct extent_map_tree *tree = bio->bi_private;
1451         u64 start;
1452         u64 end;
1453
1454 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1455         if (bio->bi_size)
1456                 return 1;
1457 #endif
1458
1459         do {
1460                 struct page *page = bvec->bv_page;
1461                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1462                         bvec->bv_offset;
1463                 end = start + bvec->bv_len - 1;
1464
1465                 if (--bvec >= bio->bi_io_vec)
1466                         prefetchw(&bvec->bv_page->flags);
1467
1468                 if (uptodate) {
1469                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1470                 } else {
1471                         ClearPageUptodate(page);
1472                         SetPageError(page);
1473                 }
1474
1475                 unlock_extent(tree, start, end, GFP_ATOMIC);
1476
1477         } while (bvec >= bio->bi_io_vec);
1478
1479         bio_put(bio);
1480 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1481         return 0;
1482 #endif
1483 }
1484
1485 static struct bio *
1486 extent_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1487                  gfp_t gfp_flags)
1488 {
1489         struct bio *bio;
1490
1491         bio = bio_alloc(gfp_flags, nr_vecs);
1492
1493         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1494                 while (!bio && (nr_vecs /= 2))
1495                         bio = bio_alloc(gfp_flags, nr_vecs);
1496         }
1497
1498         if (bio) {
1499                 bio->bi_bdev = bdev;
1500                 bio->bi_sector = first_sector;
1501         }
1502         return bio;
1503 }
1504
1505 static int submit_one_bio(int rw, struct bio *bio)
1506 {
1507         int ret = 0;
1508         bio_get(bio);
1509         submit_bio(rw, bio);
1510         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1511                 ret = -EOPNOTSUPP;
1512         bio_put(bio);
1513         return ret;
1514 }
1515
1516 static int submit_extent_page(int rw, struct extent_map_tree *tree,
1517                               struct page *page, sector_t sector,
1518                               size_t size, unsigned long offset,
1519                               struct block_device *bdev,
1520                               struct bio **bio_ret,
1521                               unsigned long max_pages,
1522                               bio_end_io_t end_io_func)
1523 {
1524         int ret = 0;
1525         struct bio *bio;
1526         int nr;
1527
1528         if (bio_ret && *bio_ret) {
1529                 bio = *bio_ret;
1530                 if (bio->bi_sector + (bio->bi_size >> 9) != sector ||
1531                     bio_add_page(bio, page, size, offset) < size) {
1532                         ret = submit_one_bio(rw, bio);
1533                         bio = NULL;
1534                 } else {
1535                         return 0;
1536                 }
1537         }
1538         nr = min_t(int, max_pages, bio_get_nr_vecs(bdev));
1539         bio = extent_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1540         if (!bio) {
1541                 printk("failed to allocate bio nr %d\n", nr);
1542         }
1543         bio_add_page(bio, page, size, offset);
1544         bio->bi_end_io = end_io_func;
1545         bio->bi_private = tree;
1546         if (bio_ret) {
1547                 *bio_ret = bio;
1548         } else {
1549                 ret = submit_one_bio(rw, bio);
1550         }
1551
1552         return ret;
1553 }
1554
1555 void set_page_extent_mapped(struct page *page)
1556 {
1557         if (!PagePrivate(page)) {
1558                 SetPagePrivate(page);
1559                 WARN_ON(!page->mapping->a_ops->invalidatepage);
1560                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1561                 page_cache_get(page);
1562         }
1563 }
1564
1565 /*
1566  * basic readpage implementation.  Locked extent state structs are inserted
1567  * into the tree that are removed when the IO is done (by the end_io
1568  * handlers)
1569  */
1570 static int __extent_read_full_page(struct extent_map_tree *tree,
1571                                    struct page *page,
1572                                    get_extent_t *get_extent,
1573                                    struct bio **bio)
1574 {
1575         struct inode *inode = page->mapping->host;
1576         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1577         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1578         u64 end;
1579         u64 cur = start;
1580         u64 extent_offset;
1581         u64 last_byte = i_size_read(inode);
1582         u64 block_start;
1583         u64 cur_end;
1584         sector_t sector;
1585         struct extent_map *em;
1586         struct block_device *bdev;
1587         int ret;
1588         int nr = 0;
1589         size_t page_offset = 0;
1590         size_t iosize;
1591         size_t blocksize = inode->i_sb->s_blocksize;
1592
1593         set_page_extent_mapped(page);
1594
1595         end = page_end;
1596         lock_extent(tree, start, end, GFP_NOFS);
1597
1598         while (cur <= end) {
1599                 if (cur >= last_byte) {
1600                         iosize = PAGE_CACHE_SIZE - page_offset;
1601                         zero_user_page(page, page_offset, iosize, KM_USER0);
1602                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1603                                             GFP_NOFS);
1604                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1605                         break;
1606                 }
1607                 em = get_extent(inode, page, page_offset, cur, end, 0);
1608                 if (IS_ERR(em) || !em) {
1609                         SetPageError(page);
1610                         unlock_extent(tree, cur, end, GFP_NOFS);
1611                         break;
1612                 }
1613
1614                 extent_offset = cur - em->start;
1615                 BUG_ON(em->end < cur);
1616                 BUG_ON(end < cur);
1617
1618                 iosize = min(em->end - cur, end - cur) + 1;
1619                 cur_end = min(em->end, end);
1620                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1621                 sector = (em->block_start + extent_offset) >> 9;
1622                 bdev = em->bdev;
1623                 block_start = em->block_start;
1624                 free_extent_map(em);
1625                 em = NULL;
1626
1627                 /* we've found a hole, just zero and go on */
1628                 if (block_start == EXTENT_MAP_HOLE) {
1629                         zero_user_page(page, page_offset, iosize, KM_USER0);
1630                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1631                                             GFP_NOFS);
1632                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1633                         cur = cur + iosize;
1634                         page_offset += iosize;
1635                         continue;
1636                 }
1637                 /* the get_extent function already copied into the page */
1638                 if (test_range_bit(tree, cur, cur_end, EXTENT_UPTODATE, 1)) {
1639                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1640                         cur = cur + iosize;
1641                         page_offset += iosize;
1642                         continue;
1643                 }
1644
1645                 ret = 0;
1646                 if (tree->ops && tree->ops->readpage_io_hook) {
1647                         ret = tree->ops->readpage_io_hook(page, cur,
1648                                                           cur + iosize - 1);
1649                 }
1650                 if (!ret) {
1651                         unsigned long nr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
1652                         nr -= page->index;
1653                         ret = submit_extent_page(READ, tree, page,
1654                                          sector, iosize, page_offset,
1655                                          bdev, bio, nr,
1656                                          end_bio_extent_readpage);
1657                 }
1658                 if (ret)
1659                         SetPageError(page);
1660                 cur = cur + iosize;
1661                 page_offset += iosize;
1662                 nr++;
1663         }
1664         if (!nr) {
1665                 if (!PageError(page))
1666                         SetPageUptodate(page);
1667                 unlock_page(page);
1668         }
1669         return 0;
1670 }
1671
1672 int extent_read_full_page(struct extent_map_tree *tree, struct page *page,
1673                             get_extent_t *get_extent)
1674 {
1675         struct bio *bio = NULL;
1676         int ret;
1677
1678         ret = __extent_read_full_page(tree, page, get_extent, &bio);
1679         if (bio)
1680                 submit_one_bio(READ, bio);
1681         return ret;
1682 }
1683 EXPORT_SYMBOL(extent_read_full_page);
1684
1685 /*
1686  * the writepage semantics are similar to regular writepage.  extent
1687  * records are inserted to lock ranges in the tree, and as dirty areas
1688  * are found, they are marked writeback.  Then the lock bits are removed
1689  * and the end_io handler clears the writeback ranges
1690  */
1691 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
1692                               void *data)
1693 {
1694         struct inode *inode = page->mapping->host;
1695         struct extent_page_data *epd = data;
1696         struct extent_map_tree *tree = epd->tree;
1697         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1698         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1699         u64 end;
1700         u64 cur = start;
1701         u64 extent_offset;
1702         u64 last_byte = i_size_read(inode);
1703         u64 block_start;
1704         u64 iosize;
1705         sector_t sector;
1706         struct extent_map *em;
1707         struct block_device *bdev;
1708         int ret;
1709         int nr = 0;
1710         size_t page_offset = 0;
1711         size_t blocksize;
1712         loff_t i_size = i_size_read(inode);
1713         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
1714         u64 nr_delalloc;
1715         u64 delalloc_end;
1716
1717         WARN_ON(!PageLocked(page));
1718         if (page->index > end_index) {
1719                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
1720                 unlock_page(page);
1721                 return 0;
1722         }
1723
1724         if (page->index == end_index) {
1725                 size_t offset = i_size & (PAGE_CACHE_SIZE - 1);
1726                 zero_user_page(page, offset,
1727                                PAGE_CACHE_SIZE - offset, KM_USER0);
1728         }
1729
1730         set_page_extent_mapped(page);
1731
1732         lock_extent(tree, start, page_end, GFP_NOFS);
1733         nr_delalloc = find_lock_delalloc_range(tree, start, page_end + 1,
1734                                                &delalloc_end,
1735                                                128 * 1024 * 1024);
1736         if (nr_delalloc) {
1737                 tree->ops->fill_delalloc(inode, start, delalloc_end);
1738                 if (delalloc_end >= page_end + 1) {
1739                         clear_extent_bit(tree, page_end + 1, delalloc_end,
1740                                          EXTENT_LOCKED | EXTENT_DELALLOC,
1741                                          1, 0, GFP_NOFS);
1742                 }
1743                 clear_extent_bit(tree, start, page_end, EXTENT_DELALLOC,
1744                                  0, 0, GFP_NOFS);
1745                 if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0)) {
1746                         printk("found delalloc bits after clear extent_bit\n");
1747                 }
1748         } else if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0)) {
1749                 printk("found delalloc bits after find_delalloc_range returns 0\n");
1750         }
1751
1752         end = page_end;
1753         if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0)) {
1754                 printk("found delalloc bits after lock_extent\n");
1755         }
1756
1757         if (last_byte <= start) {
1758                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
1759                 goto done;
1760         }
1761
1762         set_extent_uptodate(tree, start, page_end, GFP_NOFS);
1763         blocksize = inode->i_sb->s_blocksize;
1764
1765         while (cur <= end) {
1766                 if (cur >= last_byte) {
1767                         clear_extent_dirty(tree, cur, page_end, GFP_NOFS);
1768                         break;
1769                 }
1770                 em = epd->get_extent(inode, page, page_offset, cur, end, 1);
1771                 if (IS_ERR(em) || !em) {
1772                         SetPageError(page);
1773                         break;
1774                 }
1775
1776                 extent_offset = cur - em->start;
1777                 BUG_ON(em->end < cur);
1778                 BUG_ON(end < cur);
1779                 iosize = min(em->end - cur, end - cur) + 1;
1780                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1781                 sector = (em->block_start + extent_offset) >> 9;
1782                 bdev = em->bdev;
1783                 block_start = em->block_start;
1784                 free_extent_map(em);
1785                 em = NULL;
1786
1787                 if (block_start == EXTENT_MAP_HOLE ||
1788                     block_start == EXTENT_MAP_INLINE) {
1789                         clear_extent_dirty(tree, cur,
1790                                            cur + iosize - 1, GFP_NOFS);
1791                         cur = cur + iosize;
1792                         page_offset += iosize;
1793                         continue;
1794                 }
1795
1796                 /* leave this out until we have a page_mkwrite call */
1797                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
1798                                    EXTENT_DIRTY, 0)) {
1799                         cur = cur + iosize;
1800                         page_offset += iosize;
1801                         continue;
1802                 }
1803                 clear_extent_dirty(tree, cur, cur + iosize - 1, GFP_NOFS);
1804                 if (tree->ops && tree->ops->writepage_io_hook) {
1805                         ret = tree->ops->writepage_io_hook(page, cur,
1806                                                 cur + iosize - 1);
1807                 } else {
1808                         ret = 0;
1809                 }
1810                 if (ret)
1811                         SetPageError(page);
1812                 else {
1813                         unsigned long nr = end_index + 1;
1814                         set_range_writeback(tree, cur, cur + iosize - 1);
1815
1816                         ret = submit_extent_page(WRITE, tree, page, sector,
1817                                                  iosize, page_offset, bdev,
1818                                                  &epd->bio, nr,
1819                                                  end_bio_extent_writepage);
1820                         if (ret)
1821                                 SetPageError(page);
1822                 }
1823                 cur = cur + iosize;
1824                 page_offset += iosize;
1825                 nr++;
1826         }
1827 done:
1828         unlock_extent(tree, start, page_end, GFP_NOFS);
1829         unlock_page(page);
1830         return 0;
1831 }
1832
1833 int extent_write_full_page(struct extent_map_tree *tree, struct page *page,
1834                           get_extent_t *get_extent,
1835                           struct writeback_control *wbc)
1836 {
1837         int ret;
1838         struct extent_page_data epd = {
1839                 .bio = NULL,
1840                 .tree = tree,
1841                 .get_extent = get_extent,
1842         };
1843
1844         ret = __extent_writepage(page, wbc, &epd);
1845         if (epd.bio)
1846                 submit_one_bio(WRITE, epd.bio);
1847         return ret;
1848 }
1849 EXPORT_SYMBOL(extent_write_full_page);
1850
1851 int extent_writepages(struct extent_map_tree *tree,
1852                       struct address_space *mapping,
1853                       get_extent_t *get_extent,
1854                       struct writeback_control *wbc)
1855 {
1856         int ret;
1857         struct extent_page_data epd = {
1858                 .bio = NULL,
1859                 .tree = tree,
1860                 .get_extent = get_extent,
1861         };
1862
1863         ret = write_cache_pages(mapping, wbc, __extent_writepage, &epd);
1864         if (epd.bio)
1865                 submit_one_bio(WRITE, epd.bio);
1866         return ret;
1867 }
1868 EXPORT_SYMBOL(extent_writepages);
1869
1870 int extent_readpages(struct extent_map_tree *tree,
1871                      struct address_space *mapping,
1872                      struct list_head *pages, unsigned nr_pages,
1873                      get_extent_t get_extent)
1874 {
1875         struct bio *bio = NULL;
1876         unsigned page_idx;
1877         struct pagevec pvec;
1878
1879         pagevec_init(&pvec, 0);
1880         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
1881                 struct page *page = list_entry(pages->prev, struct page, lru);
1882
1883                 prefetchw(&page->flags);
1884                 list_del(&page->lru);
1885                 /*
1886                  * what we want to do here is call add_to_page_cache_lru,
1887                  * but that isn't exported, so we reproduce it here
1888                  */
1889                 if (!add_to_page_cache(page, mapping,
1890                                         page->index, GFP_KERNEL)) {
1891
1892                         /* open coding of lru_cache_add, also not exported */
1893                         page_cache_get(page);
1894                         if (!pagevec_add(&pvec, page))
1895                                 __pagevec_lru_add(&pvec);
1896                         __extent_read_full_page(tree, page, get_extent, &bio);
1897                 }
1898                 page_cache_release(page);
1899         }
1900         if (pagevec_count(&pvec))
1901                 __pagevec_lru_add(&pvec);
1902         BUG_ON(!list_empty(pages));
1903         if (bio)
1904                 submit_one_bio(READ, bio);
1905         return 0;
1906 }
1907 EXPORT_SYMBOL(extent_readpages);
1908
1909 /*
1910  * basic invalidatepage code, this waits on any locked or writeback
1911  * ranges corresponding to the page, and then deletes any extent state
1912  * records from the tree
1913  */
1914 int extent_invalidatepage(struct extent_map_tree *tree,
1915                           struct page *page, unsigned long offset)
1916 {
1917         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
1918         u64 end = start + PAGE_CACHE_SIZE - 1;
1919         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
1920
1921         start += (offset + blocksize -1) & ~(blocksize - 1);
1922         if (start > end)
1923                 return 0;
1924
1925         lock_extent(tree, start, end, GFP_NOFS);
1926         wait_on_extent_writeback(tree, start, end);
1927         clear_extent_bit(tree, start, end,
1928                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC,
1929                          1, 1, GFP_NOFS);
1930         return 0;
1931 }
1932 EXPORT_SYMBOL(extent_invalidatepage);
1933
1934 /*
1935  * simple commit_write call, set_range_dirty is used to mark both
1936  * the pages and the extent records as dirty
1937  */
1938 int extent_commit_write(struct extent_map_tree *tree,
1939                         struct inode *inode, struct page *page,
1940                         unsigned from, unsigned to)
1941 {
1942         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
1943
1944         set_page_extent_mapped(page);
1945         set_page_dirty(page);
1946
1947         if (pos > inode->i_size) {
1948                 i_size_write(inode, pos);
1949                 mark_inode_dirty(inode);
1950         }
1951         return 0;
1952 }
1953 EXPORT_SYMBOL(extent_commit_write);
1954
1955 int extent_prepare_write(struct extent_map_tree *tree,
1956                          struct inode *inode, struct page *page,
1957                          unsigned from, unsigned to, get_extent_t *get_extent)
1958 {
1959         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
1960         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
1961         u64 block_start;
1962         u64 orig_block_start;
1963         u64 block_end;
1964         u64 cur_end;
1965         struct extent_map *em;
1966         unsigned blocksize = 1 << inode->i_blkbits;
1967         size_t page_offset = 0;
1968         size_t block_off_start;
1969         size_t block_off_end;
1970         int err = 0;
1971         int iocount = 0;
1972         int ret = 0;
1973         int isnew;
1974
1975         set_page_extent_mapped(page);
1976
1977         block_start = (page_start + from) & ~((u64)blocksize - 1);
1978         block_end = (page_start + to - 1) | (blocksize - 1);
1979         orig_block_start = block_start;
1980
1981         lock_extent(tree, page_start, page_end, GFP_NOFS);
1982         while(block_start <= block_end) {
1983                 em = get_extent(inode, page, page_offset, block_start,
1984                                 block_end, 1);
1985                 if (IS_ERR(em) || !em) {
1986                         goto err;
1987                 }
1988                 cur_end = min(block_end, em->end);
1989                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
1990                 block_off_end = block_off_start + blocksize;
1991                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
1992
1993                 if (!PageUptodate(page) && isnew &&
1994                     (block_off_end > to || block_off_start < from)) {
1995                         void *kaddr;
1996
1997                         kaddr = kmap_atomic(page, KM_USER0);
1998                         if (block_off_end > to)
1999                                 memset(kaddr + to, 0, block_off_end - to);
2000                         if (block_off_start < from)
2001                                 memset(kaddr + block_off_start, 0,
2002                                        from - block_off_start);
2003                         flush_dcache_page(page);
2004                         kunmap_atomic(kaddr, KM_USER0);
2005                 }
2006                 if (!isnew && !PageUptodate(page) &&
2007                     (block_off_end > to || block_off_start < from) &&
2008                     !test_range_bit(tree, block_start, cur_end,
2009                                     EXTENT_UPTODATE, 1)) {
2010                         u64 sector;
2011                         u64 extent_offset = block_start - em->start;
2012                         size_t iosize;
2013                         sector = (em->block_start + extent_offset) >> 9;
2014                         iosize = (cur_end - block_start + blocksize - 1) &
2015                                 ~((u64)blocksize - 1);
2016                         /*
2017                          * we've already got the extent locked, but we
2018                          * need to split the state such that our end_bio
2019                          * handler can clear the lock.
2020                          */
2021                         set_extent_bit(tree, block_start,
2022                                        block_start + iosize - 1,
2023                                        EXTENT_LOCKED, 0, NULL, GFP_NOFS);
2024                         ret = submit_extent_page(READ, tree, page,
2025                                          sector, iosize, page_offset, em->bdev,
2026                                          NULL, 1,
2027                                          end_bio_extent_preparewrite);
2028                         iocount++;
2029                         block_start = block_start + iosize;
2030                 } else {
2031                         set_extent_uptodate(tree, block_start, cur_end,
2032                                             GFP_NOFS);
2033                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2034                         block_start = cur_end + 1;
2035                 }
2036                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2037                 free_extent_map(em);
2038         }
2039         if (iocount) {
2040                 wait_extent_bit(tree, orig_block_start,
2041                                 block_end, EXTENT_LOCKED);
2042         }
2043         check_page_uptodate(tree, page);
2044 err:
2045         /* FIXME, zero out newly allocated blocks on error */
2046         return err;
2047 }
2048 EXPORT_SYMBOL(extent_prepare_write);
2049
2050 /*
2051  * a helper for releasepage.  As long as there are no locked extents
2052  * in the range corresponding to the page, both state records and extent
2053  * map records are removed
2054  */
2055 int try_release_extent_mapping(struct extent_map_tree *tree, struct page *page)
2056 {
2057         struct extent_map *em;
2058         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2059         u64 end = start + PAGE_CACHE_SIZE - 1;
2060         u64 orig_start = start;
2061         int ret = 1;
2062
2063         while (start <= end) {
2064                 em = lookup_extent_mapping(tree, start, end);
2065                 if (!em || IS_ERR(em))
2066                         break;
2067                 if (!test_range_bit(tree, em->start, em->end,
2068                                     EXTENT_LOCKED, 0)) {
2069                         remove_extent_mapping(tree, em);
2070                         /* once for the rb tree */
2071                         free_extent_map(em);
2072                 }
2073                 start = em->end + 1;
2074                 /* once for us */
2075                 free_extent_map(em);
2076         }
2077         if (test_range_bit(tree, orig_start, end, EXTENT_LOCKED, 0))
2078                 ret = 0;
2079         else
2080                 clear_extent_bit(tree, orig_start, end, EXTENT_UPTODATE,
2081                                  1, 1, GFP_NOFS);
2082         return ret;
2083 }
2084 EXPORT_SYMBOL(try_release_extent_mapping);
2085
2086 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2087                 get_extent_t *get_extent)
2088 {
2089         struct inode *inode = mapping->host;
2090         u64 start = iblock << inode->i_blkbits;
2091         u64 end = start + (1 << inode->i_blkbits) - 1;
2092         sector_t sector = 0;
2093         struct extent_map *em;
2094
2095         em = get_extent(inode, NULL, 0, start, end, 0);
2096         if (!em || IS_ERR(em))
2097                 return 0;
2098
2099         if (em->block_start == EXTENT_MAP_INLINE ||
2100             em->block_start == EXTENT_MAP_HOLE)
2101                 goto out;
2102
2103         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2104 out:
2105         free_extent_map(em);
2106         return sector;
2107 }
2108
2109 static int add_lru(struct extent_map_tree *tree, struct extent_buffer *eb)
2110 {
2111         if (list_empty(&eb->lru)) {
2112                 extent_buffer_get(eb);
2113                 list_add(&eb->lru, &tree->buffer_lru);
2114                 tree->lru_size++;
2115                 if (tree->lru_size >= BUFFER_LRU_MAX) {
2116                         struct extent_buffer *rm;
2117                         rm = list_entry(tree->buffer_lru.prev,
2118                                         struct extent_buffer, lru);
2119                         tree->lru_size--;
2120                         list_del_init(&rm->lru);
2121                         free_extent_buffer(rm);
2122                 }
2123         } else
2124                 list_move(&eb->lru, &tree->buffer_lru);
2125         return 0;
2126 }
2127 static struct extent_buffer *find_lru(struct extent_map_tree *tree,
2128                                       u64 start, unsigned long len)
2129 {
2130         struct list_head *lru = &tree->buffer_lru;
2131         struct list_head *cur = lru->next;
2132         struct extent_buffer *eb;
2133
2134         if (list_empty(lru))
2135                 return NULL;
2136
2137         do {
2138                 eb = list_entry(cur, struct extent_buffer, lru);
2139                 if (eb->start == start && eb->len == len) {
2140                         extent_buffer_get(eb);
2141                         return eb;
2142                 }
2143                 cur = cur->next;
2144         } while (cur != lru);
2145         return NULL;
2146 }
2147
2148 static inline unsigned long num_extent_pages(u64 start, u64 len)
2149 {
2150         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
2151                 (start >> PAGE_CACHE_SHIFT);
2152 }
2153
2154 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2155                                               unsigned long i)
2156 {
2157         struct page *p;
2158         struct address_space *mapping;
2159
2160         if (i == 0)
2161                 return eb->first_page;
2162         i += eb->start >> PAGE_CACHE_SHIFT;
2163         mapping = eb->first_page->mapping;
2164         read_lock_irq(&mapping->tree_lock);
2165         p = radix_tree_lookup(&mapping->page_tree, i);
2166         read_unlock_irq(&mapping->tree_lock);
2167         return p;
2168 }
2169
2170 static struct extent_buffer *__alloc_extent_buffer(struct extent_map_tree *tree,
2171                                                    u64 start,
2172                                                    unsigned long len,
2173                                                    gfp_t mask)
2174 {
2175         struct extent_buffer *eb = NULL;
2176
2177         spin_lock(&tree->lru_lock);
2178         eb = find_lru(tree, start, len);
2179         spin_unlock(&tree->lru_lock);
2180         if (eb) {
2181                 return eb;
2182         }
2183
2184         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
2185         INIT_LIST_HEAD(&eb->lru);
2186         eb->start = start;
2187         eb->len = len;
2188         atomic_set(&eb->refs, 1);
2189
2190         return eb;
2191 }
2192
2193 static void __free_extent_buffer(struct extent_buffer *eb)
2194 {
2195         kmem_cache_free(extent_buffer_cache, eb);
2196 }
2197
2198 struct extent_buffer *alloc_extent_buffer(struct extent_map_tree *tree,
2199                                           u64 start, unsigned long len,
2200                                           struct page *page0,
2201                                           gfp_t mask)
2202 {
2203         unsigned long num_pages = num_extent_pages(start, len);
2204         unsigned long i;
2205         unsigned long index = start >> PAGE_CACHE_SHIFT;
2206         struct extent_buffer *eb;
2207         struct page *p;
2208         struct address_space *mapping = tree->mapping;
2209         int uptodate = 1;
2210
2211         eb = __alloc_extent_buffer(tree, start, len, mask);
2212         if (!eb || IS_ERR(eb))
2213                 return NULL;
2214
2215         if (eb->flags & EXTENT_BUFFER_FILLED)
2216                 goto lru_add;
2217
2218         if (page0) {
2219                 eb->first_page = page0;
2220                 i = 1;
2221                 index++;
2222                 page_cache_get(page0);
2223                 mark_page_accessed(page0);
2224                 set_page_extent_mapped(page0);
2225                 WARN_ON(!PageUptodate(page0));
2226                 set_page_private(page0, EXTENT_PAGE_PRIVATE_FIRST_PAGE |
2227                                  len << 2);
2228         } else {
2229                 i = 0;
2230         }
2231         for (; i < num_pages; i++, index++) {
2232                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
2233                 if (!p) {
2234                         WARN_ON(1);
2235                         goto fail;
2236                 }
2237                 set_page_extent_mapped(p);
2238                 mark_page_accessed(p);
2239                 if (i == 0) {
2240                         eb->first_page = p;
2241                         set_page_private(p, EXTENT_PAGE_PRIVATE_FIRST_PAGE |
2242                                          len << 2);
2243                 } else {
2244                         set_page_private(p, EXTENT_PAGE_PRIVATE);
2245                 }
2246                 if (!PageUptodate(p))
2247                         uptodate = 0;
2248                 unlock_page(p);
2249         }
2250         if (uptodate)
2251                 eb->flags |= EXTENT_UPTODATE;
2252         eb->flags |= EXTENT_BUFFER_FILLED;
2253
2254 lru_add:
2255         spin_lock(&tree->lru_lock);
2256         add_lru(tree, eb);
2257         spin_unlock(&tree->lru_lock);
2258         return eb;
2259
2260 fail:
2261         spin_lock(&tree->lru_lock);
2262         list_del_init(&eb->lru);
2263         spin_unlock(&tree->lru_lock);
2264         if (!atomic_dec_and_test(&eb->refs))
2265                 return NULL;
2266         for (index = 1; index < i; index++) {
2267                 page_cache_release(extent_buffer_page(eb, index));
2268         }
2269         if (i > 0)
2270                 page_cache_release(extent_buffer_page(eb, 0));
2271         __free_extent_buffer(eb);
2272         return NULL;
2273 }
2274 EXPORT_SYMBOL(alloc_extent_buffer);
2275
2276 struct extent_buffer *find_extent_buffer(struct extent_map_tree *tree,
2277                                          u64 start, unsigned long len,
2278                                           gfp_t mask)
2279 {
2280         unsigned long num_pages = num_extent_pages(start, len);
2281         unsigned long i;
2282         unsigned long index = start >> PAGE_CACHE_SHIFT;
2283         struct extent_buffer *eb;
2284         struct page *p;
2285         struct address_space *mapping = tree->mapping;
2286         int uptodate = 1;
2287
2288         eb = __alloc_extent_buffer(tree, start, len, mask);
2289         if (!eb || IS_ERR(eb))
2290                 return NULL;
2291
2292         if (eb->flags & EXTENT_BUFFER_FILLED)
2293                 goto lru_add;
2294
2295         for (i = 0; i < num_pages; i++, index++) {
2296                 p = find_lock_page(mapping, index);
2297                 if (!p) {
2298                         goto fail;
2299                 }
2300                 set_page_extent_mapped(p);
2301                 mark_page_accessed(p);
2302
2303                 if (i == 0) {
2304                         eb->first_page = p;
2305                         set_page_private(p, EXTENT_PAGE_PRIVATE_FIRST_PAGE |
2306                                          len << 2);
2307                 } else {
2308                         set_page_private(p, EXTENT_PAGE_PRIVATE);
2309                 }
2310
2311                 if (!PageUptodate(p))
2312                         uptodate = 0;
2313                 unlock_page(p);
2314         }
2315         if (uptodate)
2316                 eb->flags |= EXTENT_UPTODATE;
2317         eb->flags |= EXTENT_BUFFER_FILLED;
2318
2319 lru_add:
2320         spin_lock(&tree->lru_lock);
2321         add_lru(tree, eb);
2322         spin_unlock(&tree->lru_lock);
2323         return eb;
2324 fail:
2325         spin_lock(&tree->lru_lock);
2326         list_del_init(&eb->lru);
2327         spin_unlock(&tree->lru_lock);
2328         if (!atomic_dec_and_test(&eb->refs))
2329                 return NULL;
2330         for (index = 1; index < i; index++) {
2331                 page_cache_release(extent_buffer_page(eb, index));
2332         }
2333         if (i > 0)
2334                 page_cache_release(extent_buffer_page(eb, 0));
2335         __free_extent_buffer(eb);
2336         return NULL;
2337 }
2338 EXPORT_SYMBOL(find_extent_buffer);
2339
2340 void free_extent_buffer(struct extent_buffer *eb)
2341 {
2342         unsigned long i;
2343         unsigned long num_pages;
2344
2345         if (!eb)
2346                 return;
2347
2348         if (!atomic_dec_and_test(&eb->refs))
2349                 return;
2350
2351         WARN_ON(!list_empty(&eb->lru));
2352         num_pages = num_extent_pages(eb->start, eb->len);
2353
2354         for (i = 1; i < num_pages; i++) {
2355                 page_cache_release(extent_buffer_page(eb, i));
2356         }
2357         page_cache_release(extent_buffer_page(eb, 0));
2358         __free_extent_buffer(eb);
2359 }
2360 EXPORT_SYMBOL(free_extent_buffer);
2361
2362 int clear_extent_buffer_dirty(struct extent_map_tree *tree,
2363                               struct extent_buffer *eb)
2364 {
2365         int set;
2366         unsigned long i;
2367         unsigned long num_pages;
2368         struct page *page;
2369
2370         u64 start = eb->start;
2371         u64 end = start + eb->len - 1;
2372
2373         set = clear_extent_dirty(tree, start, end, GFP_NOFS);
2374         num_pages = num_extent_pages(eb->start, eb->len);
2375
2376         for (i = 0; i < num_pages; i++) {
2377                 page = extent_buffer_page(eb, i);
2378                 lock_page(page);
2379                 /*
2380                  * if we're on the last page or the first page and the
2381                  * block isn't aligned on a page boundary, do extra checks
2382                  * to make sure we don't clean page that is partially dirty
2383                  */
2384                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
2385                     ((i == num_pages - 1) &&
2386                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
2387                         start = (u64)page->index << PAGE_CACHE_SHIFT;
2388                         end  = start + PAGE_CACHE_SIZE - 1;
2389                         if (test_range_bit(tree, start, end,
2390                                            EXTENT_DIRTY, 0)) {
2391                                 unlock_page(page);
2392                                 continue;
2393                         }
2394                 }
2395                 clear_page_dirty_for_io(page);
2396                 unlock_page(page);
2397         }
2398         return 0;
2399 }
2400 EXPORT_SYMBOL(clear_extent_buffer_dirty);
2401
2402 int wait_on_extent_buffer_writeback(struct extent_map_tree *tree,
2403                                     struct extent_buffer *eb)
2404 {
2405         return wait_on_extent_writeback(tree, eb->start,
2406                                         eb->start + eb->len - 1);
2407 }
2408 EXPORT_SYMBOL(wait_on_extent_buffer_writeback);
2409
2410 int set_extent_buffer_dirty(struct extent_map_tree *tree,
2411                              struct extent_buffer *eb)
2412 {
2413         unsigned long i;
2414         unsigned long num_pages;
2415
2416         num_pages = num_extent_pages(eb->start, eb->len);
2417         for (i = 0; i < num_pages; i++) {
2418                 struct page *page = extent_buffer_page(eb, i);
2419                 /* writepage may need to do something special for the
2420                  * first page, we have to make sure page->private is
2421                  * properly set.  releasepage may drop page->private
2422                  * on us if the page isn't already dirty.
2423                  */
2424                 if (i == 0) {
2425                         lock_page(page);
2426                         set_page_private(page,
2427                                          EXTENT_PAGE_PRIVATE_FIRST_PAGE |
2428                                          eb->len << 2);
2429                 }
2430                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
2431                 if (i == 0)
2432                         unlock_page(page);
2433         }
2434         return set_extent_dirty(tree, eb->start,
2435                                 eb->start + eb->len - 1, GFP_NOFS);
2436 }
2437 EXPORT_SYMBOL(set_extent_buffer_dirty);
2438
2439 int set_extent_buffer_uptodate(struct extent_map_tree *tree,
2440                                 struct extent_buffer *eb)
2441 {
2442         unsigned long i;
2443         struct page *page;
2444         unsigned long num_pages;
2445
2446         num_pages = num_extent_pages(eb->start, eb->len);
2447
2448         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
2449                             GFP_NOFS);
2450         for (i = 0; i < num_pages; i++) {
2451                 page = extent_buffer_page(eb, i);
2452                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
2453                     ((i == num_pages - 1) &&
2454                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
2455                         check_page_uptodate(tree, page);
2456                         continue;
2457                 }
2458                 SetPageUptodate(page);
2459         }
2460         return 0;
2461 }
2462 EXPORT_SYMBOL(set_extent_buffer_uptodate);
2463
2464 int extent_buffer_uptodate(struct extent_map_tree *tree,
2465                              struct extent_buffer *eb)
2466 {
2467         if (eb->flags & EXTENT_UPTODATE)
2468                 return 1;
2469         return test_range_bit(tree, eb->start, eb->start + eb->len - 1,
2470                            EXTENT_UPTODATE, 1);
2471 }
2472 EXPORT_SYMBOL(extent_buffer_uptodate);
2473
2474 int read_extent_buffer_pages(struct extent_map_tree *tree,
2475                              struct extent_buffer *eb,
2476                              u64 start,
2477                              int wait)
2478 {
2479         unsigned long i;
2480         unsigned long start_i;
2481         struct page *page;
2482         int err;
2483         int ret = 0;
2484         unsigned long num_pages;
2485
2486         if (eb->flags & EXTENT_UPTODATE)
2487                 return 0;
2488
2489         if (0 && test_range_bit(tree, eb->start, eb->start + eb->len - 1,
2490                            EXTENT_UPTODATE, 1)) {
2491                 return 0;
2492         }
2493
2494         if (start) {
2495                 WARN_ON(start < eb->start);
2496                 start_i = (start >> PAGE_CACHE_SHIFT) -
2497                         (eb->start >> PAGE_CACHE_SHIFT);
2498         } else {
2499                 start_i = 0;
2500         }
2501
2502         num_pages = num_extent_pages(eb->start, eb->len);
2503         for (i = start_i; i < num_pages; i++) {
2504                 page = extent_buffer_page(eb, i);
2505                 if (PageUptodate(page)) {
2506                         continue;
2507                 }
2508                 if (!wait) {
2509                         if (TestSetPageLocked(page)) {
2510                                 continue;
2511                         }
2512                 } else {
2513                         lock_page(page);
2514                 }
2515                 if (!PageUptodate(page)) {
2516                         err = page->mapping->a_ops->readpage(NULL, page);
2517                         if (err) {
2518                                 ret = err;
2519                         }
2520                 } else {
2521                         unlock_page(page);
2522                 }
2523         }
2524
2525         if (ret || !wait) {
2526                 return ret;
2527         }
2528
2529         for (i = start_i; i < num_pages; i++) {
2530                 page = extent_buffer_page(eb, i);
2531                 wait_on_page_locked(page);
2532                 if (!PageUptodate(page)) {
2533                         ret = -EIO;
2534                 }
2535         }
2536         if (!ret)
2537                 eb->flags |= EXTENT_UPTODATE;
2538         return ret;
2539 }
2540 EXPORT_SYMBOL(read_extent_buffer_pages);
2541
2542 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
2543                         unsigned long start,
2544                         unsigned long len)
2545 {
2546         size_t cur;
2547         size_t offset;
2548         struct page *page;
2549         char *kaddr;
2550         char *dst = (char *)dstv;
2551         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2552         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2553         unsigned long num_pages = num_extent_pages(eb->start, eb->len);
2554
2555         WARN_ON(start > eb->len);
2556         WARN_ON(start + len > eb->start + eb->len);
2557
2558         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
2559
2560         while(len > 0) {
2561                 page = extent_buffer_page(eb, i);
2562                 if (!PageUptodate(page)) {
2563                         printk("page %lu not up to date i %lu, total %lu, len %lu\n", page->index, i, num_pages, eb->len);
2564                         WARN_ON(1);
2565                 }
2566                 WARN_ON(!PageUptodate(page));
2567
2568                 cur = min(len, (PAGE_CACHE_SIZE - offset));
2569                 kaddr = kmap_atomic(page, KM_USER1);
2570                 memcpy(dst, kaddr + offset, cur);
2571                 kunmap_atomic(kaddr, KM_USER1);
2572
2573                 dst += cur;
2574                 len -= cur;
2575                 offset = 0;
2576                 i++;
2577         }
2578 }
2579 EXPORT_SYMBOL(read_extent_buffer);
2580
2581 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
2582                                unsigned long min_len, char **token, char **map,
2583                                unsigned long *map_start,
2584                                unsigned long *map_len, int km)
2585 {
2586         size_t offset = start & (PAGE_CACHE_SIZE - 1);
2587         char *kaddr;
2588         struct page *p;
2589         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2590         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2591         unsigned long end_i = (start_offset + start + min_len - 1) >>
2592                 PAGE_CACHE_SHIFT;
2593
2594         if (i != end_i)
2595                 return -EINVAL;
2596
2597         if (i == 0) {
2598                 offset = start_offset;
2599                 *map_start = 0;
2600         } else {
2601                 offset = 0;
2602                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
2603         }
2604         if (start + min_len > eb->len) {
2605 printk("bad mapping eb start %Lu len %lu, wanted %lu %lu\n", eb->start, eb->len, start, min_len);
2606                 WARN_ON(1);
2607         }
2608
2609         p = extent_buffer_page(eb, i);
2610         WARN_ON(!PageUptodate(p));
2611         kaddr = kmap_atomic(p, km);
2612         *token = kaddr;
2613         *map = kaddr + offset;
2614         *map_len = PAGE_CACHE_SIZE - offset;
2615         return 0;
2616 }
2617 EXPORT_SYMBOL(map_private_extent_buffer);
2618
2619 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
2620                       unsigned long min_len,
2621                       char **token, char **map,
2622                       unsigned long *map_start,
2623                       unsigned long *map_len, int km)
2624 {
2625         int err;
2626         int save = 0;
2627         if (eb->map_token) {
2628                 unmap_extent_buffer(eb, eb->map_token, km);
2629                 eb->map_token = NULL;
2630                 save = 1;
2631         }
2632         err = map_private_extent_buffer(eb, start, min_len, token, map,
2633                                        map_start, map_len, km);
2634         if (!err && save) {
2635                 eb->map_token = *token;
2636                 eb->kaddr = *map;
2637                 eb->map_start = *map_start;
2638                 eb->map_len = *map_len;
2639         }
2640         return err;
2641 }
2642 EXPORT_SYMBOL(map_extent_buffer);
2643
2644 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
2645 {
2646         kunmap_atomic(token, km);
2647 }
2648 EXPORT_SYMBOL(unmap_extent_buffer);
2649
2650 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
2651                           unsigned long start,
2652                           unsigned long len)
2653 {
2654         size_t cur;
2655         size_t offset;
2656         struct page *page;
2657         char *kaddr;
2658         char *ptr = (char *)ptrv;
2659         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2660         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2661         int ret = 0;
2662
2663         WARN_ON(start > eb->len);
2664         WARN_ON(start + len > eb->start + eb->len);
2665
2666         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
2667
2668         while(len > 0) {
2669                 page = extent_buffer_page(eb, i);
2670                 WARN_ON(!PageUptodate(page));
2671
2672                 cur = min(len, (PAGE_CACHE_SIZE - offset));
2673
2674                 kaddr = kmap_atomic(page, KM_USER0);
2675                 ret = memcmp(ptr, kaddr + offset, cur);
2676                 kunmap_atomic(kaddr, KM_USER0);
2677                 if (ret)
2678                         break;
2679
2680                 ptr += cur;
2681                 len -= cur;
2682                 offset = 0;
2683                 i++;
2684         }
2685         return ret;
2686 }
2687 EXPORT_SYMBOL(memcmp_extent_buffer);
2688
2689 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
2690                          unsigned long start, unsigned long len)
2691 {
2692         size_t cur;
2693         size_t offset;
2694         struct page *page;
2695         char *kaddr;
2696         char *src = (char *)srcv;
2697         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2698         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2699
2700         WARN_ON(start > eb->len);
2701         WARN_ON(start + len > eb->start + eb->len);
2702
2703         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
2704
2705         while(len > 0) {
2706                 page = extent_buffer_page(eb, i);
2707                 WARN_ON(!PageUptodate(page));
2708
2709                 cur = min(len, PAGE_CACHE_SIZE - offset);
2710                 kaddr = kmap_atomic(page, KM_USER1);
2711                 memcpy(kaddr + offset, src, cur);
2712                 kunmap_atomic(kaddr, KM_USER1);
2713
2714                 src += cur;
2715                 len -= cur;
2716                 offset = 0;
2717                 i++;
2718         }
2719 }
2720 EXPORT_SYMBOL(write_extent_buffer);
2721
2722 void memset_extent_buffer(struct extent_buffer *eb, char c,
2723                           unsigned long start, unsigned long len)
2724 {
2725         size_t cur;
2726         size_t offset;
2727         struct page *page;
2728         char *kaddr;
2729         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2730         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2731
2732         WARN_ON(start > eb->len);
2733         WARN_ON(start + len > eb->start + eb->len);
2734
2735         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
2736
2737         while(len > 0) {
2738                 page = extent_buffer_page(eb, i);
2739                 WARN_ON(!PageUptodate(page));
2740
2741                 cur = min(len, PAGE_CACHE_SIZE - offset);
2742                 kaddr = kmap_atomic(page, KM_USER0);
2743                 memset(kaddr + offset, c, cur);
2744                 kunmap_atomic(kaddr, KM_USER0);
2745
2746                 len -= cur;
2747                 offset = 0;
2748                 i++;
2749         }
2750 }
2751 EXPORT_SYMBOL(memset_extent_buffer);
2752
2753 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
2754                         unsigned long dst_offset, unsigned long src_offset,
2755                         unsigned long len)
2756 {
2757         u64 dst_len = dst->len;
2758         size_t cur;
2759         size_t offset;
2760         struct page *page;
2761         char *kaddr;
2762         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
2763         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
2764
2765         WARN_ON(src->len != dst_len);
2766
2767         offset = (start_offset + dst_offset) &
2768                 ((unsigned long)PAGE_CACHE_SIZE - 1);
2769
2770         while(len > 0) {
2771                 page = extent_buffer_page(dst, i);
2772                 WARN_ON(!PageUptodate(page));
2773
2774                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
2775
2776                 kaddr = kmap_atomic(page, KM_USER0);
2777                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
2778                 kunmap_atomic(kaddr, KM_USER0);
2779
2780                 src_offset += cur;
2781                 len -= cur;
2782                 offset = 0;
2783                 i++;
2784         }
2785 }
2786 EXPORT_SYMBOL(copy_extent_buffer);
2787
2788 static void move_pages(struct page *dst_page, struct page *src_page,
2789                        unsigned long dst_off, unsigned long src_off,
2790                        unsigned long len)
2791 {
2792         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
2793         if (dst_page == src_page) {
2794                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
2795         } else {
2796                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
2797                 char *p = dst_kaddr + dst_off + len;
2798                 char *s = src_kaddr + src_off + len;
2799
2800                 while (len--)
2801                         *--p = *--s;
2802
2803                 kunmap_atomic(src_kaddr, KM_USER1);
2804         }
2805         kunmap_atomic(dst_kaddr, KM_USER0);
2806 }
2807
2808 static void copy_pages(struct page *dst_page, struct page *src_page,
2809                        unsigned long dst_off, unsigned long src_off,
2810                        unsigned long len)
2811 {
2812         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
2813         char *src_kaddr;
2814
2815         if (dst_page != src_page)
2816                 src_kaddr = kmap_atomic(src_page, KM_USER1);
2817         else
2818                 src_kaddr = dst_kaddr;
2819
2820         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
2821         kunmap_atomic(dst_kaddr, KM_USER0);
2822         if (dst_page != src_page)
2823                 kunmap_atomic(src_kaddr, KM_USER1);
2824 }
2825
2826 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
2827                            unsigned long src_offset, unsigned long len)
2828 {
2829         size_t cur;
2830         size_t dst_off_in_page;
2831         size_t src_off_in_page;
2832         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
2833         unsigned long dst_i;
2834         unsigned long src_i;
2835
2836         if (src_offset + len > dst->len) {
2837                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
2838                        src_offset, len, dst->len);
2839                 BUG_ON(1);
2840         }
2841         if (dst_offset + len > dst->len) {
2842                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
2843                        dst_offset, len, dst->len);
2844                 BUG_ON(1);
2845         }
2846
2847         while(len > 0) {
2848                 dst_off_in_page = (start_offset + dst_offset) &
2849                         ((unsigned long)PAGE_CACHE_SIZE - 1);
2850                 src_off_in_page = (start_offset + src_offset) &
2851                         ((unsigned long)PAGE_CACHE_SIZE - 1);
2852
2853                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
2854                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
2855
2856                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
2857                                                src_off_in_page));
2858                 cur = min_t(unsigned long, cur,
2859                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
2860
2861                 copy_pages(extent_buffer_page(dst, dst_i),
2862                            extent_buffer_page(dst, src_i),
2863                            dst_off_in_page, src_off_in_page, cur);
2864
2865                 src_offset += cur;
2866                 dst_offset += cur;
2867                 len -= cur;
2868         }
2869 }
2870 EXPORT_SYMBOL(memcpy_extent_buffer);
2871
2872 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
2873                            unsigned long src_offset, unsigned long len)
2874 {
2875         size_t cur;
2876         size_t dst_off_in_page;
2877         size_t src_off_in_page;
2878         unsigned long dst_end = dst_offset + len - 1;
2879         unsigned long src_end = src_offset + len - 1;
2880         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
2881         unsigned long dst_i;
2882         unsigned long src_i;
2883
2884         if (src_offset + len > dst->len) {
2885                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
2886                        src_offset, len, dst->len);
2887                 BUG_ON(1);
2888         }
2889         if (dst_offset + len > dst->len) {
2890                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
2891                        dst_offset, len, dst->len);
2892                 BUG_ON(1);
2893         }
2894         if (dst_offset < src_offset) {
2895                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
2896                 return;
2897         }
2898         while(len > 0) {
2899                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
2900                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
2901
2902                 dst_off_in_page = (start_offset + dst_end) &
2903                         ((unsigned long)PAGE_CACHE_SIZE - 1);
2904                 src_off_in_page = (start_offset + src_end) &
2905                         ((unsigned long)PAGE_CACHE_SIZE - 1);
2906
2907                 cur = min_t(unsigned long, len, src_off_in_page + 1);
2908                 cur = min(cur, dst_off_in_page + 1);
2909                 move_pages(extent_buffer_page(dst, dst_i),
2910                            extent_buffer_page(dst, src_i),
2911                            dst_off_in_page - cur + 1,
2912                            src_off_in_page - cur + 1, cur);
2913
2914                 dst_end -= cur;
2915                 src_end -= cur;
2916                 len -= cur;
2917         }
2918 }
2919 EXPORT_SYMBOL(memmove_extent_buffer);