SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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
  46 int __init extent_map_init(void)
  47 {
  48         extent_map_cache = btrfs_cache_create("extent_map",
  49                                             sizeof(struct extent_map), 0,
  50                                             NULL);
  51         if (!extent_map_cache)
  52                 return -ENOMEM;
  53         extent_state_cache = btrfs_cache_create("extent_state",
  54                                             sizeof(struct extent_state), 0,
  55                                             NULL);
  56         if (!extent_state_cache)
  57                 goto free_map_cache;
  58         extent_buffer_cache = btrfs_cache_create("extent_buffers",
  59                                             sizeof(struct extent_buffer), 0,
  60                                             NULL);
  61         if (!extent_buffer_cache)
  62                 goto free_state_cache;
  63         return 0;
  64
  65 free_state_cache:
  66         kmem_cache_destroy(extent_state_cache);
  67 free_map_cache:
  68         kmem_cache_destroy(extent_map_cache);
  69         return -ENOMEM;
  70 }
  71
  72 void extent_map_exit(void)
  73 {
  74         struct extent_state *state;
  75
  76         while (!list_empty(&states)) {
  77                 state = list_entry(states.next, struct extent_state, list);
  78                 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));
  79                 list_del(&state->list);
  80                 kmem_cache_free(extent_state_cache, state);
  81
  82         }
  83
  84         if (extent_map_cache)
  85                 kmem_cache_destroy(extent_map_cache);
  86         if (extent_state_cache)
  87                 kmem_cache_destroy(extent_state_cache);
  88         if (extent_buffer_cache)
  89                 kmem_cache_destroy(extent_buffer_cache);
  90 }
  91
  92 void extent_map_tree_init(struct extent_map_tree *tree,
  93                           struct address_space *mapping, gfp_t mask)
  94 {
  95         tree->map.rb_node = NULL;
  96         tree->state.rb_node = NULL;
  97         tree->ops = NULL;
  98         tree->dirty_bytes = 0;
  99         rwlock_init(&tree->lock);
 100         spin_lock_init(&tree->lru_lock);
 101         tree->mapping = mapping;
 102         INIT_LIST_HEAD(&tree->buffer_lru);
 103         tree->lru_size = 0;
 104 }
 105 EXPORT_SYMBOL(extent_map_tree_init);
 106
 107 void extent_map_tree_empty_lru(struct extent_map_tree *tree)
 108 {
 109         struct extent_buffer *eb;
 110         while(!list_empty(&tree->buffer_lru)) {
 111                 eb = list_entry(tree->buffer_lru.next, struct extent_buffer,
 112                                 lru);
 113                 list_del_init(&eb->lru);
 114                 free_extent_buffer(eb);
 115         }
 116 }
 117 EXPORT_SYMBOL(extent_map_tree_empty_lru);
 118
 119 struct extent_map *alloc_extent_map(gfp_t mask)
 120 {
 121         struct extent_map *em;
 122         em = kmem_cache_alloc(extent_map_cache, mask);
 123         if (!em || IS_ERR(em))
 124                 return em;
 125         em->in_tree = 0;
 126         atomic_set(&em->refs, 1);
 127         return em;
 128 }
 129 EXPORT_SYMBOL(alloc_extent_map);
 130
 131 void free_extent_map(struct extent_map *em)
 132 {
 133         if (!em)
 134                 return;
 135         if (atomic_dec_and_test(&em->refs)) {
 136                 WARN_ON(em->in_tree);
 137                 kmem_cache_free(extent_map_cache, em);
 138         }
 139 }
 140 EXPORT_SYMBOL(free_extent_map);
 141
 142
 143 struct extent_state *alloc_extent_state(gfp_t mask)
 144 {
 145         struct extent_state *state;
 146         unsigned long flags;
 147
 148         state = kmem_cache_alloc(extent_state_cache, mask);
 149         if (!state || IS_ERR(state))
 150                 return state;
 151         state->state = 0;
 152         state->in_tree = 0;
 153         state->private = 0;
 154
 155         spin_lock_irqsave(&state_lock, flags);
 156         list_add(&state->list, &states);
 157         spin_unlock_irqrestore(&state_lock, flags);
 158
 159         atomic_set(&state->refs, 1);
 160         init_waitqueue_head(&state->wq);
 161         return state;
 162 }
 163 EXPORT_SYMBOL(alloc_extent_state);
 164
 165 void free_extent_state(struct extent_state *state)
 166 {
 167         unsigned long flags;
 168         if (!state)
 169                 return;
 170         if (atomic_dec_and_test(&state->refs)) {
 171                 WARN_ON(state->in_tree);
 172                 spin_lock_irqsave(&state_lock, flags);
 173                 list_del(&state->list);
 174                 spin_unlock_irqrestore(&state_lock, flags);
 175                 kmem_cache_free(extent_state_cache, state);
 176         }
 177 }
 178 EXPORT_SYMBOL(free_extent_state);
 179
 180 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 181                                    struct rb_node *node)
 182 {
 183         struct rb_node ** p = &root->rb_node;
 184         struct rb_node * parent = NULL;
 185         struct tree_entry *entry;
 186
 187         while(*p) {
 188                 parent = *p;
 189                 entry = rb_entry(parent, struct tree_entry, rb_node);
 190
 191                 if (offset < entry->start)
 192                         p = &(*p)->rb_left;
 193                 else if (offset > entry->end)
 194                         p = &(*p)->rb_right;
 195                 else
 196                         return parent;
 197         }
 198
 199         entry = rb_entry(node, struct tree_entry, rb_node);
 200         entry->in_tree = 1;
 201         rb_link_node(node, parent, p);
 202         rb_insert_color(node, root);
 203         return NULL;
 204 }
 205
 206 static struct rb_node *__tree_search(struct rb_root *root, u64 offset,
 207                                    struct rb_node **prev_ret)
 208 {
 209         struct rb_node * n = root->rb_node;
 210         struct rb_node *prev = NULL;
 211         struct tree_entry *entry;
 212         struct tree_entry *prev_entry = NULL;
 213
 214         while(n) {
 215                 entry = rb_entry(n, struct tree_entry, rb_node);
 216                 prev = n;
 217                 prev_entry = entry;
 218
 219                 if (offset < entry->start)
 220                         n = n->rb_left;
 221                 else if (offset > entry->end)
 222                         n = n->rb_right;
 223                 else
 224                         return n;
 225         }
 226         if (!prev_ret)
 227                 return NULL;
 228         while(prev && offset > prev_entry->end) {
 229                 prev = rb_next(prev);
 230                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 231         }
 232         *prev_ret = prev;
 233         return NULL;
 234 }
 235
 236 static inline struct rb_node *tree_search(struct rb_root *root, u64 offset)
 237 {
 238         struct rb_node *prev;
 239         struct rb_node *ret;
 240         ret = __tree_search(root, offset, &prev);
 241         if (!ret)
 242                 return prev;
 243         return ret;
 244 }
 245
 246 static int tree_delete(struct rb_root *root, u64 offset)
 247 {
 248         struct rb_node *node;
 249         struct tree_entry *entry;
 250
 251         node = __tree_search(root, offset, NULL);
 252         if (!node)
 253                 return -ENOENT;
 254         entry = rb_entry(node, struct tree_entry, rb_node);
 255         entry->in_tree = 0;
 256         rb_erase(node, root);
 257         return 0;
 258 }
 259
 260 /*
 261  * add_extent_mapping tries a simple backward merge with existing
 262  * mappings.  The extent_map struct passed in will be inserted into
 263  * the tree directly (no copies made, just a reference taken).
 264  */
 265 int add_extent_mapping(struct extent_map_tree *tree,
 266                        struct extent_map *em)
 267 {
 268         int ret = 0;
 269         struct extent_map *prev = NULL;
 270         struct rb_node *rb;
 271
 272         write_lock_irq(&tree->lock);
 273         rb = tree_insert(&tree->map, em->end, &em->rb_node);
 274         if (rb) {
 275                 prev = rb_entry(rb, struct extent_map, rb_node);
 276                 printk("found extent map %Lu %Lu on insert of %Lu %Lu\n", prev->start, prev->end, em->start, em->end);
 277                 ret = -EEXIST;
 278                 goto out;
 279         }
 280         atomic_inc(&em->refs);
 281         if (em->start != 0) {
 282                 rb = rb_prev(&em->rb_node);
 283                 if (rb)
 284                         prev = rb_entry(rb, struct extent_map, rb_node);
 285                 if (prev && prev->end + 1 == em->start &&
 286                     ((em->block_start == EXTENT_MAP_HOLE &&
 287                       prev->block_start == EXTENT_MAP_HOLE) ||
 288                      (em->block_start == EXTENT_MAP_INLINE &&
 289                       prev->block_start == EXTENT_MAP_INLINE) ||
 290                      (em->block_start == EXTENT_MAP_DELALLOC &&
 291                       prev->block_start == EXTENT_MAP_DELALLOC) ||
 292                      (em->block_start < EXTENT_MAP_DELALLOC - 1 &&
 293                       em->block_start == prev->block_end + 1))) {
 294                         em->start = prev->start;
 295                         em->block_start = prev->block_start;
 296                         rb_erase(&prev->rb_node, &tree->map);
 297                         prev->in_tree = 0;
 298                         free_extent_map(prev);
 299                 }
 300          }
 301 out:
 302         write_unlock_irq(&tree->lock);
 303         return ret;
 304 }
 305 EXPORT_SYMBOL(add_extent_mapping);
 306
 307 /*
 308  * lookup_extent_mapping returns the first extent_map struct in the
 309  * tree that intersects the [start, end] (inclusive) range.  There may
 310  * be additional objects in the tree that intersect, so check the object
 311  * returned carefully to make sure you don't need additional lookups.
 312  */
 313 struct extent_map *lookup_extent_mapping(struct extent_map_tree *tree,
 314                                          u64 start, u64 end)
 315 {
 316         struct extent_map *em;
 317         struct rb_node *rb_node;
 318
 319         read_lock_irq(&tree->lock);
 320         rb_node = tree_search(&tree->map, start);
 321         if (!rb_node) {
 322                 em = NULL;
 323                 goto out;
 324         }
 325         if (IS_ERR(rb_node)) {
 326                 em = ERR_PTR(PTR_ERR(rb_node));
 327                 goto out;
 328         }
 329         em = rb_entry(rb_node, struct extent_map, rb_node);
 330         if (em->end < start || em->start > end) {
 331                 em = NULL;
 332                 goto out;
 333         }
 334         atomic_inc(&em->refs);
 335 out:
 336         read_unlock_irq(&tree->lock);
 337         return em;
 338 }
 339 EXPORT_SYMBOL(lookup_extent_mapping);
 340
 341 /*
 342  * removes an extent_map struct from the tree.  No reference counts are
 343  * dropped, and no checks are done to  see if the range is in use
 344  */
 345 int remove_extent_mapping(struct extent_map_tree *tree, struct extent_map *em)
 346 {
 347         int ret;
 348
 349         write_lock_irq(&tree->lock);
 350         ret = tree_delete(&tree->map, em->end);
 351         write_unlock_irq(&tree->lock);
 352         return ret;
 353 }
 354 EXPORT_SYMBOL(remove_extent_mapping);
 355
 356 /*
 357  * utility function to look for merge candidates inside a given range.
 358  * Any extents with matching state are merged together into a single
 359  * extent in the tree.  Extents with EXTENT_IO in their state field
 360  * are not merged because the end_io handlers need to be able to do
 361  * operations on them without sleeping (or doing allocations/splits).
 362  *
 363  * This should be called with the tree lock held.
 364  */
 365 static int merge_state(struct extent_map_tree *tree,
 366                        struct extent_state *state)
 367 {
 368         struct extent_state *other;
 369         struct rb_node *other_node;
 370
 371         if (state->state & EXTENT_IOBITS)
 372                 return 0;
 373
 374         other_node = rb_prev(&state->rb_node);
 375         if (other_node) {
 376                 other = rb_entry(other_node, struct extent_state, rb_node);
 377                 if (other->end == state->start - 1 &&
 378                     other->state == state->state) {
 379                         state->start = other->start;
 380                         other->in_tree = 0;
 381                         rb_erase(&other->rb_node, &tree->state);
 382                         free_extent_state(other);
 383                 }
 384         }
 385         other_node = rb_next(&state->rb_node);
 386         if (other_node) {
 387                 other = rb_entry(other_node, struct extent_state, rb_node);
 388                 if (other->start == state->end + 1 &&
 389                     other->state == state->state) {
 390                         other->start = state->start;
 391                         state->in_tree = 0;
 392                         rb_erase(&state->rb_node, &tree->state);
 393                         free_extent_state(state);
 394                 }
 395         }
 396         return 0;
 397 }
 398
 399 /*
 400  * insert an extent_state struct into the tree.  'bits' are set on the
 401  * struct before it is inserted.
 402  *
 403  * This may return -EEXIST if the extent is already there, in which case the
 404  * state struct is freed.
 405  *
 406  * The tree lock is not taken internally.  This is a utility function and
 407  * probably isn't what you want to call (see set/clear_extent_bit).
 408  */
 409 static int insert_state(struct extent_map_tree *tree,
 410                         struct extent_state *state, u64 start, u64 end,
 411                         int bits)
 412 {
 413         struct rb_node *node;
 414
 415         if (end < start) {
 416                 printk("end < start %Lu %Lu\n", end, start);
 417                 WARN_ON(1);
 418         }
 419         if (bits & EXTENT_DIRTY)
 420                 tree->dirty_bytes += end - start + 1;
 421         state->state |= bits;
 422         state->start = start;
 423         state->end = end;
 424         node = tree_insert(&tree->state, end, &state->rb_node);
 425         if (node) {
 426                 struct extent_state *found;
 427                 found = rb_entry(node, struct extent_state, rb_node);
 428                 printk("found node %Lu %Lu on insert of %Lu %Lu\n", found->start, found->end, start, end);
 429                 free_extent_state(state);
 430                 return -EEXIST;
 431         }
 432         merge_state(tree, state);
 433         return 0;
 434 }
 435
 436 /*
 437  * split a given extent state struct in two, inserting the preallocated
 438  * struct 'prealloc' as the newly created second half.  'split' indicates an
 439  * offset inside 'orig' where it should be split.
 440  *
 441  * Before calling,
 442  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 443  * are two extent state structs in the tree:
 444  * prealloc: [orig->start, split - 1]
 445  * orig: [ split, orig->end ]
 446  *
 447  * The tree locks are not taken by this function. They need to be held
 448  * by the caller.
 449  */
 450 static int split_state(struct extent_map_tree *tree, struct extent_state *orig,
 451                        struct extent_state *prealloc, u64 split)
 452 {
 453         struct rb_node *node;
 454         prealloc->start = orig->start;
 455         prealloc->end = split - 1;
 456         prealloc->state = orig->state;
 457         orig->start = split;
 458
 459         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 460         if (node) {
 461                 struct extent_state *found;
 462                 found = rb_entry(node, struct extent_state, rb_node);
 463                 printk("found node %Lu %Lu on insert of %Lu %Lu\n", found->start, found->end, prealloc->start, prealloc->end);
 464                 free_extent_state(prealloc);
 465                 return -EEXIST;
 466         }
 467         return 0;
 468 }
 469
 470 /*
 471  * utility function to clear some bits in an extent state struct.
 472  * it will optionally wake up any one waiting on this state (wake == 1), or
 473  * forcibly remove the state from the tree (delete == 1).
 474  *
 475  * If no bits are set on the state struct after clearing things, the
 476  * struct is freed and removed from the tree
 477  */
 478 static int clear_state_bit(struct extent_map_tree *tree,
 479                             struct extent_state *state, int bits, int wake,
 480                             int delete)
 481 {
 482         int ret = state->state & bits;
 483
 484         if ((bits & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 485                 u64 range = state->end - state->start + 1;
 486                 WARN_ON(range > tree->dirty_bytes);
 487                 tree->dirty_bytes -= range;
 488         }
 489         state->state &= ~bits;
 490         if (wake)
 491                 wake_up(&state->wq);
 492         if (delete || state->state == 0) {
 493                 if (state->in_tree) {
 494                         rb_erase(&state->rb_node, &tree->state);
 495                         state->in_tree = 0;
 496                         free_extent_state(state);
 497                 } else {
 498                         WARN_ON(1);
 499                 }
 500         } else {
 501                 merge_state(tree, state);
 502         }
 503         return ret;
 504 }
 505
 506 /*
 507  * clear some bits on a range in the tree.  This may require splitting
 508  * or inserting elements in the tree, so the gfp mask is used to
 509  * indicate which allocations or sleeping are allowed.
 510  *
 511  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 512  * the given range from the tree regardless of state (ie for truncate).
 513  *
 514  * the range [start, end] is inclusive.
 515  *
 516  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 517  * bits were already set, or zero if none of the bits were already set.
 518  */
 519 int clear_extent_bit(struct extent_map_tree *tree, u64 start, u64 end,
 520                      int bits, int wake, int delete, gfp_t mask)
 521 {
 522         struct extent_state *state;
 523         struct extent_state *prealloc = NULL;
 524         struct rb_node *node;
 525         unsigned long flags;
 526         int err;
 527         int set = 0;
 528
 529 again:
 530         if (!prealloc && (mask & __GFP_WAIT)) {
 531                 prealloc = alloc_extent_state(mask);
 532                 if (!prealloc)
 533                         return -ENOMEM;
 534         }
 535
 536         write_lock_irqsave(&tree->lock, flags);
 537         /*
 538          * this search will find the extents that end after
 539          * our range starts
 540          */
 541         node = tree_search(&tree->state, start);
 542         if (!node)
 543                 goto out;
 544         state = rb_entry(node, struct extent_state, rb_node);
 545         if (state->start > end)
 546                 goto out;
 547         WARN_ON(state->end < start);
 548
 549         /*
 550          *     | ---- desired range ---- |
 551          *  | state | or
 552          *  | ------------- state -------------- |
 553          *
 554          * We need to split the extent we found, and may flip
 555          * bits on second half.
 556          *
 557          * If the extent we found extends past our range, we
 558          * just split and search again.  It'll get split again
 559          * the next time though.
 560          *
 561          * If the extent we found is inside our range, we clear
 562          * the desired bit on it.
 563          */
 564
 565         if (state->start < start) {
 566                 err = split_state(tree, state, prealloc, start);
 567                 BUG_ON(err == -EEXIST);
 568                 prealloc = NULL;
 569                 if (err)
 570                         goto out;
 571                 if (state->end <= end) {
 572                         start = state->end + 1;
 573                         set |= clear_state_bit(tree, state, bits,
 574                                         wake, delete);
 575                 } else {
 576                         start = state->start;
 577                 }
 578                 goto search_again;
 579         }
 580         /*
 581          * | ---- desired range ---- |
 582          *                        | state |
 583          * We need to split the extent, and clear the bit
 584          * on the first half
 585          */
 586         if (state->start <= end && state->end > end) {
 587                 err = split_state(tree, state, prealloc, end + 1);
 588                 BUG_ON(err == -EEXIST);
 589
 590                 if (wake)
 591                         wake_up(&state->wq);
 592                 set |= clear_state_bit(tree, prealloc, bits,
 593                                        wake, delete);
 594                 prealloc = NULL;
 595                 goto out;
 596         }
 597
 598         start = state->end + 1;
 599         set |= clear_state_bit(tree, state, bits, wake, delete);
 600         goto search_again;
 601
 602 out:
 603         write_unlock_irqrestore(&tree->lock, flags);
 604         if (prealloc)
 605                 free_extent_state(prealloc);
 606
 607         return set;
 608
 609 search_again:
 610         if (start > end)
 611                 goto out;
 612         write_unlock_irqrestore(&tree->lock, flags);
 613         if (mask & __GFP_WAIT)
 614                 cond_resched();
 615         goto again;
 616 }
 617 EXPORT_SYMBOL(clear_extent_bit);
 618
 619 static int wait_on_state(struct extent_map_tree *tree,
 620                          struct extent_state *state)
 621 {
 622         DEFINE_WAIT(wait);
 623         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 624         read_unlock_irq(&tree->lock);
 625         schedule();
 626         read_lock_irq(&tree->lock);
 627         finish_wait(&state->wq, &wait);
 628         return 0;
 629 }
 630
 631 /*
 632  * waits for one or more bits to clear on a range in the state tree.
 633  * The range [start, end] is inclusive.
 634  * The tree lock is taken by this function
 635  */
 636 int wait_extent_bit(struct extent_map_tree *tree, u64 start, u64 end, int bits)
 637 {
 638         struct extent_state *state;
 639         struct rb_node *node;
 640
 641         read_lock_irq(&tree->lock);
 642 again:
 643         while (1) {
 644                 /*
 645                  * this search will find all the extents that end after
 646                  * our range starts
 647                  */
 648                 node = tree_search(&tree->state, start);
 649                 if (!node)
 650                         break;
 651
 652                 state = rb_entry(node, struct extent_state, rb_node);
 653
 654                 if (state->start > end)
 655                         goto out;
 656
 657                 if (state->state & bits) {
 658                         start = state->start;
 659                         atomic_inc(&state->refs);
 660                         wait_on_state(tree, state);
 661                         free_extent_state(state);
 662                         goto again;
 663                 }
 664                 start = state->end + 1;
 665
 666                 if (start > end)
 667                         break;
 668
 669                 if (need_resched()) {
 670                         read_unlock_irq(&tree->lock);
 671                         cond_resched();
 672                         read_lock_irq(&tree->lock);
 673                 }
 674         }
 675 out:
 676         read_unlock_irq(&tree->lock);
 677         return 0;
 678 }
 679 EXPORT_SYMBOL(wait_extent_bit);
 680
 681 static void set_state_bits(struct extent_map_tree *tree,
 682                            struct extent_state *state,
 683                            int bits)
 684 {
 685         if ((bits & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 686                 u64 range = state->end - state->start + 1;
 687                 tree->dirty_bytes += range;
 688         }
 689         state->state |= bits;
 690 }
 691
 692 /*
 693  * set some bits on a range in the tree.  This may require allocations
 694  * or sleeping, so the gfp mask is used to indicate what is allowed.
 695  *
 696  * If 'exclusive' == 1, this will fail with -EEXIST if some part of the
 697  * range already has the desired bits set.  The start of the existing
 698  * range is returned in failed_start in this case.
 699  *
 700  * [start, end] is inclusive
 701  * This takes the tree lock.
 702  */
 703 int set_extent_bit(struct extent_map_tree *tree, u64 start, u64 end, int bits,
 704                    int exclusive, u64 *failed_start, gfp_t mask)
 705 {
 706         struct extent_state *state;
 707         struct extent_state *prealloc = NULL;
 708         struct rb_node *node;
 709         unsigned long flags;
 710         int err = 0;
 711         int set;
 712         u64 last_start;
 713         u64 last_end;
 714 again:
 715         if (!prealloc && (mask & __GFP_WAIT)) {
 716                 prealloc = alloc_extent_state(mask);
 717                 if (!prealloc)
 718                         return -ENOMEM;
 719         }
 720
 721         write_lock_irqsave(&tree->lock, flags);
 722         /*
 723          * this search will find all the extents that end after
 724          * our range starts.
 725          */
 726         node = tree_search(&tree->state, start);
 727         if (!node) {
 728                 err = insert_state(tree, prealloc, start, end, bits);
 729                 prealloc = NULL;
 730                 BUG_ON(err == -EEXIST);
 731                 goto out;
 732         }
 733
 734         state = rb_entry(node, struct extent_state, rb_node);
 735         last_start = state->start;
 736         last_end = state->end;
 737
 738         /*
 739          * | ---- desired range ---- |
 740          * | state |
 741          *
 742          * Just lock what we found and keep going
 743          */
 744         if (state->start == start && state->end <= end) {
 745                 set = state->state & bits;
 746                 if (set && exclusive) {
 747                         *failed_start = state->start;
 748                         err = -EEXIST;
 749                         goto out;
 750                 }
 751                 set_state_bits(tree, state, bits);
 752                 start = state->end + 1;
 753                 merge_state(tree, state);
 754                 goto search_again;
 755         }
 756
 757         /*
 758          *     | ---- desired range ---- |
 759          * | state |
 760          *   or
 761          * | ------------- state -------------- |
 762          *
 763          * We need to split the extent we found, and may flip bits on
 764          * second half.
 765          *
 766          * If the extent we found extends past our
 767          * range, we just split and search again.  It'll get split
 768          * again the next time though.
 769          *
 770          * If the extent we found is inside our range, we set the
 771          * desired bit on it.
 772          */
 773         if (state->start < start) {
 774                 set = state->state & bits;
 775                 if (exclusive && set) {
 776                         *failed_start = start;
 777                         err = -EEXIST;
 778                         goto out;
 779                 }
 780                 err = split_state(tree, state, prealloc, start);
 781                 BUG_ON(err == -EEXIST);
 782                 prealloc = NULL;
 783                 if (err)
 784                         goto out;
 785                 if (state->end <= end) {
 786                         set_state_bits(tree, state, bits);
 787                         start = state->end + 1;
 788                         merge_state(tree, state);
 789                 } else {
 790                         start = state->start;
 791                 }
 792                 goto search_again;
 793         }
 794         /*
 795          * | ---- desired range ---- |
 796          *     | state | or               | state |
 797          *
 798          * There's a hole, we need to insert something in it and
 799          * ignore the extent we found.
 800          */
 801         if (state->start > start) {
 802                 u64 this_end;
 803                 if (end < last_start)
 804                         this_end = end;
 805                 else
 806                         this_end = last_start -1;
 807                 err = insert_state(tree, prealloc, start, this_end,
 808                                    bits);
 809                 prealloc = NULL;
 810                 BUG_ON(err == -EEXIST);
 811                 if (err)
 812                         goto out;
 813                 start = this_end + 1;
 814                 goto search_again;
 815         }
 816         /*
 817          * | ---- desired range ---- |
 818          *                        | state |
 819          * We need to split the extent, and set the bit
 820          * on the first half
 821          */
 822         if (state->start <= end && state->end > end) {
 823                 set = state->state & bits;
 824                 if (exclusive && set) {
 825                         *failed_start = start;
 826                         err = -EEXIST;
 827                         goto out;
 828                 }
 829                 err = split_state(tree, state, prealloc, end + 1);
 830                 BUG_ON(err == -EEXIST);
 831
 832                 set_state_bits(tree, prealloc, bits);
 833                 merge_state(tree, prealloc);
 834                 prealloc = NULL;
 835                 goto out;
 836         }
 837
 838         goto search_again;
 839
 840 out:
 841         write_unlock_irqrestore(&tree->lock, flags);
 842         if (prealloc)
 843                 free_extent_state(prealloc);
 844
 845         return err;
 846
 847 search_again:
 848         if (start > end)
 849                 goto out;
 850         write_unlock_irqrestore(&tree->lock, flags);
 851         if (mask & __GFP_WAIT)
 852                 cond_resched();
 853         goto again;
 854 }
 855 EXPORT_SYMBOL(set_extent_bit);
 856
 857 /* wrappers around set/clear extent bit */
 858 int set_extent_dirty(struct extent_map_tree *tree, u64 start, u64 end,
 859                      gfp_t mask)
 860 {
 861         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 862                               mask);
 863 }
 864 EXPORT_SYMBOL(set_extent_dirty);
 865
 866 int set_extent_bits(struct extent_map_tree *tree, u64 start, u64 end,
 867                     int bits, gfp_t mask)
 868 {
 869         return set_extent_bit(tree, start, end, bits, 0, NULL,
 870                               mask);
 871 }
 872 EXPORT_SYMBOL(set_extent_bits);
 873
 874 int clear_extent_bits(struct extent_map_tree *tree, u64 start, u64 end,
 875                       int bits, gfp_t mask)
 876 {
 877         return clear_extent_bit(tree, start, end, bits, 0, 0, mask);
 878 }
 879 EXPORT_SYMBOL(clear_extent_bits);
 880
 881 int set_extent_delalloc(struct extent_map_tree *tree, u64 start, u64 end,
 882                      gfp_t mask)
 883 {
 884         return set_extent_bit(tree, start, end,
 885                               EXTENT_DELALLOC | EXTENT_DIRTY, 0, NULL,
 886                               mask);
 887 }
 888 EXPORT_SYMBOL(set_extent_delalloc);
 889
 890 int clear_extent_dirty(struct extent_map_tree *tree, u64 start, u64 end,
 891                        gfp_t mask)
 892 {
 893         return clear_extent_bit(tree, start, end,
 894                                 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, mask);
 895 }
 896 EXPORT_SYMBOL(clear_extent_dirty);
 897
 898 int set_extent_new(struct extent_map_tree *tree, u64 start, u64 end,
 899                      gfp_t mask)
 900 {
 901         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 902                               mask);
 903 }
 904 EXPORT_SYMBOL(set_extent_new);
 905
 906 int clear_extent_new(struct extent_map_tree *tree, u64 start, u64 end,
 907                        gfp_t mask)
 908 {
 909         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0, mask);
 910 }
 911 EXPORT_SYMBOL(clear_extent_new);
 912
 913 int set_extent_uptodate(struct extent_map_tree *tree, u64 start, u64 end,
 914                         gfp_t mask)
 915 {
 916         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, NULL,
 917                               mask);
 918 }
 919 EXPORT_SYMBOL(set_extent_uptodate);
 920
 921 int clear_extent_uptodate(struct extent_map_tree *tree, u64 start, u64 end,
 922                           gfp_t mask)
 923 {
 924         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0, mask);
 925 }
 926 EXPORT_SYMBOL(clear_extent_uptodate);
 927
 928 int set_extent_writeback(struct extent_map_tree *tree, u64 start, u64 end,
 929                          gfp_t mask)
 930 {
 931         return set_extent_bit(tree, start, end, EXTENT_WRITEBACK,
 932                               0, NULL, mask);
 933 }
 934 EXPORT_SYMBOL(set_extent_writeback);
 935
 936 int clear_extent_writeback(struct extent_map_tree *tree, u64 start, u64 end,
 937                            gfp_t mask)
 938 {
 939         return clear_extent_bit(tree, start, end, EXTENT_WRITEBACK, 1, 0, mask);
 940 }
 941 EXPORT_SYMBOL(clear_extent_writeback);
 942
 943 int wait_on_extent_writeback(struct extent_map_tree *tree, u64 start, u64 end)
 944 {
 945         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 946 }
 947 EXPORT_SYMBOL(wait_on_extent_writeback);
 948
 949 /*
 950  * locks a range in ascending order, waiting for any locked regions
 951  * it hits on the way.  [start,end] are inclusive, and this will sleep.
 952  */
 953 int lock_extent(struct extent_map_tree *tree, u64 start, u64 end, gfp_t mask)
 954 {
 955         int err;
 956         u64 failed_start;
 957         while (1) {
 958                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED, 1,
 959                                      &failed_start, mask);
 960                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
 961                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
 962                         start = failed_start;
 963                 } else {
 964                         break;
 965                 }
 966                 WARN_ON(start > end);
 967         }
 968         return err;
 969 }
 970 EXPORT_SYMBOL(lock_extent);
 971
 972 int unlock_extent(struct extent_map_tree *tree, u64 start, u64 end,
 973                   gfp_t mask)
 974 {
 975         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, mask);
 976 }
 977 EXPORT_SYMBOL(unlock_extent);
 978
 979 /*
 980  * helper function to set pages and extents in the tree dirty
 981  */
 982 int set_range_dirty(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_dirty_nobuffers(page);
 992                 page_cache_release(page);
 993                 index++;
 994         }
 995         set_extent_dirty(tree, start, end, GFP_NOFS);
 996         return 0;
 997 }
 998 EXPORT_SYMBOL(set_range_dirty);
 999
1000 /*
1001  * helper function to set both pages and extents in the tree writeback
1002  */
1003 int set_range_writeback(struct extent_map_tree *tree, u64 start, u64 end)
1004 {
1005         unsigned long index = start >> PAGE_CACHE_SHIFT;
1006         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1007         struct page *page;
1008
1009         while (index <= end_index) {
1010                 page = find_get_page(tree->mapping, index);
1011                 BUG_ON(!page);
1012                 set_page_writeback(page);
1013                 page_cache_release(page);
1014                 index++;
1015         }
1016         set_extent_writeback(tree, start, end, GFP_NOFS);
1017         return 0;
1018 }
1019 EXPORT_SYMBOL(set_range_writeback);
1020
1021 int find_first_extent_bit(struct extent_map_tree *tree, u64 start,
1022                           u64 *start_ret, u64 *end_ret, int bits)
1023 {
1024         struct rb_node *node;
1025         struct extent_state *state;
1026         int ret = 1;
1027
1028         read_lock_irq(&tree->lock);
1029         /*
1030          * this search will find all the extents that end after
1031          * our range starts.
1032          */
1033         node = tree_search(&tree->state, start);
1034         if (!node || IS_ERR(node)) {
1035                 goto out;
1036         }
1037
1038         while(1) {
1039                 state = rb_entry(node, struct extent_state, rb_node);
1040                 if (state->end >= start && (state->state & bits)) {
1041                         *start_ret = state->start;
1042                         *end_ret = state->end;
1043                         ret = 0;
1044                         break;
1045                 }
1046                 node = rb_next(node);
1047                 if (!node)
1048                         break;
1049         }
1050 out:
1051         read_unlock_irq(&tree->lock);
1052         return ret;
1053 }
1054 EXPORT_SYMBOL(find_first_extent_bit);
1055
1056 u64 find_lock_delalloc_range(struct extent_map_tree *tree,
1057                              u64 *start, u64 *end, u64 max_bytes)
1058 {
1059         struct rb_node *node;
1060         struct extent_state *state;
1061         u64 cur_start = *start;
1062         u64 found = 0;
1063         u64 total_bytes = 0;
1064
1065         write_lock_irq(&tree->lock);
1066         /*
1067          * this search will find all the extents that end after
1068          * our range starts.
1069          */
1070 search_again:
1071         node = tree_search(&tree->state, cur_start);
1072         if (!node || IS_ERR(node)) {
1073                 goto out;
1074         }
1075
1076         while(1) {
1077                 state = rb_entry(node, struct extent_state, rb_node);
1078                 if (found && state->start != cur_start) {
1079                         goto out;
1080                 }
1081                 if (!(state->state & EXTENT_DELALLOC)) {
1082                         goto out;
1083                 }
1084                 if (!found) {
1085                         struct extent_state *prev_state;
1086                         struct rb_node *prev_node = node;
1087                         while(1) {
1088                                 prev_node = rb_prev(prev_node);
1089                                 if (!prev_node)
1090                                         break;
1091                                 prev_state = rb_entry(prev_node,
1092                                                       struct extent_state,
1093                                                       rb_node);
1094                                 if (!(prev_state->state & EXTENT_DELALLOC))
1095                                         break;
1096                                 state = prev_state;
1097                                 node = prev_node;
1098                         }
1099                 }
1100                 if (state->state & EXTENT_LOCKED) {
1101                         DEFINE_WAIT(wait);
1102                         atomic_inc(&state->refs);
1103                         prepare_to_wait(&state->wq, &wait,
1104                                         TASK_UNINTERRUPTIBLE);
1105                         write_unlock_irq(&tree->lock);
1106                         schedule();
1107                         write_lock_irq(&tree->lock);
1108                         finish_wait(&state->wq, &wait);
1109                         free_extent_state(state);
1110                         goto search_again;
1111                 }
1112                 state->state |= EXTENT_LOCKED;
1113                 if (!found)
1114                         *start = state->start;
1115                 found++;
1116                 *end = state->end;
1117                 cur_start = state->end + 1;
1118                 node = rb_next(node);
1119                 if (!node)
1120                         break;
1121                 total_bytes += state->end - state->start + 1;
1122                 if (total_bytes >= max_bytes)
1123                         break;
1124         }
1125 out:
1126         write_unlock_irq(&tree->lock);
1127         return found;
1128 }
1129
1130 u64 count_range_bits(struct extent_map_tree *tree,
1131                      u64 *start, u64 max_bytes, unsigned long bits)
1132 {
1133         struct rb_node *node;
1134         struct extent_state *state;
1135         u64 cur_start = *start;
1136         u64 total_bytes = 0;
1137         int found = 0;
1138
1139         write_lock_irq(&tree->lock);
1140         if (bits == EXTENT_DIRTY) {
1141                 *start = 0;
1142                 total_bytes = tree->dirty_bytes;
1143                 goto out;
1144         }
1145         /*
1146          * this search will find all the extents that end after
1147          * our range starts.
1148          */
1149         node = tree_search(&tree->state, cur_start);
1150         if (!node || IS_ERR(node)) {
1151                 goto out;
1152         }
1153
1154         while(1) {
1155                 state = rb_entry(node, struct extent_state, rb_node);
1156                 if ((state->state & bits)) {
1157                         total_bytes += state->end - state->start + 1;
1158                         if (total_bytes >= max_bytes)
1159                                 break;
1160                         if (!found) {
1161                                 *start = state->start;
1162                                 found = 1;
1163                         }
1164                 }
1165                 node = rb_next(node);
1166                 if (!node)
1167                         break;
1168         }
1169 out:
1170         write_unlock_irq(&tree->lock);
1171         return total_bytes;
1172 }
1173
1174 /*
1175  * helper function to lock both pages and extents in the tree.
1176  * pages must be locked first.
1177  */
1178 int lock_range(struct extent_map_tree *tree, u64 start, u64 end)
1179 {
1180         unsigned long index = start >> PAGE_CACHE_SHIFT;
1181         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1182         struct page *page;
1183         int err;
1184
1185         while (index <= end_index) {
1186                 page = grab_cache_page(tree->mapping, index);
1187                 if (!page) {
1188                         err = -ENOMEM;
1189                         goto failed;
1190                 }
1191                 if (IS_ERR(page)) {
1192                         err = PTR_ERR(page);
1193                         goto failed;
1194                 }
1195                 index++;
1196         }
1197         lock_extent(tree, start, end, GFP_NOFS);
1198         return 0;
1199
1200 failed:
1201         /*
1202          * we failed above in getting the page at 'index', so we undo here
1203          * up to but not including the page at 'index'
1204          */
1205         end_index = index;
1206         index = start >> PAGE_CACHE_SHIFT;
1207         while (index < end_index) {
1208                 page = find_get_page(tree->mapping, index);
1209                 unlock_page(page);
1210                 page_cache_release(page);
1211                 index++;
1212         }
1213         return err;
1214 }
1215 EXPORT_SYMBOL(lock_range);
1216
1217 /*
1218  * helper function to unlock both pages and extents in the tree.
1219  */
1220 int unlock_range(struct extent_map_tree *tree, u64 start, u64 end)
1221 {
1222         unsigned long index = start >> PAGE_CACHE_SHIFT;
1223         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1224         struct page *page;
1225
1226         while (index <= end_index) {
1227                 page = find_get_page(tree->mapping, index);
1228                 unlock_page(page);
1229                 page_cache_release(page);
1230                 index++;
1231         }
1232         unlock_extent(tree, start, end, GFP_NOFS);
1233         return 0;
1234 }
1235 EXPORT_SYMBOL(unlock_range);
1236
1237 int set_state_private(struct extent_map_tree *tree, u64 start, u64 private)
1238 {
1239         struct rb_node *node;
1240         struct extent_state *state;
1241         int ret = 0;
1242
1243         write_lock_irq(&tree->lock);
1244         /*
1245          * this search will find all the extents that end after
1246          * our range starts.
1247          */
1248         node = tree_search(&tree->state, start);
1249         if (!node || IS_ERR(node)) {
1250                 ret = -ENOENT;
1251                 goto out;
1252         }
1253         state = rb_entry(node, struct extent_state, rb_node);
1254         if (state->start != start) {
1255                 ret = -ENOENT;
1256                 goto out;
1257         }
1258         state->private = private;
1259 out:
1260         write_unlock_irq(&tree->lock);
1261         return ret;
1262 }
1263
1264 int get_state_private(struct extent_map_tree *tree, u64 start, u64 *private)
1265 {
1266         struct rb_node *node;
1267         struct extent_state *state;
1268         int ret = 0;
1269
1270         read_lock_irq(&tree->lock);
1271         /*
1272          * this search will find all the extents that end after
1273          * our range starts.
1274          */
1275         node = tree_search(&tree->state, start);
1276         if (!node || IS_ERR(node)) {
1277                 ret = -ENOENT;
1278                 goto out;
1279         }
1280         state = rb_entry(node, struct extent_state, rb_node);
1281         if (state->start != start) {
1282                 ret = -ENOENT;
1283                 goto out;
1284         }
1285         *private = state->private;
1286 out:
1287         read_unlock_irq(&tree->lock);
1288         return ret;
1289 }
1290
1291 /*
1292  * searches a range in the state tree for a given mask.
1293  * If 'filled' == 1, this returns 1 only if ever extent in the tree
1294  * has the bits set.  Otherwise, 1 is returned if any bit in the
1295  * range is found set.
1296  */
1297 int test_range_bit(struct extent_map_tree *tree, u64 start, u64 end,
1298                    int bits, int filled)
1299 {
1300         struct extent_state *state = NULL;
1301         struct rb_node *node;
1302         int bitset = 0;
1303
1304         read_lock_irq(&tree->lock);
1305         node = tree_search(&tree->state, start);
1306         while (node && start <= end) {
1307                 state = rb_entry(node, struct extent_state, rb_node);
1308
1309                 if (filled && state->start > start) {
1310                         bitset = 0;
1311                         break;
1312                 }
1313
1314                 if (state->start > end)
1315                         break;
1316
1317                 if (state->state & bits) {
1318                         bitset = 1;
1319                         if (!filled)
1320                                 break;
1321                 } else if (filled) {
1322                         bitset = 0;
1323                         break;
1324                 }
1325                 start = state->end + 1;
1326                 if (start > end)
1327                         break;
1328                 node = rb_next(node);
1329         }
1330         read_unlock_irq(&tree->lock);
1331         return bitset;
1332 }
1333 EXPORT_SYMBOL(test_range_bit);
1334
1335 /*
1336  * helper function to set a given page up to date if all the
1337  * extents in the tree for that page are up to date
1338  */
1339 static int check_page_uptodate(struct extent_map_tree *tree,
1340                                struct page *page)
1341 {
1342         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1343         u64 end = start + PAGE_CACHE_SIZE - 1;
1344         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1))
1345                 SetPageUptodate(page);
1346         return 0;
1347 }
1348
1349 /*
1350  * helper function to unlock a page if all the extents in the tree
1351  * for that page are unlocked
1352  */
1353 static int check_page_locked(struct extent_map_tree *tree,
1354                              struct page *page)
1355 {
1356         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1357         u64 end = start + PAGE_CACHE_SIZE - 1;
1358         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0))
1359                 unlock_page(page);
1360         return 0;
1361 }
1362
1363 /*
1364  * helper function to end page writeback if all the extents
1365  * in the tree for that page are done with writeback
1366  */
1367 static int check_page_writeback(struct extent_map_tree *tree,
1368                              struct page *page)
1369 {
1370         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1371         u64 end = start + PAGE_CACHE_SIZE - 1;
1372         if (!test_range_bit(tree, start, end, EXTENT_WRITEBACK, 0))
1373                 end_page_writeback(page);
1374         return 0;
1375 }
1376
1377 /* lots and lots of room for performance fixes in the end_bio funcs */
1378
1379 /*
1380  * after a writepage IO is done, we need to:
1381  * clear the uptodate bits on error
1382  * clear the writeback bits in the extent tree for this IO
1383  * end_page_writeback if the page has no more pending IO
1384  *
1385  * Scheduling is not allowed, so the extent state tree is expected
1386  * to have one and only one object corresponding to this IO.
1387  */
1388 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1389 static void end_bio_extent_writepage(struct bio *bio, int err)
1390 #else
1391 static int end_bio_extent_writepage(struct bio *bio,
1392                                    unsigned int bytes_done, int err)
1393 #endif
1394 {
1395         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1396         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1397         struct extent_map_tree *tree = bio->bi_private;
1398         u64 start;
1399         u64 end;
1400         int whole_page;
1401
1402 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1403         if (bio->bi_size)
1404                 return 1;
1405 #endif
1406
1407         do {
1408                 struct page *page = bvec->bv_page;
1409                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1410                          bvec->bv_offset;
1411                 end = start + bvec->bv_len - 1;
1412
1413                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1414                         whole_page = 1;
1415                 else
1416                         whole_page = 0;
1417
1418                 if (--bvec >= bio->bi_io_vec)
1419                         prefetchw(&bvec->bv_page->flags);
1420
1421                 if (!uptodate) {
1422                         clear_extent_uptodate(tree, start, end, GFP_ATOMIC);
1423                         ClearPageUptodate(page);
1424                         SetPageError(page);
1425                 }
1426                 clear_extent_writeback(tree, start, end, GFP_ATOMIC);
1427
1428                 if (whole_page)
1429                         end_page_writeback(page);
1430                 else
1431                         check_page_writeback(tree, page);
1432                 if (tree->ops && tree->ops->writepage_end_io_hook)
1433                         tree->ops->writepage_end_io_hook(page, start, end);
1434         } while (bvec >= bio->bi_io_vec);
1435
1436         bio_put(bio);
1437 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1438         return 0;
1439 #endif
1440 }
1441
1442 /*
1443  * after a readpage IO is done, we need to:
1444  * clear the uptodate bits on error
1445  * set the uptodate bits if things worked
1446  * set the page up to date if all extents in the tree are uptodate
1447  * clear the lock bit in the extent tree
1448  * unlock the page if there are no other extents locked for it
1449  *
1450  * Scheduling is not allowed, so the extent state tree is expected
1451  * to have one and only one object corresponding to this IO.
1452  */
1453 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1454 static void end_bio_extent_readpage(struct bio *bio, int err)
1455 #else
1456 static int end_bio_extent_readpage(struct bio *bio,
1457                                    unsigned int bytes_done, int err)
1458 #endif
1459 {
1460         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1461         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1462         struct extent_map_tree *tree = bio->bi_private;
1463         u64 start;
1464         u64 end;
1465         int whole_page;
1466         int ret;
1467
1468 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1469         if (bio->bi_size)
1470                 return 1;
1471 #endif
1472
1473         do {
1474                 struct page *page = bvec->bv_page;
1475                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1476                         bvec->bv_offset;
1477                 end = start + bvec->bv_len - 1;
1478
1479                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1480                         whole_page = 1;
1481                 else
1482                         whole_page = 0;
1483
1484                 if (--bvec >= bio->bi_io_vec)
1485                         prefetchw(&bvec->bv_page->flags);
1486
1487                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1488                         ret = tree->ops->readpage_end_io_hook(page, start, end);
1489                         if (ret)
1490                                 uptodate = 0;
1491                 }
1492                 if (uptodate) {
1493                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1494                         if (whole_page)
1495                                 SetPageUptodate(page);
1496                         else
1497                                 check_page_uptodate(tree, page);
1498                 } else {
1499                         ClearPageUptodate(page);
1500                         SetPageError(page);
1501                 }
1502
1503                 unlock_extent(tree, start, end, GFP_ATOMIC);
1504
1505                 if (whole_page)
1506                         unlock_page(page);
1507                 else
1508                         check_page_locked(tree, page);
1509         } while (bvec >= bio->bi_io_vec);
1510
1511         bio_put(bio);
1512 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1513         return 0;
1514 #endif
1515 }
1516
1517 /*
1518  * IO done from prepare_write is pretty simple, we just unlock
1519  * the structs in the extent tree when done, and set the uptodate bits
1520  * as appropriate.
1521  */
1522 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1523 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1524 #else
1525 static int end_bio_extent_preparewrite(struct bio *bio,
1526                                        unsigned int bytes_done, int err)
1527 #endif
1528 {
1529         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1530         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1531         struct extent_map_tree *tree = bio->bi_private;
1532         u64 start;
1533         u64 end;
1534
1535 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1536         if (bio->bi_size)
1537                 return 1;
1538 #endif
1539
1540         do {
1541                 struct page *page = bvec->bv_page;
1542                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1543                         bvec->bv_offset;
1544                 end = start + bvec->bv_len - 1;
1545
1546                 if (--bvec >= bio->bi_io_vec)
1547                         prefetchw(&bvec->bv_page->flags);
1548
1549                 if (uptodate) {
1550                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1551                 } else {
1552                         ClearPageUptodate(page);
1553                         SetPageError(page);
1554                 }
1555
1556                 unlock_extent(tree, start, end, GFP_ATOMIC);
1557
1558         } while (bvec >= bio->bi_io_vec);
1559
1560         bio_put(bio);
1561 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1562         return 0;
1563 #endif
1564 }
1565
1566 static struct bio *
1567 extent_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1568                  gfp_t gfp_flags)
1569 {
1570         struct bio *bio;
1571
1572         bio = bio_alloc(gfp_flags, nr_vecs);
1573
1574         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1575                 while (!bio && (nr_vecs /= 2))
1576                         bio = bio_alloc(gfp_flags, nr_vecs);
1577         }
1578
1579         if (bio) {
1580                 bio->bi_bdev = bdev;
1581                 bio->bi_sector = first_sector;
1582         }
1583         return bio;
1584 }
1585
1586 static int submit_one_bio(int rw, struct bio *bio)
1587 {
1588         int ret = 0;
1589         bio_get(bio);
1590         submit_bio(rw, bio);
1591         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1592                 ret = -EOPNOTSUPP;
1593         bio_put(bio);
1594         return ret;
1595 }
1596
1597 static int submit_extent_page(int rw, struct extent_map_tree *tree,
1598                               struct page *page, sector_t sector,
1599                               size_t size, unsigned long offset,
1600                               struct block_device *bdev,
1601                               struct bio **bio_ret,
1602                               unsigned long max_pages,
1603                               bio_end_io_t end_io_func)
1604 {
1605         int ret = 0;
1606         struct bio *bio;
1607         int nr;
1608
1609         if (bio_ret && *bio_ret) {
1610                 bio = *bio_ret;
1611                 if (bio->bi_sector + (bio->bi_size >> 9) != sector ||
1612                     bio_add_page(bio, page, size, offset) < size) {
1613                         ret = submit_one_bio(rw, bio);
1614                         bio = NULL;
1615                 } else {
1616                         return 0;
1617                 }
1618         }
1619         nr = min_t(int, max_pages, bio_get_nr_vecs(bdev));
1620         bio = extent_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1621         if (!bio) {
1622                 printk("failed to allocate bio nr %d\n", nr);
1623         }
1624         bio_add_page(bio, page, size, offset);
1625         bio->bi_end_io = end_io_func;
1626         bio->bi_private = tree;
1627         if (bio_ret) {
1628                 *bio_ret = bio;
1629         } else {
1630                 ret = submit_one_bio(rw, bio);
1631         }
1632
1633         return ret;
1634 }
1635
1636 void set_page_extent_mapped(struct page *page)
1637 {
1638         if (!PagePrivate(page)) {
1639                 SetPagePrivate(page);
1640                 WARN_ON(!page->mapping->a_ops->invalidatepage);
1641                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1642                 page_cache_get(page);
1643         }
1644 }
1645
1646 /*
1647  * basic readpage implementation.  Locked extent state structs are inserted
1648  * into the tree that are removed when the IO is done (by the end_io
1649  * handlers)
1650  */
1651 static int __extent_read_full_page(struct extent_map_tree *tree,
1652                                    struct page *page,
1653                                    get_extent_t *get_extent,
1654                                    struct bio **bio)
1655 {
1656         struct inode *inode = page->mapping->host;
1657         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1658         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1659         u64 end;
1660         u64 cur = start;
1661         u64 extent_offset;
1662         u64 last_byte = i_size_read(inode);
1663         u64 block_start;
1664         u64 cur_end;
1665         sector_t sector;
1666         struct extent_map *em;
1667         struct block_device *bdev;
1668         int ret;
1669         int nr = 0;
1670         size_t page_offset = 0;
1671         size_t iosize;
1672         size_t blocksize = inode->i_sb->s_blocksize;
1673
1674         set_page_extent_mapped(page);
1675
1676         end = page_end;
1677         lock_extent(tree, start, end, GFP_NOFS);
1678
1679         while (cur <= end) {
1680                 if (cur >= last_byte) {
1681                         iosize = PAGE_CACHE_SIZE - page_offset;
1682                         zero_user_page(page, page_offset, iosize, KM_USER0);
1683                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1684                                             GFP_NOFS);
1685                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1686                         break;
1687                 }
1688                 em = get_extent(inode, page, page_offset, cur, end, 0);
1689                 if (IS_ERR(em) || !em) {
1690                         SetPageError(page);
1691                         unlock_extent(tree, cur, end, GFP_NOFS);
1692                         break;
1693                 }
1694
1695                 extent_offset = cur - em->start;
1696                 BUG_ON(em->end < cur);
1697                 BUG_ON(end < cur);
1698
1699                 iosize = min(em->end - cur, end - cur) + 1;
1700                 cur_end = min(em->end, end);
1701                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1702                 sector = (em->block_start + extent_offset) >> 9;
1703                 bdev = em->bdev;
1704                 block_start = em->block_start;
1705                 free_extent_map(em);
1706                 em = NULL;
1707
1708                 /* we've found a hole, just zero and go on */
1709                 if (block_start == EXTENT_MAP_HOLE) {
1710                         zero_user_page(page, page_offset, iosize, KM_USER0);
1711                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1712                                             GFP_NOFS);
1713                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1714                         cur = cur + iosize;
1715                         page_offset += iosize;
1716                         continue;
1717                 }
1718                 /* the get_extent function already copied into the page */
1719                 if (test_range_bit(tree, cur, cur_end, EXTENT_UPTODATE, 1)) {
1720                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1721                         cur = cur + iosize;
1722                         page_offset += iosize;
1723                         continue;
1724                 }
1725
1726                 ret = 0;
1727                 if (tree->ops && tree->ops->readpage_io_hook) {
1728                         ret = tree->ops->readpage_io_hook(page, cur,
1729                                                           cur + iosize - 1);
1730                 }
1731                 if (!ret) {
1732                         unsigned long nr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
1733                         nr -= page->index;
1734                         ret = submit_extent_page(READ, tree, page,
1735                                          sector, iosize, page_offset,
1736                                          bdev, bio, nr,
1737                                          end_bio_extent_readpage);
1738                 }
1739                 if (ret)
1740                         SetPageError(page);
1741                 cur = cur + iosize;
1742                 page_offset += iosize;
1743                 nr++;
1744         }
1745         if (!nr) {
1746                 if (!PageError(page))
1747                         SetPageUptodate(page);
1748                 unlock_page(page);
1749         }
1750         return 0;
1751 }
1752
1753 int extent_read_full_page(struct extent_map_tree *tree, struct page *page,
1754                             get_extent_t *get_extent)
1755 {
1756         struct bio *bio = NULL;
1757         int ret;
1758
1759         ret = __extent_read_full_page(tree, page, get_extent, &bio);
1760         if (bio)
1761                 submit_one_bio(READ, bio);
1762         return ret;
1763 }
1764 EXPORT_SYMBOL(extent_read_full_page);
1765
1766 /*
1767  * the writepage semantics are similar to regular writepage.  extent
1768  * records are inserted to lock ranges in the tree, and as dirty areas
1769  * are found, they are marked writeback.  Then the lock bits are removed
1770  * and the end_io handler clears the writeback ranges
1771  */
1772 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
1773                               void *data)
1774 {
1775         struct inode *inode = page->mapping->host;
1776         struct extent_page_data *epd = data;
1777         struct extent_map_tree *tree = epd->tree;
1778         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1779         u64 delalloc_start;
1780         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1781         u64 end;
1782         u64 cur = start;
1783         u64 extent_offset;
1784         u64 last_byte = i_size_read(inode);
1785         u64 block_start;
1786         u64 iosize;
1787         sector_t sector;
1788         struct extent_map *em;
1789         struct block_device *bdev;
1790         int ret;
1791         int nr = 0;
1792         size_t page_offset = 0;
1793         size_t blocksize;
1794         loff_t i_size = i_size_read(inode);
1795         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
1796         u64 nr_delalloc;
1797         u64 delalloc_end;
1798
1799         WARN_ON(!PageLocked(page));
1800         if (page->index > end_index) {
1801                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
1802                 unlock_page(page);
1803                 return 0;
1804         }
1805
1806         if (page->index == end_index) {
1807                 size_t offset = i_size & (PAGE_CACHE_SIZE - 1);
1808                 zero_user_page(page, offset,
1809                                PAGE_CACHE_SIZE - offset, KM_USER0);
1810         }
1811
1812         set_page_extent_mapped(page);
1813
1814         delalloc_start = start;
1815         delalloc_end = 0;
1816         while(delalloc_end < page_end) {
1817                 nr_delalloc = find_lock_delalloc_range(tree, &delalloc_start,
1818                                                        &delalloc_end,
1819                                                        128 * 1024 * 1024);
1820                 if (nr_delalloc <= 0)
1821                         break;
1822                 tree->ops->fill_delalloc(inode, delalloc_start,
1823                                          delalloc_end);
1824                 clear_extent_bit(tree, delalloc_start,
1825                                  delalloc_end,
1826                                  EXTENT_LOCKED | EXTENT_DELALLOC,
1827                                  1, 0, GFP_NOFS);
1828                 delalloc_start = delalloc_end + 1;
1829         }
1830         lock_extent(tree, start, page_end, GFP_NOFS);
1831
1832         end = page_end;
1833         if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0)) {
1834                 printk("found delalloc bits after lock_extent\n");
1835         }
1836
1837         if (last_byte <= start) {
1838                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
1839                 goto done;
1840         }
1841
1842         set_extent_uptodate(tree, start, page_end, GFP_NOFS);
1843         blocksize = inode->i_sb->s_blocksize;
1844
1845         while (cur <= end) {
1846                 if (cur >= last_byte) {
1847                         clear_extent_dirty(tree, cur, page_end, GFP_NOFS);
1848                         break;
1849                 }
1850                 em = epd->get_extent(inode, page, page_offset, cur, end, 1);
1851                 if (IS_ERR(em) || !em) {
1852                         SetPageError(page);
1853                         break;
1854                 }
1855
1856                 extent_offset = cur - em->start;
1857                 BUG_ON(em->end < cur);
1858                 BUG_ON(end < cur);
1859                 iosize = min(em->end - cur, end - cur) + 1;
1860                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1861                 sector = (em->block_start + extent_offset) >> 9;
1862                 bdev = em->bdev;
1863                 block_start = em->block_start;
1864                 free_extent_map(em);
1865                 em = NULL;
1866
1867                 if (block_start == EXTENT_MAP_HOLE ||
1868                     block_start == EXTENT_MAP_INLINE) {
1869                         clear_extent_dirty(tree, cur,
1870                                            cur + iosize - 1, GFP_NOFS);
1871                         cur = cur + iosize;
1872                         page_offset += iosize;
1873                         continue;
1874                 }
1875
1876                 /* leave this out until we have a page_mkwrite call */
1877                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
1878                                    EXTENT_DIRTY, 0)) {
1879                         cur = cur + iosize;
1880                         page_offset += iosize;
1881                         continue;
1882                 }
1883                 clear_extent_dirty(tree, cur, cur + iosize - 1, GFP_NOFS);
1884                 if (tree->ops && tree->ops->writepage_io_hook) {
1885                         ret = tree->ops->writepage_io_hook(page, cur,
1886                                                 cur + iosize - 1);
1887                 } else {
1888                         ret = 0;
1889                 }
1890                 if (ret)
1891                         SetPageError(page);
1892                 else {
1893                         unsigned long max_nr = end_index + 1;
1894                         set_range_writeback(tree, cur, cur + iosize - 1);
1895                         if (!PageWriteback(page)) {
1896                                 printk("warning page %lu not writeback, "
1897                                        "cur %llu end %llu\n", page->index,
1898                                        (unsigned long long)cur,
1899                                        (unsigned long long)end);
1900                         }
1901
1902                         ret = submit_extent_page(WRITE, tree, page, sector,
1903                                                  iosize, page_offset, bdev,
1904                                                  &epd->bio, max_nr,
1905                                                  end_bio_extent_writepage);
1906                         if (ret)
1907                                 SetPageError(page);
1908                 }
1909                 cur = cur + iosize;
1910                 page_offset += iosize;
1911                 nr++;
1912         }
1913 done:
1914         if (nr == 0) {
1915                 /* make sure the mapping tag for page dirty gets cleared */
1916                 set_page_writeback(page);
1917                 end_page_writeback(page);
1918         }
1919         unlock_extent(tree, start, page_end, GFP_NOFS);
1920         unlock_page(page);
1921         return 0;
1922 }
1923
1924 int extent_write_full_page(struct extent_map_tree *tree, struct page *page,
1925                           get_extent_t *get_extent,
1926                           struct writeback_control *wbc)
1927 {
1928         int ret;
1929         struct address_space *mapping = page->mapping;
1930         struct extent_page_data epd = {
1931                 .bio = NULL,
1932                 .tree = tree,
1933                 .get_extent = get_extent,
1934         };
1935         struct writeback_control wbc_writepages = {
1936                 .bdi            = wbc->bdi,
1937                 .sync_mode      = WB_SYNC_NONE,
1938                 .older_than_this = NULL,
1939                 .nr_to_write    = 64,
1940                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
1941                 .range_end      = (loff_t)-1,
1942         };
1943
1944
1945         ret = __extent_writepage(page, wbc, &epd);
1946
1947         write_cache_pages(mapping, &wbc_writepages, __extent_writepage, &epd);
1948         if (epd.bio)
1949                 submit_one_bio(WRITE, epd.bio);
1950         return ret;
1951 }
1952 EXPORT_SYMBOL(extent_write_full_page);
1953
1954 int extent_writepages(struct extent_map_tree *tree,
1955                       struct address_space *mapping,
1956                       get_extent_t *get_extent,
1957                       struct writeback_control *wbc)
1958 {
1959         int ret;
1960         struct extent_page_data epd = {
1961                 .bio = NULL,
1962                 .tree = tree,
1963                 .get_extent = get_extent,
1964         };
1965
1966         ret = write_cache_pages(mapping, wbc, __extent_writepage, &epd);
1967         if (epd.bio)
1968                 submit_one_bio(WRITE, epd.bio);
1969         return ret;
1970 }
1971 EXPORT_SYMBOL(extent_writepages);
1972
1973 int extent_readpages(struct extent_map_tree *tree,
1974                      struct address_space *mapping,
1975                      struct list_head *pages, unsigned nr_pages,
1976                      get_extent_t get_extent)
1977 {
1978         struct bio *bio = NULL;
1979         unsigned page_idx;
1980         struct pagevec pvec;
1981
1982         pagevec_init(&pvec, 0);
1983         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
1984                 struct page *page = list_entry(pages->prev, struct page, lru);
1985
1986                 prefetchw(&page->flags);
1987                 list_del(&page->lru);
1988                 /*
1989                  * what we want to do here is call add_to_page_cache_lru,
1990                  * but that isn't exported, so we reproduce it here
1991                  */
1992                 if (!add_to_page_cache(page, mapping,
1993                                         page->index, GFP_KERNEL)) {
1994
1995                         /* open coding of lru_cache_add, also not exported */
1996                         page_cache_get(page);
1997                         if (!pagevec_add(&pvec, page))
1998                                 __pagevec_lru_add(&pvec);
1999                         __extent_read_full_page(tree, page, get_extent, &bio);
2000                 }
2001                 page_cache_release(page);
2002         }
2003         if (pagevec_count(&pvec))
2004                 __pagevec_lru_add(&pvec);
2005         BUG_ON(!list_empty(pages));
2006         if (bio)
2007                 submit_one_bio(READ, bio);
2008         return 0;
2009 }
2010 EXPORT_SYMBOL(extent_readpages);
2011
2012 /*
2013  * basic invalidatepage code, this waits on any locked or writeback
2014  * ranges corresponding to the page, and then deletes any extent state
2015  * records from the tree
2016  */
2017 int extent_invalidatepage(struct extent_map_tree *tree,
2018                           struct page *page, unsigned long offset)
2019 {
2020         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2021         u64 end = start + PAGE_CACHE_SIZE - 1;
2022         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2023
2024         start += (offset + blocksize -1) & ~(blocksize - 1);
2025         if (start > end)
2026                 return 0;
2027
2028         lock_extent(tree, start, end, GFP_NOFS);
2029         wait_on_extent_writeback(tree, start, end);
2030         clear_extent_bit(tree, start, end,
2031                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC,
2032                          1, 1, GFP_NOFS);
2033         return 0;
2034 }
2035 EXPORT_SYMBOL(extent_invalidatepage);
2036
2037 /*
2038  * simple commit_write call, set_range_dirty is used to mark both
2039  * the pages and the extent records as dirty
2040  */
2041 int extent_commit_write(struct extent_map_tree *tree,
2042                         struct inode *inode, struct page *page,
2043                         unsigned from, unsigned to)
2044 {
2045         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2046
2047         set_page_extent_mapped(page);
2048         set_page_dirty(page);
2049
2050         if (pos > inode->i_size) {
2051                 i_size_write(inode, pos);
2052                 mark_inode_dirty(inode);
2053         }
2054         return 0;
2055 }
2056 EXPORT_SYMBOL(extent_commit_write);
2057
2058 int extent_prepare_write(struct extent_map_tree *tree,
2059                          struct inode *inode, struct page *page,
2060                          unsigned from, unsigned to, get_extent_t *get_extent)
2061 {
2062         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2063         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2064         u64 block_start;
2065         u64 orig_block_start;
2066         u64 block_end;
2067         u64 cur_end;
2068         struct extent_map *em;
2069         unsigned blocksize = 1 << inode->i_blkbits;
2070         size_t page_offset = 0;
2071         size_t block_off_start;
2072         size_t block_off_end;
2073         int err = 0;
2074         int iocount = 0;
2075         int ret = 0;
2076         int isnew;
2077
2078         set_page_extent_mapped(page);
2079
2080         block_start = (page_start + from) & ~((u64)blocksize - 1);
2081         block_end = (page_start + to - 1) | (blocksize - 1);
2082         orig_block_start = block_start;
2083
2084         lock_extent(tree, page_start, page_end, GFP_NOFS);
2085         while(block_start <= block_end) {
2086                 em = get_extent(inode, page, page_offset, block_start,
2087                                 block_end, 1);
2088                 if (IS_ERR(em) || !em) {
2089                         goto err;
2090                 }
2091                 cur_end = min(block_end, em->end);
2092                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2093                 block_off_end = block_off_start + blocksize;
2094                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2095
2096                 if (!PageUptodate(page) && isnew &&
2097                     (block_off_end > to || block_off_start < from)) {
2098                         void *kaddr;
2099
2100                         kaddr = kmap_atomic(page, KM_USER0);
2101                         if (block_off_end > to)
2102                                 memset(kaddr + to, 0, block_off_end - to);
2103                         if (block_off_start < from)
2104                                 memset(kaddr + block_off_start, 0,
2105                                        from - block_off_start);
2106                         flush_dcache_page(page);
2107                         kunmap_atomic(kaddr, KM_USER0);
2108                 }
2109                 if (!isnew && !PageUptodate(page) &&
2110                     (block_off_end > to || block_off_start < from) &&
2111                     !test_range_bit(tree, block_start, cur_end,
2112                                     EXTENT_UPTODATE, 1)) {
2113                         u64 sector;
2114                         u64 extent_offset = block_start - em->start;
2115                         size_t iosize;
2116                         sector = (em->block_start + extent_offset) >> 9;
2117                         iosize = (cur_end - block_start + blocksize - 1) &
2118                                 ~((u64)blocksize - 1);
2119                         /*
2120                          * we've already got the extent locked, but we
2121                          * need to split the state such that our end_bio
2122                          * handler can clear the lock.
2123                          */
2124                         set_extent_bit(tree, block_start,
2125                                        block_start + iosize - 1,
2126                                        EXTENT_LOCKED, 0, NULL, GFP_NOFS);
2127                         ret = submit_extent_page(READ, tree, page,
2128                                          sector, iosize, page_offset, em->bdev,
2129                                          NULL, 1,
2130                                          end_bio_extent_preparewrite);
2131                         iocount++;
2132                         block_start = block_start + iosize;
2133                 } else {
2134                         set_extent_uptodate(tree, block_start, cur_end,
2135                                             GFP_NOFS);
2136                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2137                         block_start = cur_end + 1;
2138                 }
2139                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2140                 free_extent_map(em);
2141         }
2142         if (iocount) {
2143                 wait_extent_bit(tree, orig_block_start,
2144                                 block_end, EXTENT_LOCKED);
2145         }
2146         check_page_uptodate(tree, page);
2147 err:
2148         /* FIXME, zero out newly allocated blocks on error */
2149         return err;
2150 }
2151 EXPORT_SYMBOL(extent_prepare_write);
2152
2153 /*
2154  * a helper for releasepage.  As long as there are no locked extents
2155  * in the range corresponding to the page, both state records and extent
2156  * map records are removed
2157  */
2158 int try_release_extent_mapping(struct extent_map_tree *tree, struct page *page)
2159 {
2160         struct extent_map *em;
2161         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2162         u64 end = start + PAGE_CACHE_SIZE - 1;
2163         u64 orig_start = start;
2164         int ret = 1;
2165
2166         while (start <= end) {
2167                 em = lookup_extent_mapping(tree, start, end);
2168                 if (!em || IS_ERR(em))
2169                         break;
2170                 if (!test_range_bit(tree, em->start, em->end,
2171                                     EXTENT_LOCKED, 0)) {
2172                         remove_extent_mapping(tree, em);
2173                         /* once for the rb tree */
2174                         free_extent_map(em);
2175                 }
2176                 start = em->end + 1;
2177                 /* once for us */
2178                 free_extent_map(em);
2179         }
2180         if (test_range_bit(tree, orig_start, end, EXTENT_LOCKED, 0))
2181                 ret = 0;
2182         else
2183                 clear_extent_bit(tree, orig_start, end, EXTENT_UPTODATE,
2184                                  1, 1, GFP_NOFS);
2185         return ret;
2186 }
2187 EXPORT_SYMBOL(try_release_extent_mapping);
2188
2189 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2190                 get_extent_t *get_extent)
2191 {
2192         struct inode *inode = mapping->host;
2193         u64 start = iblock << inode->i_blkbits;
2194         u64 end = start + (1 << inode->i_blkbits) - 1;
2195         sector_t sector = 0;
2196         struct extent_map *em;
2197
2198         em = get_extent(inode, NULL, 0, start, end, 0);
2199         if (!em || IS_ERR(em))
2200                 return 0;
2201
2202         if (em->block_start == EXTENT_MAP_INLINE ||
2203             em->block_start == EXTENT_MAP_HOLE)
2204                 goto out;
2205
2206         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2207 out:
2208         free_extent_map(em);
2209         return sector;
2210 }
2211
2212 static int add_lru(struct extent_map_tree *tree, struct extent_buffer *eb)
2213 {
2214         if (list_empty(&eb->lru)) {
2215                 extent_buffer_get(eb);
2216                 list_add(&eb->lru, &tree->buffer_lru);
2217                 tree->lru_size++;
2218                 if (tree->lru_size >= BUFFER_LRU_MAX) {
2219                         struct extent_buffer *rm;
2220                         rm = list_entry(tree->buffer_lru.prev,
2221                                         struct extent_buffer, lru);
2222                         tree->lru_size--;
2223                         list_del_init(&rm->lru);
2224                         free_extent_buffer(rm);
2225                 }
2226         } else
2227                 list_move(&eb->lru, &tree->buffer_lru);
2228         return 0;
2229 }
2230 static struct extent_buffer *find_lru(struct extent_map_tree *tree,
2231                                       u64 start, unsigned long len)
2232 {
2233         struct list_head *lru = &tree->buffer_lru;
2234         struct list_head *cur = lru->next;
2235         struct extent_buffer *eb;
2236
2237         if (list_empty(lru))
2238                 return NULL;
2239
2240         do {
2241                 eb = list_entry(cur, struct extent_buffer, lru);
2242                 if (eb->start == start && eb->len == len) {
2243                         extent_buffer_get(eb);
2244                         return eb;
2245                 }
2246                 cur = cur->next;
2247         } while (cur != lru);
2248         return NULL;
2249 }
2250
2251 static inline unsigned long num_extent_pages(u64 start, u64 len)
2252 {
2253         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
2254                 (start >> PAGE_CACHE_SHIFT);
2255 }
2256
2257 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2258                                               unsigned long i)
2259 {
2260         struct page *p;
2261         struct address_space *mapping;
2262
2263         if (i == 0)
2264                 return eb->first_page;
2265         i += eb->start >> PAGE_CACHE_SHIFT;
2266         mapping = eb->first_page->mapping;
2267         read_lock_irq(&mapping->tree_lock);
2268         p = radix_tree_lookup(&mapping->page_tree, i);
2269         read_unlock_irq(&mapping->tree_lock);
2270         return p;
2271 }
2272
2273 static struct extent_buffer *__alloc_extent_buffer(struct extent_map_tree *tree,
2274                                                    u64 start,
2275                                                    unsigned long len,
2276                                                    gfp_t mask)
2277 {
2278         struct extent_buffer *eb = NULL;
2279
2280         spin_lock(&tree->lru_lock);
2281         eb = find_lru(tree, start, len);
2282         spin_unlock(&tree->lru_lock);
2283         if (eb) {
2284                 return eb;
2285         }
2286
2287         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
2288         INIT_LIST_HEAD(&eb->lru);
2289         eb->start = start;
2290         eb->len = len;
2291         atomic_set(&eb->refs, 1);
2292
2293         return eb;
2294 }
2295
2296 static void __free_extent_buffer(struct extent_buffer *eb)
2297 {
2298         kmem_cache_free(extent_buffer_cache, eb);
2299 }
2300
2301 struct extent_buffer *alloc_extent_buffer(struct extent_map_tree *tree,
2302                                           u64 start, unsigned long len,
2303                                           struct page *page0,
2304                                           gfp_t mask)
2305 {
2306         unsigned long num_pages = num_extent_pages(start, len);
2307         unsigned long i;
2308         unsigned long index = start >> PAGE_CACHE_SHIFT;
2309         struct extent_buffer *eb;
2310         struct page *p;
2311         struct address_space *mapping = tree->mapping;
2312         int uptodate = 1;
2313
2314         eb = __alloc_extent_buffer(tree, start, len, mask);
2315         if (!eb || IS_ERR(eb))
2316                 return NULL;
2317
2318         if (eb->flags & EXTENT_BUFFER_FILLED)
2319                 goto lru_add;
2320
2321         if (page0) {
2322                 eb->first_page = page0;
2323                 i = 1;
2324                 index++;
2325                 page_cache_get(page0);
2326                 mark_page_accessed(page0);
2327                 set_page_extent_mapped(page0);
2328                 WARN_ON(!PageUptodate(page0));
2329                 set_page_private(page0, EXTENT_PAGE_PRIVATE_FIRST_PAGE |
2330                                  len << 2);
2331         } else {
2332                 i = 0;
2333         }
2334         for (; i < num_pages; i++, index++) {
2335                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
2336                 if (!p) {
2337                         WARN_ON(1);
2338                         goto fail;
2339                 }
2340                 set_page_extent_mapped(p);
2341                 mark_page_accessed(p);
2342                 if (i == 0) {
2343                         eb->first_page = p;
2344                         set_page_private(p, EXTENT_PAGE_PRIVATE_FIRST_PAGE |
2345                                          len << 2);
2346                 } else {
2347                         set_page_private(p, EXTENT_PAGE_PRIVATE);
2348                 }
2349                 if (!PageUptodate(p))
2350                         uptodate = 0;
2351                 unlock_page(p);
2352         }
2353         if (uptodate)
2354                 eb->flags |= EXTENT_UPTODATE;
2355         eb->flags |= EXTENT_BUFFER_FILLED;
2356
2357 lru_add:
2358         spin_lock(&tree->lru_lock);
2359         add_lru(tree, eb);
2360         spin_unlock(&tree->lru_lock);
2361         return eb;
2362
2363 fail:
2364         spin_lock(&tree->lru_lock);
2365         list_del_init(&eb->lru);
2366         spin_unlock(&tree->lru_lock);
2367         if (!atomic_dec_and_test(&eb->refs))
2368                 return NULL;
2369         for (index = 1; index < i; index++) {
2370                 page_cache_release(extent_buffer_page(eb, index));
2371         }
2372         if (i > 0)
2373                 page_cache_release(extent_buffer_page(eb, 0));
2374         __free_extent_buffer(eb);
2375         return NULL;
2376 }
2377 EXPORT_SYMBOL(alloc_extent_buffer);
2378
2379 struct extent_buffer *find_extent_buffer(struct extent_map_tree *tree,
2380                                          u64 start, unsigned long len,
2381                                           gfp_t mask)
2382 {
2383         unsigned long num_pages = num_extent_pages(start, len);
2384         unsigned long i;
2385         unsigned long index = start >> PAGE_CACHE_SHIFT;
2386         struct extent_buffer *eb;
2387         struct page *p;
2388         struct address_space *mapping = tree->mapping;
2389         int uptodate = 1;
2390
2391         eb = __alloc_extent_buffer(tree, start, len, mask);
2392         if (!eb || IS_ERR(eb))
2393                 return NULL;
2394
2395         if (eb->flags & EXTENT_BUFFER_FILLED)
2396                 goto lru_add;
2397
2398         for (i = 0; i < num_pages; i++, index++) {
2399                 p = find_lock_page(mapping, index);
2400                 if (!p) {
2401                         goto fail;
2402                 }
2403                 set_page_extent_mapped(p);
2404                 mark_page_accessed(p);
2405
2406                 if (i == 0) {
2407                         eb->first_page = p;
2408                         set_page_private(p, EXTENT_PAGE_PRIVATE_FIRST_PAGE |
2409                                          len << 2);
2410                 } else {
2411                         set_page_private(p, EXTENT_PAGE_PRIVATE);
2412                 }
2413
2414                 if (!PageUptodate(p))
2415                         uptodate = 0;
2416                 unlock_page(p);
2417         }
2418         if (uptodate)
2419                 eb->flags |= EXTENT_UPTODATE;
2420         eb->flags |= EXTENT_BUFFER_FILLED;
2421
2422 lru_add:
2423         spin_lock(&tree->lru_lock);
2424         add_lru(tree, eb);
2425         spin_unlock(&tree->lru_lock);
2426         return eb;
2427 fail:
2428         spin_lock(&tree->lru_lock);
2429         list_del_init(&eb->lru);
2430         spin_unlock(&tree->lru_lock);
2431         if (!atomic_dec_and_test(&eb->refs))
2432                 return NULL;
2433         for (index = 1; index < i; index++) {
2434                 page_cache_release(extent_buffer_page(eb, index));
2435         }
2436         if (i > 0)
2437                 page_cache_release(extent_buffer_page(eb, 0));
2438         __free_extent_buffer(eb);
2439         return NULL;
2440 }
2441 EXPORT_SYMBOL(find_extent_buffer);
2442
2443 void free_extent_buffer(struct extent_buffer *eb)
2444 {
2445         unsigned long i;
2446         unsigned long num_pages;
2447
2448         if (!eb)
2449                 return;
2450
2451         if (!atomic_dec_and_test(&eb->refs))
2452                 return;
2453
2454         WARN_ON(!list_empty(&eb->lru));
2455         num_pages = num_extent_pages(eb->start, eb->len);
2456
2457         for (i = 1; i < num_pages; i++) {
2458                 page_cache_release(extent_buffer_page(eb, i));
2459         }
2460         page_cache_release(extent_buffer_page(eb, 0));
2461         __free_extent_buffer(eb);
2462 }
2463 EXPORT_SYMBOL(free_extent_buffer);
2464
2465 int clear_extent_buffer_dirty(struct extent_map_tree *tree,
2466                               struct extent_buffer *eb)
2467 {
2468         int set;
2469         unsigned long i;
2470         unsigned long num_pages;
2471         struct page *page;
2472
2473         u64 start = eb->start;
2474         u64 end = start + eb->len - 1;
2475
2476         set = clear_extent_dirty(tree, start, end, GFP_NOFS);
2477         num_pages = num_extent_pages(eb->start, eb->len);
2478
2479         for (i = 0; i < num_pages; i++) {
2480                 page = extent_buffer_page(eb, i);
2481                 lock_page(page);
2482                 /*
2483                  * if we're on the last page or the first page and the
2484                  * block isn't aligned on a page boundary, do extra checks
2485                  * to make sure we don't clean page that is partially dirty
2486                  */
2487                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
2488                     ((i == num_pages - 1) &&
2489                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
2490                         start = (u64)page->index << PAGE_CACHE_SHIFT;
2491                         end  = start + PAGE_CACHE_SIZE - 1;
2492                         if (test_range_bit(tree, start, end,
2493                                            EXTENT_DIRTY, 0)) {
2494                                 unlock_page(page);
2495                                 continue;
2496                         }
2497                 }
2498                 clear_page_dirty_for_io(page);
2499                 write_lock_irq(&page->mapping->tree_lock);
2500                 if (!PageDirty(page)) {
2501                         radix_tree_tag_clear(&page->mapping->page_tree,
2502                                                 page_index(page),
2503                                                 PAGECACHE_TAG_DIRTY);
2504                 }
2505                 write_unlock_irq(&page->mapping->tree_lock);
2506                 unlock_page(page);
2507         }
2508         return 0;
2509 }
2510 EXPORT_SYMBOL(clear_extent_buffer_dirty);
2511
2512 int wait_on_extent_buffer_writeback(struct extent_map_tree *tree,
2513                                     struct extent_buffer *eb)
2514 {
2515         return wait_on_extent_writeback(tree, eb->start,
2516                                         eb->start + eb->len - 1);
2517 }
2518 EXPORT_SYMBOL(wait_on_extent_buffer_writeback);
2519
2520 int set_extent_buffer_dirty(struct extent_map_tree *tree,
2521                              struct extent_buffer *eb)
2522 {
2523         unsigned long i;
2524         unsigned long num_pages;
2525
2526         num_pages = num_extent_pages(eb->start, eb->len);
2527         for (i = 0; i < num_pages; i++) {
2528                 struct page *page = extent_buffer_page(eb, i);
2529                 /* writepage may need to do something special for the
2530                  * first page, we have to make sure page->private is
2531                  * properly set.  releasepage may drop page->private
2532                  * on us if the page isn't already dirty.
2533                  */
2534                 if (i == 0) {
2535                         lock_page(page);
2536                         set_page_private(page,
2537                                          EXTENT_PAGE_PRIVATE_FIRST_PAGE |
2538                                          eb->len << 2);
2539                 }
2540                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
2541                 if (i == 0)
2542                         unlock_page(page);
2543         }
2544         return set_extent_dirty(tree, eb->start,
2545                                 eb->start + eb->len - 1, GFP_NOFS);
2546 }
2547 EXPORT_SYMBOL(set_extent_buffer_dirty);
2548
2549 int set_extent_buffer_uptodate(struct extent_map_tree *tree,
2550                                 struct extent_buffer *eb)
2551 {
2552         unsigned long i;
2553         struct page *page;
2554         unsigned long num_pages;
2555
2556         num_pages = num_extent_pages(eb->start, eb->len);
2557
2558         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
2559                             GFP_NOFS);
2560         for (i = 0; i < num_pages; i++) {
2561                 page = extent_buffer_page(eb, i);
2562                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
2563                     ((i == num_pages - 1) &&
2564                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
2565                         check_page_uptodate(tree, page);
2566                         continue;
2567                 }
2568                 SetPageUptodate(page);
2569         }
2570         return 0;
2571 }
2572 EXPORT_SYMBOL(set_extent_buffer_uptodate);
2573
2574 int extent_buffer_uptodate(struct extent_map_tree *tree,
2575                              struct extent_buffer *eb)
2576 {
2577         if (eb->flags & EXTENT_UPTODATE)
2578                 return 1;
2579         return test_range_bit(tree, eb->start, eb->start + eb->len - 1,
2580                            EXTENT_UPTODATE, 1);
2581 }
2582 EXPORT_SYMBOL(extent_buffer_uptodate);
2583
2584 int read_extent_buffer_pages(struct extent_map_tree *tree,
2585                              struct extent_buffer *eb,
2586                              u64 start,
2587                              int wait)
2588 {
2589         unsigned long i;
2590         unsigned long start_i;
2591         struct page *page;
2592         int err;
2593         int ret = 0;
2594         unsigned long num_pages;
2595
2596         if (eb->flags & EXTENT_UPTODATE)
2597                 return 0;
2598
2599         if (0 && test_range_bit(tree, eb->start, eb->start + eb->len - 1,
2600                            EXTENT_UPTODATE, 1)) {
2601                 return 0;
2602         }
2603
2604         if (start) {
2605                 WARN_ON(start < eb->start);
2606                 start_i = (start >> PAGE_CACHE_SHIFT) -
2607                         (eb->start >> PAGE_CACHE_SHIFT);
2608         } else {
2609                 start_i = 0;
2610         }
2611
2612         num_pages = num_extent_pages(eb->start, eb->len);
2613         for (i = start_i; i < num_pages; i++) {
2614                 page = extent_buffer_page(eb, i);
2615                 if (PageUptodate(page)) {
2616                         continue;
2617                 }
2618                 if (!wait) {
2619                         if (TestSetPageLocked(page)) {
2620                                 continue;
2621                         }
2622                 } else {
2623                         lock_page(page);
2624                 }
2625                 if (!PageUptodate(page)) {
2626                         err = page->mapping->a_ops->readpage(NULL, page);
2627                         if (err) {
2628                                 ret = err;
2629                         }
2630                 } else {
2631                         unlock_page(page);
2632                 }
2633         }
2634
2635         if (ret || !wait) {
2636                 return ret;
2637         }
2638
2639         for (i = start_i; i < num_pages; i++) {
2640                 page = extent_buffer_page(eb, i);
2641                 wait_on_page_locked(page);
2642                 if (!PageUptodate(page)) {
2643                         ret = -EIO;
2644                 }
2645         }
2646         if (!ret)
2647                 eb->flags |= EXTENT_UPTODATE;
2648         return ret;
2649 }
2650 EXPORT_SYMBOL(read_extent_buffer_pages);
2651
2652 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
2653                         unsigned long start,
2654                         unsigned long len)
2655 {
2656         size_t cur;
2657         size_t offset;
2658         struct page *page;
2659         char *kaddr;
2660         char *dst = (char *)dstv;
2661         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2662         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2663         unsigned long num_pages = num_extent_pages(eb->start, eb->len);
2664
2665         WARN_ON(start > eb->len);
2666         WARN_ON(start + len > eb->start + eb->len);
2667
2668         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
2669
2670         while(len > 0) {
2671                 page = extent_buffer_page(eb, i);
2672                 if (!PageUptodate(page)) {
2673                         printk("page %lu not up to date i %lu, total %lu, len %lu\n", page->index, i, num_pages, eb->len);
2674                         WARN_ON(1);
2675                 }
2676                 WARN_ON(!PageUptodate(page));
2677
2678                 cur = min(len, (PAGE_CACHE_SIZE - offset));
2679                 kaddr = kmap_atomic(page, KM_USER1);
2680                 memcpy(dst, kaddr + offset, cur);
2681                 kunmap_atomic(kaddr, KM_USER1);
2682
2683                 dst += cur;
2684                 len -= cur;
2685                 offset = 0;
2686                 i++;
2687         }
2688 }
2689 EXPORT_SYMBOL(read_extent_buffer);
2690
2691 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
2692                                unsigned long min_len, char **token, char **map,
2693                                unsigned long *map_start,
2694                                unsigned long *map_len, int km)
2695 {
2696         size_t offset = start & (PAGE_CACHE_SIZE - 1);
2697         char *kaddr;
2698         struct page *p;
2699         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2700         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2701         unsigned long end_i = (start_offset + start + min_len - 1) >>
2702                 PAGE_CACHE_SHIFT;
2703
2704         if (i != end_i)
2705                 return -EINVAL;
2706
2707         if (i == 0) {
2708                 offset = start_offset;
2709                 *map_start = 0;
2710         } else {
2711                 offset = 0;
2712                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
2713         }
2714         if (start + min_len > eb->len) {
2715 printk("bad mapping eb start %Lu len %lu, wanted %lu %lu\n", eb->start, eb->len, start, min_len);
2716                 WARN_ON(1);
2717         }
2718
2719         p = extent_buffer_page(eb, i);
2720         WARN_ON(!PageUptodate(p));
2721         kaddr = kmap_atomic(p, km);
2722         *token = kaddr;
2723         *map = kaddr + offset;
2724         *map_len = PAGE_CACHE_SIZE - offset;
2725         return 0;
2726 }
2727 EXPORT_SYMBOL(map_private_extent_buffer);
2728
2729 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
2730                       unsigned long min_len,
2731                       char **token, char **map,
2732                       unsigned long *map_start,
2733                       unsigned long *map_len, int km)
2734 {
2735         int err;
2736         int save = 0;
2737         if (eb->map_token) {
2738                 unmap_extent_buffer(eb, eb->map_token, km);
2739                 eb->map_token = NULL;
2740                 save = 1;
2741         }
2742         err = map_private_extent_buffer(eb, start, min_len, token, map,
2743                                        map_start, map_len, km);
2744         if (!err && save) {
2745                 eb->map_token = *token;
2746                 eb->kaddr = *map;
2747                 eb->map_start = *map_start;
2748                 eb->map_len = *map_len;
2749         }
2750         return err;
2751 }
2752 EXPORT_SYMBOL(map_extent_buffer);
2753
2754 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
2755 {
2756         kunmap_atomic(token, km);
2757 }
2758 EXPORT_SYMBOL(unmap_extent_buffer);
2759
2760 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
2761                           unsigned long start,
2762                           unsigned long len)
2763 {
2764         size_t cur;
2765         size_t offset;
2766         struct page *page;
2767         char *kaddr;
2768         char *ptr = (char *)ptrv;
2769         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2770         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2771         int ret = 0;
2772
2773         WARN_ON(start > eb->len);
2774         WARN_ON(start + len > eb->start + eb->len);
2775
2776         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
2777
2778         while(len > 0) {
2779                 page = extent_buffer_page(eb, i);
2780                 WARN_ON(!PageUptodate(page));
2781
2782                 cur = min(len, (PAGE_CACHE_SIZE - offset));
2783
2784                 kaddr = kmap_atomic(page, KM_USER0);
2785                 ret = memcmp(ptr, kaddr + offset, cur);
2786                 kunmap_atomic(kaddr, KM_USER0);
2787                 if (ret)
2788                         break;
2789
2790                 ptr += cur;
2791                 len -= cur;
2792                 offset = 0;
2793                 i++;
2794         }
2795         return ret;
2796 }
2797 EXPORT_SYMBOL(memcmp_extent_buffer);
2798
2799 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
2800                          unsigned long start, unsigned long len)
2801 {
2802         size_t cur;
2803         size_t offset;
2804         struct page *page;
2805         char *kaddr;
2806         char *src = (char *)srcv;
2807         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2808         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2809
2810         WARN_ON(start > eb->len);
2811         WARN_ON(start + len > eb->start + eb->len);
2812
2813         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
2814
2815         while(len > 0) {
2816                 page = extent_buffer_page(eb, i);
2817                 WARN_ON(!PageUptodate(page));
2818
2819                 cur = min(len, PAGE_CACHE_SIZE - offset);
2820                 kaddr = kmap_atomic(page, KM_USER1);
2821                 memcpy(kaddr + offset, src, cur);
2822                 kunmap_atomic(kaddr, KM_USER1);
2823
2824                 src += cur;
2825                 len -= cur;
2826                 offset = 0;
2827                 i++;
2828         }
2829 }
2830 EXPORT_SYMBOL(write_extent_buffer);
2831
2832 void memset_extent_buffer(struct extent_buffer *eb, char c,
2833                           unsigned long start, unsigned long len)
2834 {
2835         size_t cur;
2836         size_t offset;
2837         struct page *page;
2838         char *kaddr;
2839         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2840         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2841
2842         WARN_ON(start > eb->len);
2843         WARN_ON(start + len > eb->start + eb->len);
2844
2845         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
2846
2847         while(len > 0) {
2848                 page = extent_buffer_page(eb, i);
2849                 WARN_ON(!PageUptodate(page));
2850
2851                 cur = min(len, PAGE_CACHE_SIZE - offset);
2852                 kaddr = kmap_atomic(page, KM_USER0);
2853                 memset(kaddr + offset, c, cur);
2854                 kunmap_atomic(kaddr, KM_USER0);
2855
2856                 len -= cur;
2857                 offset = 0;
2858                 i++;
2859         }
2860 }
2861 EXPORT_SYMBOL(memset_extent_buffer);
2862
2863 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
2864                         unsigned long dst_offset, unsigned long src_offset,
2865                         unsigned long len)
2866 {
2867         u64 dst_len = dst->len;
2868         size_t cur;
2869         size_t offset;
2870         struct page *page;
2871         char *kaddr;
2872         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
2873         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
2874
2875         WARN_ON(src->len != dst_len);
2876
2877         offset = (start_offset + dst_offset) &
2878                 ((unsigned long)PAGE_CACHE_SIZE - 1);
2879
2880         while(len > 0) {
2881                 page = extent_buffer_page(dst, i);
2882                 WARN_ON(!PageUptodate(page));
2883
2884                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
2885
2886                 kaddr = kmap_atomic(page, KM_USER0);
2887                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
2888                 kunmap_atomic(kaddr, KM_USER0);
2889
2890                 src_offset += cur;
2891                 len -= cur;
2892                 offset = 0;
2893                 i++;
2894         }
2895 }
2896 EXPORT_SYMBOL(copy_extent_buffer);
2897
2898 static void move_pages(struct page *dst_page, struct page *src_page,
2899                        unsigned long dst_off, unsigned long src_off,
2900                        unsigned long len)
2901 {
2902         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
2903         if (dst_page == src_page) {
2904                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
2905         } else {
2906                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
2907                 char *p = dst_kaddr + dst_off + len;
2908                 char *s = src_kaddr + src_off + len;
2909
2910                 while (len--)
2911                         *--p = *--s;
2912
2913                 kunmap_atomic(src_kaddr, KM_USER1);
2914         }
2915         kunmap_atomic(dst_kaddr, KM_USER0);
2916 }
2917
2918 static void copy_pages(struct page *dst_page, struct page *src_page,
2919                        unsigned long dst_off, unsigned long src_off,
2920                        unsigned long len)
2921 {
2922         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
2923         char *src_kaddr;
2924
2925         if (dst_page != src_page)
2926                 src_kaddr = kmap_atomic(src_page, KM_USER1);
2927         else
2928                 src_kaddr = dst_kaddr;
2929
2930         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
2931         kunmap_atomic(dst_kaddr, KM_USER0);
2932         if (dst_page != src_page)
2933                 kunmap_atomic(src_kaddr, KM_USER1);
2934 }
2935
2936 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
2937                            unsigned long src_offset, unsigned long len)
2938 {
2939         size_t cur;
2940         size_t dst_off_in_page;
2941         size_t src_off_in_page;
2942         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
2943         unsigned long dst_i;
2944         unsigned long src_i;
2945
2946         if (src_offset + len > dst->len) {
2947                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
2948                        src_offset, len, dst->len);
2949                 BUG_ON(1);
2950         }
2951         if (dst_offset + len > dst->len) {
2952                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
2953                        dst_offset, len, dst->len);
2954                 BUG_ON(1);
2955         }
2956
2957         while(len > 0) {
2958                 dst_off_in_page = (start_offset + dst_offset) &
2959                         ((unsigned long)PAGE_CACHE_SIZE - 1);
2960                 src_off_in_page = (start_offset + src_offset) &
2961                         ((unsigned long)PAGE_CACHE_SIZE - 1);
2962
2963                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
2964                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
2965
2966                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
2967                                                src_off_in_page));
2968                 cur = min_t(unsigned long, cur,
2969                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
2970
2971                 copy_pages(extent_buffer_page(dst, dst_i),
2972                            extent_buffer_page(dst, src_i),
2973                            dst_off_in_page, src_off_in_page, cur);
2974
2975                 src_offset += cur;
2976                 dst_offset += cur;
2977                 len -= cur;
2978         }
2979 }
2980 EXPORT_SYMBOL(memcpy_extent_buffer);
2981
2982 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
2983                            unsigned long src_offset, unsigned long len)
2984 {
2985         size_t cur;
2986         size_t dst_off_in_page;
2987         size_t src_off_in_page;
2988         unsigned long dst_end = dst_offset + len - 1;
2989         unsigned long src_end = src_offset + len - 1;
2990         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
2991         unsigned long dst_i;
2992         unsigned long src_i;
2993
2994         if (src_offset + len > dst->len) {
2995                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
2996                        src_offset, len, dst->len);
2997                 BUG_ON(1);
2998         }
2999         if (dst_offset + len > dst->len) {
3000                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
3001                        dst_offset, len, dst->len);
3002                 BUG_ON(1);
3003         }
3004         if (dst_offset < src_offset) {
3005                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3006                 return;
3007         }
3008         while(len > 0) {
3009                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3010                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3011
3012                 dst_off_in_page = (start_offset + dst_end) &
3013                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3014                 src_off_in_page = (start_offset + src_end) &
3015                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3016
3017                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3018                 cur = min(cur, dst_off_in_page + 1);
3019                 move_pages(extent_buffer_page(dst, dst_i),
3020                            extent_buffer_page(dst, src_i),
3021                            dst_off_in_page - cur + 1,
3022                            src_off_in_page - cur + 1, cur);
3023
3024                 dst_end -= cur;
3025                 src_end -= cur;
3026                 len -= cur;
3027         }
3028 }
3029 EXPORT_SYMBOL(memmove_extent_buffer);