SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/btrfs/extent_io.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_io.h"
  16 #include "extent_map.h"
  17 #include "compat.h"
  18 #include "ctree.h"
  19 #include "btrfs_inode.h"
  20
  21 /* temporary define until extent_map moves out of btrfs */
  22 struct kmem_cache *btrfs_cache_create(const char *name, size_t size,
  23                                        unsigned long extra_flags,
  24                                        void (*ctor)(void *, struct kmem_cache *,
  25                                                     unsigned long));
  26
  27 static struct kmem_cache *extent_state_cache;
  28 static struct kmem_cache *extent_buffer_cache;
  29
  30 static LIST_HEAD(buffers);
  31 static LIST_HEAD(states);
  32
  33 #ifdef LEAK_DEBUG
  34 static spinlock_t leak_lock = SPIN_LOCK_UNLOCKED;
  35 #endif
  36
  37 #define BUFFER_LRU_MAX 64
  38
  39 struct tree_entry {
  40         u64 start;
  41         u64 end;
  42         struct rb_node rb_node;
  43 };
  44
  45 struct extent_page_data {
  46         struct bio *bio;
  47         struct extent_io_tree *tree;
  48         get_extent_t *get_extent;
  49 };
  50
  51 int __init extent_io_init(void)
  52 {
  53         extent_state_cache = btrfs_cache_create("extent_state",
  54                                             sizeof(struct extent_state), 0,
  55                                             NULL);
  56         if (!extent_state_cache)
  57                 return -ENOMEM;
  58
  59         extent_buffer_cache = btrfs_cache_create("extent_buffers",
  60                                             sizeof(struct extent_buffer), 0,
  61                                             NULL);
  62         if (!extent_buffer_cache)
  63                 goto free_state_cache;
  64         return 0;
  65
  66 free_state_cache:
  67         kmem_cache_destroy(extent_state_cache);
  68         return -ENOMEM;
  69 }
  70
  71 void extent_io_exit(void)
  72 {
  73         struct extent_state *state;
  74         struct extent_buffer *eb;
  75
  76         while (!list_empty(&states)) {
  77                 state = list_entry(states.next, struct extent_state, leak_list);
  78                 printk("state leak: start %Lu end %Lu state %lu in tree %p refs %d\n", state->start, state->end, state->state, state->tree, atomic_read(&state->refs));
  79                 list_del(&state->leak_list);
  80                 kmem_cache_free(extent_state_cache, state);
  81
  82         }
  83
  84         while (!list_empty(&buffers)) {
  85                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  86                 printk("buffer leak start %Lu len %lu refs %d\n", eb->start, eb->len, atomic_read(&eb->refs));
  87                 list_del(&eb->leak_list);
  88                 kmem_cache_free(extent_buffer_cache, eb);
  89         }
  90         if (extent_state_cache)
  91                 kmem_cache_destroy(extent_state_cache);
  92         if (extent_buffer_cache)
  93                 kmem_cache_destroy(extent_buffer_cache);
  94 }
  95
  96 void extent_io_tree_init(struct extent_io_tree *tree,
  97                           struct address_space *mapping, gfp_t mask)
  98 {
  99         tree->state.rb_node = NULL;
 100         tree->buffer.rb_node = NULL;
 101         tree->ops = NULL;
 102         tree->dirty_bytes = 0;
 103         spin_lock_init(&tree->lock);
 104         spin_lock_init(&tree->buffer_lock);
 105         tree->mapping = mapping;
 106 }
 107 EXPORT_SYMBOL(extent_io_tree_init);
 108
 109 struct extent_state *alloc_extent_state(gfp_t mask)
 110 {
 111         struct extent_state *state;
 112 #ifdef LEAK_DEBUG
 113         unsigned long flags;
 114 #endif
 115
 116         state = kmem_cache_alloc(extent_state_cache, mask);
 117         if (!state)
 118                 return state;
 119         state->state = 0;
 120         state->private = 0;
 121         state->tree = NULL;
 122 #ifdef LEAK_DEBUG
 123         spin_lock_irqsave(&leak_lock, flags);
 124         list_add(&state->leak_list, &states);
 125         spin_unlock_irqrestore(&leak_lock, flags);
 126 #endif
 127         atomic_set(&state->refs, 1);
 128         init_waitqueue_head(&state->wq);
 129         return state;
 130 }
 131 EXPORT_SYMBOL(alloc_extent_state);
 132
 133 void free_extent_state(struct extent_state *state)
 134 {
 135         if (!state)
 136                 return;
 137         if (atomic_dec_and_test(&state->refs)) {
 138 #ifdef LEAK_DEBUG
 139                 unsigned long flags;
 140 #endif
 141                 WARN_ON(state->tree);
 142 #ifdef LEAK_DEBUG
 143                 spin_lock_irqsave(&leak_lock, flags);
 144                 list_del(&state->leak_list);
 145                 spin_unlock_irqrestore(&leak_lock, flags);
 146 #endif
 147                 kmem_cache_free(extent_state_cache, state);
 148         }
 149 }
 150 EXPORT_SYMBOL(free_extent_state);
 151
 152 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 153                                    struct rb_node *node)
 154 {
 155         struct rb_node ** p = &root->rb_node;
 156         struct rb_node * parent = NULL;
 157         struct tree_entry *entry;
 158
 159         while(*p) {
 160                 parent = *p;
 161                 entry = rb_entry(parent, struct tree_entry, rb_node);
 162
 163                 if (offset < entry->start)
 164                         p = &(*p)->rb_left;
 165                 else if (offset > entry->end)
 166                         p = &(*p)->rb_right;
 167                 else
 168                         return parent;
 169         }
 170
 171         entry = rb_entry(node, struct tree_entry, rb_node);
 172         rb_link_node(node, parent, p);
 173         rb_insert_color(node, root);
 174         return NULL;
 175 }
 176
 177 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 178                                      struct rb_node **prev_ret,
 179                                      struct rb_node **next_ret)
 180 {
 181         struct rb_root *root = &tree->state;
 182         struct rb_node * n = root->rb_node;
 183         struct rb_node *prev = NULL;
 184         struct rb_node *orig_prev = NULL;
 185         struct tree_entry *entry;
 186         struct tree_entry *prev_entry = NULL;
 187
 188         while(n) {
 189                 entry = rb_entry(n, struct tree_entry, rb_node);
 190                 prev = n;
 191                 prev_entry = entry;
 192
 193                 if (offset < entry->start)
 194                         n = n->rb_left;
 195                 else if (offset > entry->end)
 196                         n = n->rb_right;
 197                 else {
 198                         return n;
 199                 }
 200         }
 201
 202         if (prev_ret) {
 203                 orig_prev = prev;
 204                 while(prev && offset > prev_entry->end) {
 205                         prev = rb_next(prev);
 206                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 207                 }
 208                 *prev_ret = prev;
 209                 prev = orig_prev;
 210         }
 211
 212         if (next_ret) {
 213                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 214                 while(prev && offset < prev_entry->start) {
 215                         prev = rb_prev(prev);
 216                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 217                 }
 218                 *next_ret = prev;
 219         }
 220         return NULL;
 221 }
 222
 223 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 224                                           u64 offset)
 225 {
 226         struct rb_node *prev = NULL;
 227         struct rb_node *ret;
 228
 229         ret = __etree_search(tree, offset, &prev, NULL);
 230         if (!ret) {
 231                 return prev;
 232         }
 233         return ret;
 234 }
 235
 236 static struct extent_buffer *buffer_tree_insert(struct extent_io_tree *tree,
 237                                           u64 offset, struct rb_node *node)
 238 {
 239         struct rb_root *root = &tree->buffer;
 240         struct rb_node ** p = &root->rb_node;
 241         struct rb_node * parent = NULL;
 242         struct extent_buffer *eb;
 243
 244         while(*p) {
 245                 parent = *p;
 246                 eb = rb_entry(parent, struct extent_buffer, rb_node);
 247
 248                 if (offset < eb->start)
 249                         p = &(*p)->rb_left;
 250                 else if (offset > eb->start)
 251                         p = &(*p)->rb_right;
 252                 else
 253                         return eb;
 254         }
 255
 256         rb_link_node(node, parent, p);
 257         rb_insert_color(node, root);
 258         return NULL;
 259 }
 260
 261 static struct extent_buffer *buffer_search(struct extent_io_tree *tree,
 262                                            u64 offset)
 263 {
 264         struct rb_root *root = &tree->buffer;
 265         struct rb_node * n = root->rb_node;
 266         struct extent_buffer *eb;
 267
 268         while(n) {
 269                 eb = rb_entry(n, struct extent_buffer, rb_node);
 270                 if (offset < eb->start)
 271                         n = n->rb_left;
 272                 else if (offset > eb->start)
 273                         n = n->rb_right;
 274                 else
 275                         return eb;
 276         }
 277         return NULL;
 278 }
 279
 280 /*
 281  * utility function to look for merge candidates inside a given range.
 282  * Any extents with matching state are merged together into a single
 283  * extent in the tree.  Extents with EXTENT_IO in their state field
 284  * are not merged because the end_io handlers need to be able to do
 285  * operations on them without sleeping (or doing allocations/splits).
 286  *
 287  * This should be called with the tree lock held.
 288  */
 289 static int merge_state(struct extent_io_tree *tree,
 290                        struct extent_state *state)
 291 {
 292         struct extent_state *other;
 293         struct rb_node *other_node;
 294
 295         if (state->state & EXTENT_IOBITS)
 296                 return 0;
 297
 298         other_node = rb_prev(&state->rb_node);
 299         if (other_node) {
 300                 other = rb_entry(other_node, struct extent_state, rb_node);
 301                 if (other->end == state->start - 1 &&
 302                     other->state == state->state) {
 303                         state->start = other->start;
 304                         other->tree = NULL;
 305                         rb_erase(&other->rb_node, &tree->state);
 306                         free_extent_state(other);
 307                 }
 308         }
 309         other_node = rb_next(&state->rb_node);
 310         if (other_node) {
 311                 other = rb_entry(other_node, struct extent_state, rb_node);
 312                 if (other->start == state->end + 1 &&
 313                     other->state == state->state) {
 314                         other->start = state->start;
 315                         state->tree = NULL;
 316                         rb_erase(&state->rb_node, &tree->state);
 317                         free_extent_state(state);
 318                 }
 319         }
 320         return 0;
 321 }
 322
 323 static void set_state_cb(struct extent_io_tree *tree,
 324                          struct extent_state *state,
 325                          unsigned long bits)
 326 {
 327         if (tree->ops && tree->ops->set_bit_hook) {
 328                 tree->ops->set_bit_hook(tree->mapping->host, state->start,
 329                                         state->end, state->state, bits);
 330         }
 331 }
 332
 333 static void clear_state_cb(struct extent_io_tree *tree,
 334                            struct extent_state *state,
 335                            unsigned long bits)
 336 {
 337         if (tree->ops && tree->ops->set_bit_hook) {
 338                 tree->ops->clear_bit_hook(tree->mapping->host, state->start,
 339                                           state->end, state->state, bits);
 340         }
 341 }
 342
 343 /*
 344  * insert an extent_state struct into the tree.  'bits' are set on the
 345  * struct before it is inserted.
 346  *
 347  * This may return -EEXIST if the extent is already there, in which case the
 348  * state struct is freed.
 349  *
 350  * The tree lock is not taken internally.  This is a utility function and
 351  * probably isn't what you want to call (see set/clear_extent_bit).
 352  */
 353 static int insert_state(struct extent_io_tree *tree,
 354                         struct extent_state *state, u64 start, u64 end,
 355                         int bits)
 356 {
 357         struct rb_node *node;
 358
 359         if (end < start) {
 360                 printk("end < start %Lu %Lu\n", end, start);
 361                 WARN_ON(1);
 362         }
 363         if (bits & EXTENT_DIRTY)
 364                 tree->dirty_bytes += end - start + 1;
 365         set_state_cb(tree, state, bits);
 366         state->state |= bits;
 367         state->start = start;
 368         state->end = end;
 369         node = tree_insert(&tree->state, end, &state->rb_node);
 370         if (node) {
 371                 struct extent_state *found;
 372                 found = rb_entry(node, struct extent_state, rb_node);
 373                 printk("found node %Lu %Lu on insert of %Lu %Lu\n", found->start, found->end, start, end);
 374                 free_extent_state(state);
 375                 return -EEXIST;
 376         }
 377         state->tree = tree;
 378         merge_state(tree, state);
 379         return 0;
 380 }
 381
 382 /*
 383  * split a given extent state struct in two, inserting the preallocated
 384  * struct 'prealloc' as the newly created second half.  'split' indicates an
 385  * offset inside 'orig' where it should be split.
 386  *
 387  * Before calling,
 388  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 389  * are two extent state structs in the tree:
 390  * prealloc: [orig->start, split - 1]
 391  * orig: [ split, orig->end ]
 392  *
 393  * The tree locks are not taken by this function. They need to be held
 394  * by the caller.
 395  */
 396 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 397                        struct extent_state *prealloc, u64 split)
 398 {
 399         struct rb_node *node;
 400         prealloc->start = orig->start;
 401         prealloc->end = split - 1;
 402         prealloc->state = orig->state;
 403         orig->start = split;
 404
 405         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 406         if (node) {
 407                 struct extent_state *found;
 408                 found = rb_entry(node, struct extent_state, rb_node);
 409                 printk("found node %Lu %Lu on insert of %Lu %Lu\n", found->start, found->end, prealloc->start, prealloc->end);
 410                 free_extent_state(prealloc);
 411                 return -EEXIST;
 412         }
 413         prealloc->tree = tree;
 414         return 0;
 415 }
 416
 417 /*
 418  * utility function to clear some bits in an extent state struct.
 419  * it will optionally wake up any one waiting on this state (wake == 1), or
 420  * forcibly remove the state from the tree (delete == 1).
 421  *
 422  * If no bits are set on the state struct after clearing things, the
 423  * struct is freed and removed from the tree
 424  */
 425 static int clear_state_bit(struct extent_io_tree *tree,
 426                             struct extent_state *state, int bits, int wake,
 427                             int delete)
 428 {
 429         int ret = state->state & bits;
 430
 431         if ((bits & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 432                 u64 range = state->end - state->start + 1;
 433                 WARN_ON(range > tree->dirty_bytes);
 434                 tree->dirty_bytes -= range;
 435         }
 436         clear_state_cb(tree, state, bits);
 437         state->state &= ~bits;
 438         if (wake)
 439                 wake_up(&state->wq);
 440         if (delete || state->state == 0) {
 441                 if (state->tree) {
 442                         clear_state_cb(tree, state, state->state);
 443                         rb_erase(&state->rb_node, &tree->state);
 444                         state->tree = NULL;
 445                         free_extent_state(state);
 446                 } else {
 447                         WARN_ON(1);
 448                 }
 449         } else {
 450                 merge_state(tree, state);
 451         }
 452         return ret;
 453 }
 454
 455 /*
 456  * clear some bits on a range in the tree.  This may require splitting
 457  * or inserting elements in the tree, so the gfp mask is used to
 458  * indicate which allocations or sleeping are allowed.
 459  *
 460  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 461  * the given range from the tree regardless of state (ie for truncate).
 462  *
 463  * the range [start, end] is inclusive.
 464  *
 465  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 466  * bits were already set, or zero if none of the bits were already set.
 467  */
 468 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 469                      int bits, int wake, int delete, gfp_t mask)
 470 {
 471         struct extent_state *state;
 472         struct extent_state *prealloc = NULL;
 473         struct rb_node *node;
 474         unsigned long flags;
 475         int err;
 476         int set = 0;
 477
 478 again:
 479         if (!prealloc && (mask & __GFP_WAIT)) {
 480                 prealloc = alloc_extent_state(mask);
 481                 if (!prealloc)
 482                         return -ENOMEM;
 483         }
 484
 485         spin_lock_irqsave(&tree->lock, flags);
 486         /*
 487          * this search will find the extents that end after
 488          * our range starts
 489          */
 490         node = tree_search(tree, start);
 491         if (!node)
 492                 goto out;
 493         state = rb_entry(node, struct extent_state, rb_node);
 494         if (state->start > end)
 495                 goto out;
 496         WARN_ON(state->end < start);
 497
 498         /*
 499          *     | ---- desired range ---- |
 500          *  | state | or
 501          *  | ------------- state -------------- |
 502          *
 503          * We need to split the extent we found, and may flip
 504          * bits on second half.
 505          *
 506          * If the extent we found extends past our range, we
 507          * just split and search again.  It'll get split again
 508          * the next time though.
 509          *
 510          * If the extent we found is inside our range, we clear
 511          * the desired bit on it.
 512          */
 513
 514         if (state->start < start) {
 515                 if (!prealloc)
 516                         prealloc = alloc_extent_state(GFP_ATOMIC);
 517                 err = split_state(tree, state, prealloc, start);
 518                 BUG_ON(err == -EEXIST);
 519                 prealloc = NULL;
 520                 if (err)
 521                         goto out;
 522                 if (state->end <= end) {
 523                         start = state->end + 1;
 524                         set |= clear_state_bit(tree, state, bits,
 525                                         wake, delete);
 526                 } else {
 527                         start = state->start;
 528                 }
 529                 goto search_again;
 530         }
 531         /*
 532          * | ---- desired range ---- |
 533          *                        | state |
 534          * We need to split the extent, and clear the bit
 535          * on the first half
 536          */
 537         if (state->start <= end && state->end > end) {
 538                 if (!prealloc)
 539                         prealloc = alloc_extent_state(GFP_ATOMIC);
 540                 err = split_state(tree, state, prealloc, end + 1);
 541                 BUG_ON(err == -EEXIST);
 542
 543                 if (wake)
 544                         wake_up(&state->wq);
 545                 set |= clear_state_bit(tree, prealloc, bits,
 546                                        wake, delete);
 547                 prealloc = NULL;
 548                 goto out;
 549         }
 550
 551         start = state->end + 1;
 552         set |= clear_state_bit(tree, state, bits, wake, delete);
 553         goto search_again;
 554
 555 out:
 556         spin_unlock_irqrestore(&tree->lock, flags);
 557         if (prealloc)
 558                 free_extent_state(prealloc);
 559
 560         return set;
 561
 562 search_again:
 563         if (start > end)
 564                 goto out;
 565         spin_unlock_irqrestore(&tree->lock, flags);
 566         if (mask & __GFP_WAIT)
 567                 cond_resched();
 568         goto again;
 569 }
 570 EXPORT_SYMBOL(clear_extent_bit);
 571
 572 static int wait_on_state(struct extent_io_tree *tree,
 573                          struct extent_state *state)
 574 {
 575         DEFINE_WAIT(wait);
 576         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 577         spin_unlock_irq(&tree->lock);
 578         schedule();
 579         spin_lock_irq(&tree->lock);
 580         finish_wait(&state->wq, &wait);
 581         return 0;
 582 }
 583
 584 /*
 585  * waits for one or more bits to clear on a range in the state tree.
 586  * The range [start, end] is inclusive.
 587  * The tree lock is taken by this function
 588  */
 589 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 590 {
 591         struct extent_state *state;
 592         struct rb_node *node;
 593
 594         spin_lock_irq(&tree->lock);
 595 again:
 596         while (1) {
 597                 /*
 598                  * this search will find all the extents that end after
 599                  * our range starts
 600                  */
 601                 node = tree_search(tree, start);
 602                 if (!node)
 603                         break;
 604
 605                 state = rb_entry(node, struct extent_state, rb_node);
 606
 607                 if (state->start > end)
 608                         goto out;
 609
 610                 if (state->state & bits) {
 611                         start = state->start;
 612                         atomic_inc(&state->refs);
 613                         wait_on_state(tree, state);
 614                         free_extent_state(state);
 615                         goto again;
 616                 }
 617                 start = state->end + 1;
 618
 619                 if (start > end)
 620                         break;
 621
 622                 if (need_resched()) {
 623                         spin_unlock_irq(&tree->lock);
 624                         cond_resched();
 625                         spin_lock_irq(&tree->lock);
 626                 }
 627         }
 628 out:
 629         spin_unlock_irq(&tree->lock);
 630         return 0;
 631 }
 632 EXPORT_SYMBOL(wait_extent_bit);
 633
 634 static void set_state_bits(struct extent_io_tree *tree,
 635                            struct extent_state *state,
 636                            int bits)
 637 {
 638         if ((bits & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 639                 u64 range = state->end - state->start + 1;
 640                 tree->dirty_bytes += range;
 641         }
 642         set_state_cb(tree, state, bits);
 643         state->state |= bits;
 644 }
 645
 646 /*
 647  * set some bits on a range in the tree.  This may require allocations
 648  * or sleeping, so the gfp mask is used to indicate what is allowed.
 649  *
 650  * If 'exclusive' == 1, this will fail with -EEXIST if some part of the
 651  * range already has the desired bits set.  The start of the existing
 652  * range is returned in failed_start in this case.
 653  *
 654  * [start, end] is inclusive
 655  * This takes the tree lock.
 656  */
 657 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits,
 658                    int exclusive, u64 *failed_start, gfp_t mask)
 659 {
 660         struct extent_state *state;
 661         struct extent_state *prealloc = NULL;
 662         struct rb_node *node;
 663         unsigned long flags;
 664         int err = 0;
 665         int set;
 666         u64 last_start;
 667         u64 last_end;
 668 again:
 669         if (!prealloc && (mask & __GFP_WAIT)) {
 670                 prealloc = alloc_extent_state(mask);
 671                 if (!prealloc)
 672                         return -ENOMEM;
 673         }
 674
 675         spin_lock_irqsave(&tree->lock, flags);
 676         /*
 677          * this search will find all the extents that end after
 678          * our range starts.
 679          */
 680         node = tree_search(tree, start);
 681         if (!node) {
 682                 err = insert_state(tree, prealloc, start, end, bits);
 683                 prealloc = NULL;
 684                 BUG_ON(err == -EEXIST);
 685                 goto out;
 686         }
 687
 688         state = rb_entry(node, struct extent_state, rb_node);
 689         last_start = state->start;
 690         last_end = state->end;
 691
 692         /*
 693          * | ---- desired range ---- |
 694          * | state |
 695          *
 696          * Just lock what we found and keep going
 697          */
 698         if (state->start == start && state->end <= end) {
 699                 set = state->state & bits;
 700                 if (set && exclusive) {
 701                         *failed_start = state->start;
 702                         err = -EEXIST;
 703                         goto out;
 704                 }
 705                 set_state_bits(tree, state, bits);
 706                 start = state->end + 1;
 707                 merge_state(tree, state);
 708                 goto search_again;
 709         }
 710
 711         /*
 712          *     | ---- desired range ---- |
 713          * | state |
 714          *   or
 715          * | ------------- state -------------- |
 716          *
 717          * We need to split the extent we found, and may flip bits on
 718          * second half.
 719          *
 720          * If the extent we found extends past our
 721          * range, we just split and search again.  It'll get split
 722          * again the next time though.
 723          *
 724          * If the extent we found is inside our range, we set the
 725          * desired bit on it.
 726          */
 727         if (state->start < start) {
 728                 set = state->state & bits;
 729                 if (exclusive && set) {
 730                         *failed_start = start;
 731                         err = -EEXIST;
 732                         goto out;
 733                 }
 734                 err = split_state(tree, state, prealloc, start);
 735                 BUG_ON(err == -EEXIST);
 736                 prealloc = NULL;
 737                 if (err)
 738                         goto out;
 739                 if (state->end <= end) {
 740                         set_state_bits(tree, state, bits);
 741                         start = state->end + 1;
 742                         merge_state(tree, state);
 743                 } else {
 744                         start = state->start;
 745                 }
 746                 goto search_again;
 747         }
 748         /*
 749          * | ---- desired range ---- |
 750          *     | state | or               | state |
 751          *
 752          * There's a hole, we need to insert something in it and
 753          * ignore the extent we found.
 754          */
 755         if (state->start > start) {
 756                 u64 this_end;
 757                 if (end < last_start)
 758                         this_end = end;
 759                 else
 760                         this_end = last_start -1;
 761                 err = insert_state(tree, prealloc, start, this_end,
 762                                    bits);
 763                 prealloc = NULL;
 764                 BUG_ON(err == -EEXIST);
 765                 if (err)
 766                         goto out;
 767                 start = this_end + 1;
 768                 goto search_again;
 769         }
 770         /*
 771          * | ---- desired range ---- |
 772          *                        | state |
 773          * We need to split the extent, and set the bit
 774          * on the first half
 775          */
 776         if (state->start <= end && state->end > end) {
 777                 set = state->state & bits;
 778                 if (exclusive && set) {
 779                         *failed_start = start;
 780                         err = -EEXIST;
 781                         goto out;
 782                 }
 783                 err = split_state(tree, state, prealloc, end + 1);
 784                 BUG_ON(err == -EEXIST);
 785
 786                 set_state_bits(tree, prealloc, bits);
 787                 merge_state(tree, prealloc);
 788                 prealloc = NULL;
 789                 goto out;
 790         }
 791
 792         goto search_again;
 793
 794 out:
 795         spin_unlock_irqrestore(&tree->lock, flags);
 796         if (prealloc)
 797                 free_extent_state(prealloc);
 798
 799         return err;
 800
 801 search_again:
 802         if (start > end)
 803                 goto out;
 804         spin_unlock_irqrestore(&tree->lock, flags);
 805         if (mask & __GFP_WAIT)
 806                 cond_resched();
 807         goto again;
 808 }
 809 EXPORT_SYMBOL(set_extent_bit);
 810
 811 /* wrappers around set/clear extent bit */
 812 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 813                      gfp_t mask)
 814 {
 815         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 816                               mask);
 817 }
 818 EXPORT_SYMBOL(set_extent_dirty);
 819
 820 int set_extent_ordered(struct extent_io_tree *tree, u64 start, u64 end,
 821                        gfp_t mask)
 822 {
 823         return set_extent_bit(tree, start, end, EXTENT_ORDERED, 0, NULL, mask);
 824 }
 825 EXPORT_SYMBOL(set_extent_ordered);
 826
 827 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 828                     int bits, gfp_t mask)
 829 {
 830         return set_extent_bit(tree, start, end, bits, 0, NULL,
 831                               mask);
 832 }
 833 EXPORT_SYMBOL(set_extent_bits);
 834
 835 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 836                       int bits, gfp_t mask)
 837 {
 838         return clear_extent_bit(tree, start, end, bits, 0, 0, mask);
 839 }
 840 EXPORT_SYMBOL(clear_extent_bits);
 841
 842 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
 843                      gfp_t mask)
 844 {
 845         return set_extent_bit(tree, start, end,
 846                               EXTENT_DELALLOC | EXTENT_DIRTY,
 847                               0, NULL, mask);
 848 }
 849 EXPORT_SYMBOL(set_extent_delalloc);
 850
 851 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 852                        gfp_t mask)
 853 {
 854         return clear_extent_bit(tree, start, end,
 855                                 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, mask);
 856 }
 857 EXPORT_SYMBOL(clear_extent_dirty);
 858
 859 int clear_extent_ordered(struct extent_io_tree *tree, u64 start, u64 end,
 860                          gfp_t mask)
 861 {
 862         return clear_extent_bit(tree, start, end, EXTENT_ORDERED, 1, 0, mask);
 863 }
 864 EXPORT_SYMBOL(clear_extent_ordered);
 865
 866 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 867                      gfp_t mask)
 868 {
 869         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 870                               mask);
 871 }
 872 EXPORT_SYMBOL(set_extent_new);
 873
 874 int clear_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 875                        gfp_t mask)
 876 {
 877         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0, mask);
 878 }
 879 EXPORT_SYMBOL(clear_extent_new);
 880
 881 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 882                         gfp_t mask)
 883 {
 884         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, NULL,
 885                               mask);
 886 }
 887 EXPORT_SYMBOL(set_extent_uptodate);
 888
 889 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 890                           gfp_t mask)
 891 {
 892         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0, mask);
 893 }
 894 EXPORT_SYMBOL(clear_extent_uptodate);
 895
 896 int set_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end,
 897                          gfp_t mask)
 898 {
 899         return set_extent_bit(tree, start, end, EXTENT_WRITEBACK,
 900                               0, NULL, mask);
 901 }
 902 EXPORT_SYMBOL(set_extent_writeback);
 903
 904 int clear_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end,
 905                            gfp_t mask)
 906 {
 907         return clear_extent_bit(tree, start, end, EXTENT_WRITEBACK, 1, 0, mask);
 908 }
 909 EXPORT_SYMBOL(clear_extent_writeback);
 910
 911 int wait_on_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 912 {
 913         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 914 }
 915 EXPORT_SYMBOL(wait_on_extent_writeback);
 916
 917 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
 918 {
 919         int err;
 920         u64 failed_start;
 921         while (1) {
 922                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED, 1,
 923                                      &failed_start, mask);
 924                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
 925                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
 926                         start = failed_start;
 927                 } else {
 928                         break;
 929                 }
 930                 WARN_ON(start > end);
 931         }
 932         return err;
 933 }
 934 EXPORT_SYMBOL(lock_extent);
 935
 936 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end,
 937                   gfp_t mask)
 938 {
 939         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, mask);
 940 }
 941 EXPORT_SYMBOL(unlock_extent);
 942
 943 /*
 944  * helper function to set pages and extents in the tree dirty
 945  */
 946 int set_range_dirty(struct extent_io_tree *tree, u64 start, u64 end)
 947 {
 948         unsigned long index = start >> PAGE_CACHE_SHIFT;
 949         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 950         struct page *page;
 951
 952         while (index <= end_index) {
 953                 page = find_get_page(tree->mapping, index);
 954                 BUG_ON(!page);
 955                 __set_page_dirty_nobuffers(page);
 956                 page_cache_release(page);
 957                 index++;
 958         }
 959         set_extent_dirty(tree, start, end, GFP_NOFS);
 960         return 0;
 961 }
 962 EXPORT_SYMBOL(set_range_dirty);
 963
 964 /*
 965  * helper function to set both pages and extents in the tree writeback
 966  */
 967 int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 968 {
 969         unsigned long index = start >> PAGE_CACHE_SHIFT;
 970         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 971         struct page *page;
 972
 973         while (index <= end_index) {
 974                 page = find_get_page(tree->mapping, index);
 975                 BUG_ON(!page);
 976                 set_page_writeback(page);
 977                 page_cache_release(page);
 978                 index++;
 979         }
 980         set_extent_writeback(tree, start, end, GFP_NOFS);
 981         return 0;
 982 }
 983 EXPORT_SYMBOL(set_range_writeback);
 984
 985 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
 986                           u64 *start_ret, u64 *end_ret, int bits)
 987 {
 988         struct rb_node *node;
 989         struct extent_state *state;
 990         int ret = 1;
 991
 992         spin_lock_irq(&tree->lock);
 993         /*
 994          * this search will find all the extents that end after
 995          * our range starts.
 996          */
 997         node = tree_search(tree, start);
 998         if (!node) {
 999                 goto out;
1000         }
1001
1002         while(1) {
1003                 state = rb_entry(node, struct extent_state, rb_node);
1004                 if (state->end >= start && (state->state & bits)) {
1005                         *start_ret = state->start;
1006                         *end_ret = state->end;
1007                         ret = 0;
1008                         break;
1009                 }
1010                 node = rb_next(node);
1011                 if (!node)
1012                         break;
1013         }
1014 out:
1015         spin_unlock_irq(&tree->lock);
1016         return ret;
1017 }
1018 EXPORT_SYMBOL(find_first_extent_bit);
1019
1020 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1021                                                  u64 start, int bits)
1022 {
1023         struct rb_node *node;
1024         struct extent_state *state;
1025
1026         /*
1027          * this search will find all the extents that end after
1028          * our range starts.
1029          */
1030         node = tree_search(tree, start);
1031         if (!node) {
1032                 goto out;
1033         }
1034
1035         while(1) {
1036                 state = rb_entry(node, struct extent_state, rb_node);
1037                 if (state->end >= start && (state->state & bits)) {
1038                         return state;
1039                 }
1040                 node = rb_next(node);
1041                 if (!node)
1042                         break;
1043         }
1044 out:
1045         return NULL;
1046 }
1047 EXPORT_SYMBOL(find_first_extent_bit_state);
1048
1049 u64 find_lock_delalloc_range(struct extent_io_tree *tree,
1050                              u64 *start, u64 *end, u64 max_bytes)
1051 {
1052         struct rb_node *node;
1053         struct extent_state *state;
1054         u64 cur_start = *start;
1055         u64 found = 0;
1056         u64 total_bytes = 0;
1057
1058         spin_lock_irq(&tree->lock);
1059         /*
1060          * this search will find all the extents that end after
1061          * our range starts.
1062          */
1063 search_again:
1064         node = tree_search(tree, cur_start);
1065         if (!node) {
1066                 if (!found)
1067                         *end = (u64)-1;
1068                 goto out;
1069         }
1070
1071         while(1) {
1072                 state = rb_entry(node, struct extent_state, rb_node);
1073                 if (found && state->start != cur_start) {
1074                         goto out;
1075                 }
1076                 if (!(state->state & EXTENT_DELALLOC)) {
1077                         if (!found)
1078                                 *end = state->end;
1079                         goto out;
1080                 }
1081                 if (!found) {
1082                         struct extent_state *prev_state;
1083                         struct rb_node *prev_node = node;
1084                         while(1) {
1085                                 prev_node = rb_prev(prev_node);
1086                                 if (!prev_node)
1087                                         break;
1088                                 prev_state = rb_entry(prev_node,
1089                                                       struct extent_state,
1090                                                       rb_node);
1091                                 if (!(prev_state->state & EXTENT_DELALLOC))
1092                                         break;
1093                                 state = prev_state;
1094                                 node = prev_node;
1095                         }
1096                 }
1097                 if (state->state & EXTENT_LOCKED) {
1098                         DEFINE_WAIT(wait);
1099                         atomic_inc(&state->refs);
1100                         prepare_to_wait(&state->wq, &wait,
1101                                         TASK_UNINTERRUPTIBLE);
1102                         spin_unlock_irq(&tree->lock);
1103                         schedule();
1104                         spin_lock_irq(&tree->lock);
1105                         finish_wait(&state->wq, &wait);
1106                         free_extent_state(state);
1107                         goto search_again;
1108                 }
1109                 set_state_cb(tree, state, EXTENT_LOCKED);
1110                 state->state |= EXTENT_LOCKED;
1111                 if (!found)
1112                         *start = state->start;
1113                 found++;
1114                 *end = state->end;
1115                 cur_start = state->end + 1;
1116                 node = rb_next(node);
1117                 if (!node)
1118                         break;
1119                 total_bytes += state->end - state->start + 1;
1120                 if (total_bytes >= max_bytes)
1121                         break;
1122         }
1123 out:
1124         spin_unlock_irq(&tree->lock);
1125         return found;
1126 }
1127
1128 u64 count_range_bits(struct extent_io_tree *tree,
1129                      u64 *start, u64 search_end, u64 max_bytes,
1130                      unsigned long bits)
1131 {
1132         struct rb_node *node;
1133         struct extent_state *state;
1134         u64 cur_start = *start;
1135         u64 total_bytes = 0;
1136         int found = 0;
1137
1138         if (search_end <= cur_start) {
1139                 printk("search_end %Lu start %Lu\n", search_end, cur_start);
1140                 WARN_ON(1);
1141                 return 0;
1142         }
1143
1144         spin_lock_irq(&tree->lock);
1145         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1146                 total_bytes = tree->dirty_bytes;
1147                 goto out;
1148         }
1149         /*
1150          * this search will find all the extents that end after
1151          * our range starts.
1152          */
1153         node = tree_search(tree, cur_start);
1154         if (!node) {
1155                 goto out;
1156         }
1157
1158         while(1) {
1159                 state = rb_entry(node, struct extent_state, rb_node);
1160                 if (state->start > search_end)
1161                         break;
1162                 if (state->end >= cur_start && (state->state & bits)) {
1163                         total_bytes += min(search_end, state->end) + 1 -
1164                                        max(cur_start, state->start);
1165                         if (total_bytes >= max_bytes)
1166                                 break;
1167                         if (!found) {
1168                                 *start = state->start;
1169                                 found = 1;
1170                         }
1171                 }
1172                 node = rb_next(node);
1173                 if (!node)
1174                         break;
1175         }
1176 out:
1177         spin_unlock_irq(&tree->lock);
1178         return total_bytes;
1179 }
1180 /*
1181  * helper function to lock both pages and extents in the tree.
1182  * pages must be locked first.
1183  */
1184 int lock_range(struct extent_io_tree *tree, u64 start, u64 end)
1185 {
1186         unsigned long index = start >> PAGE_CACHE_SHIFT;
1187         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1188         struct page *page;
1189         int err;
1190
1191         while (index <= end_index) {
1192                 page = grab_cache_page(tree->mapping, index);
1193                 if (!page) {
1194                         err = -ENOMEM;
1195                         goto failed;
1196                 }
1197                 if (IS_ERR(page)) {
1198                         err = PTR_ERR(page);
1199                         goto failed;
1200                 }
1201                 index++;
1202         }
1203         lock_extent(tree, start, end, GFP_NOFS);
1204         return 0;
1205
1206 failed:
1207         /*
1208          * we failed above in getting the page at 'index', so we undo here
1209          * up to but not including the page at 'index'
1210          */
1211         end_index = index;
1212         index = start >> PAGE_CACHE_SHIFT;
1213         while (index < end_index) {
1214                 page = find_get_page(tree->mapping, index);
1215                 unlock_page(page);
1216                 page_cache_release(page);
1217                 index++;
1218         }
1219         return err;
1220 }
1221 EXPORT_SYMBOL(lock_range);
1222
1223 /*
1224  * helper function to unlock both pages and extents in the tree.
1225  */
1226 int unlock_range(struct extent_io_tree *tree, u64 start, u64 end)
1227 {
1228         unsigned long index = start >> PAGE_CACHE_SHIFT;
1229         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1230         struct page *page;
1231
1232         while (index <= end_index) {
1233                 page = find_get_page(tree->mapping, index);
1234                 unlock_page(page);
1235                 page_cache_release(page);
1236                 index++;
1237         }
1238         unlock_extent(tree, start, end, GFP_NOFS);
1239         return 0;
1240 }
1241 EXPORT_SYMBOL(unlock_range);
1242
1243 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1244 {
1245         struct rb_node *node;
1246         struct extent_state *state;
1247         int ret = 0;
1248
1249         spin_lock_irq(&tree->lock);
1250         /*
1251          * this search will find all the extents that end after
1252          * our range starts.
1253          */
1254         node = tree_search(tree, start);
1255         if (!node) {
1256                 ret = -ENOENT;
1257                 goto out;
1258         }
1259         state = rb_entry(node, struct extent_state, rb_node);
1260         if (state->start != start) {
1261                 ret = -ENOENT;
1262                 goto out;
1263         }
1264         state->private = private;
1265 out:
1266         spin_unlock_irq(&tree->lock);
1267         return ret;
1268 }
1269
1270 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1271 {
1272         struct rb_node *node;
1273         struct extent_state *state;
1274         int ret = 0;
1275
1276         spin_lock_irq(&tree->lock);
1277         /*
1278          * this search will find all the extents that end after
1279          * our range starts.
1280          */
1281         node = tree_search(tree, start);
1282         if (!node) {
1283                 ret = -ENOENT;
1284                 goto out;
1285         }
1286         state = rb_entry(node, struct extent_state, rb_node);
1287         if (state->start != start) {
1288                 ret = -ENOENT;
1289                 goto out;
1290         }
1291         *private = state->private;
1292 out:
1293         spin_unlock_irq(&tree->lock);
1294         return ret;
1295 }
1296
1297 /*
1298  * searches a range in the state tree for a given mask.
1299  * If 'filled' == 1, this returns 1 only if every extent in the tree
1300  * has the bits set.  Otherwise, 1 is returned if any bit in the
1301  * range is found set.
1302  */
1303 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1304                    int bits, int filled)
1305 {
1306         struct extent_state *state = NULL;
1307         struct rb_node *node;
1308         int bitset = 0;
1309         unsigned long flags;
1310
1311         spin_lock_irqsave(&tree->lock, flags);
1312         node = tree_search(tree, start);
1313         while (node && start <= end) {
1314                 state = rb_entry(node, struct extent_state, rb_node);
1315
1316                 if (filled && state->start > start) {
1317                         bitset = 0;
1318                         break;
1319                 }
1320
1321                 if (state->start > end)
1322                         break;
1323
1324                 if (state->state & bits) {
1325                         bitset = 1;
1326                         if (!filled)
1327                                 break;
1328                 } else if (filled) {
1329                         bitset = 0;
1330                         break;
1331                 }
1332                 start = state->end + 1;
1333                 if (start > end)
1334                         break;
1335                 node = rb_next(node);
1336                 if (!node) {
1337                         if (filled)
1338                                 bitset = 0;
1339                         break;
1340                 }
1341         }
1342         spin_unlock_irqrestore(&tree->lock, flags);
1343         return bitset;
1344 }
1345 EXPORT_SYMBOL(test_range_bit);
1346
1347 /*
1348  * helper function to set a given page up to date if all the
1349  * extents in the tree for that page are up to date
1350  */
1351 static int check_page_uptodate(struct extent_io_tree *tree,
1352                                struct page *page)
1353 {
1354         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1355         u64 end = start + PAGE_CACHE_SIZE - 1;
1356         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1))
1357                 SetPageUptodate(page);
1358         return 0;
1359 }
1360
1361 /*
1362  * helper function to unlock a page if all the extents in the tree
1363  * for that page are unlocked
1364  */
1365 static int check_page_locked(struct extent_io_tree *tree,
1366                              struct page *page)
1367 {
1368         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1369         u64 end = start + PAGE_CACHE_SIZE - 1;
1370         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0))
1371                 unlock_page(page);
1372         return 0;
1373 }
1374
1375 /*
1376  * helper function to end page writeback if all the extents
1377  * in the tree for that page are done with writeback
1378  */
1379 static int check_page_writeback(struct extent_io_tree *tree,
1380                              struct page *page)
1381 {
1382         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1383         u64 end = start + PAGE_CACHE_SIZE - 1;
1384         if (!test_range_bit(tree, start, end, EXTENT_WRITEBACK, 0))
1385                 end_page_writeback(page);
1386         return 0;
1387 }
1388
1389 /* lots and lots of room for performance fixes in the end_bio funcs */
1390
1391 /*
1392  * after a writepage IO is done, we need to:
1393  * clear the uptodate bits on error
1394  * clear the writeback bits in the extent tree for this IO
1395  * end_page_writeback if the page has no more pending IO
1396  *
1397  * Scheduling is not allowed, so the extent state tree is expected
1398  * to have one and only one object corresponding to this IO.
1399  */
1400 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1401 static void end_bio_extent_writepage(struct bio *bio, int err)
1402 #else
1403 static int end_bio_extent_writepage(struct bio *bio,
1404                                    unsigned int bytes_done, int err)
1405 #endif
1406 {
1407         int uptodate = err == 0;
1408         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1409         struct extent_io_tree *tree;
1410         u64 start;
1411         u64 end;
1412         int whole_page;
1413         int ret;
1414
1415 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1416         if (bio->bi_size)
1417                 return 1;
1418 #endif
1419         do {
1420                 struct page *page = bvec->bv_page;
1421                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1422
1423                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1424                          bvec->bv_offset;
1425                 end = start + bvec->bv_len - 1;
1426
1427                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1428                         whole_page = 1;
1429                 else
1430                         whole_page = 0;
1431
1432                 if (--bvec >= bio->bi_io_vec)
1433                         prefetchw(&bvec->bv_page->flags);
1434                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1435                         ret = tree->ops->writepage_end_io_hook(page, start,
1436                                                        end, NULL, uptodate);
1437                         if (ret)
1438                                 uptodate = 0;
1439                 }
1440
1441                 if (!uptodate && tree->ops &&
1442                     tree->ops->writepage_io_failed_hook) {
1443                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1444                                                          start, end, NULL);
1445                         if (ret == 0) {
1446                                 uptodate = (err == 0);
1447                                 continue;
1448                         }
1449                 }
1450
1451                 if (!uptodate) {
1452                         clear_extent_uptodate(tree, start, end, GFP_ATOMIC);
1453                         ClearPageUptodate(page);
1454                         SetPageError(page);
1455                 }
1456
1457                 clear_extent_writeback(tree, start, end, GFP_ATOMIC);
1458
1459                 if (whole_page)
1460                         end_page_writeback(page);
1461                 else
1462                         check_page_writeback(tree, page);
1463         } while (bvec >= bio->bi_io_vec);
1464         bio_put(bio);
1465 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1466         return 0;
1467 #endif
1468 }
1469
1470 /*
1471  * after a readpage IO is done, we need to:
1472  * clear the uptodate bits on error
1473  * set the uptodate bits if things worked
1474  * set the page up to date if all extents in the tree are uptodate
1475  * clear the lock bit in the extent tree
1476  * unlock the page if there are no other extents locked for it
1477  *
1478  * Scheduling is not allowed, so the extent state tree is expected
1479  * to have one and only one object corresponding to this IO.
1480  */
1481 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1482 static void end_bio_extent_readpage(struct bio *bio, int err)
1483 #else
1484 static int end_bio_extent_readpage(struct bio *bio,
1485                                    unsigned int bytes_done, int err)
1486 #endif
1487 {
1488         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1489         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1490         struct extent_io_tree *tree;
1491         u64 start;
1492         u64 end;
1493         int whole_page;
1494         int ret;
1495
1496 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1497         if (bio->bi_size)
1498                 return 1;
1499 #endif
1500
1501         do {
1502                 struct page *page = bvec->bv_page;
1503                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1504
1505                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1506                         bvec->bv_offset;
1507                 end = start + bvec->bv_len - 1;
1508
1509                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1510                         whole_page = 1;
1511                 else
1512                         whole_page = 0;
1513
1514                 if (--bvec >= bio->bi_io_vec)
1515                         prefetchw(&bvec->bv_page->flags);
1516
1517                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1518                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1519                                                               NULL);
1520                         if (ret)
1521                                 uptodate = 0;
1522                 }
1523                 if (!uptodate && tree->ops &&
1524                     tree->ops->readpage_io_failed_hook) {
1525                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1526                                                          start, end, NULL);
1527                         if (ret == 0) {
1528                                 uptodate =
1529                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1530                                 continue;
1531                         }
1532                 }
1533
1534                 if (uptodate)
1535                         set_extent_uptodate(tree, start, end,
1536                                             GFP_ATOMIC);
1537                 unlock_extent(tree, start, end, GFP_ATOMIC);
1538
1539                 if (whole_page) {
1540                         if (uptodate) {
1541                                 SetPageUptodate(page);
1542                         } else {
1543                                 ClearPageUptodate(page);
1544                                 SetPageError(page);
1545                         }
1546                         unlock_page(page);
1547                 } else {
1548                         if (uptodate) {
1549                                 check_page_uptodate(tree, page);
1550                         } else {
1551                                 ClearPageUptodate(page);
1552                                 SetPageError(page);
1553                         }
1554                         check_page_locked(tree, page);
1555                 }
1556         } while (bvec >= bio->bi_io_vec);
1557
1558         bio_put(bio);
1559 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1560         return 0;
1561 #endif
1562 }
1563
1564 /*
1565  * IO done from prepare_write is pretty simple, we just unlock
1566  * the structs in the extent tree when done, and set the uptodate bits
1567  * as appropriate.
1568  */
1569 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1570 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1571 #else
1572 static int end_bio_extent_preparewrite(struct bio *bio,
1573                                        unsigned int bytes_done, int err)
1574 #endif
1575 {
1576         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1577         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1578         struct extent_io_tree *tree;
1579         u64 start;
1580         u64 end;
1581
1582 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1583         if (bio->bi_size)
1584                 return 1;
1585 #endif
1586
1587         do {
1588                 struct page *page = bvec->bv_page;
1589                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1590
1591                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1592                         bvec->bv_offset;
1593                 end = start + bvec->bv_len - 1;
1594
1595                 if (--bvec >= bio->bi_io_vec)
1596                         prefetchw(&bvec->bv_page->flags);
1597
1598                 if (uptodate) {
1599                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1600                 } else {
1601                         ClearPageUptodate(page);
1602                         SetPageError(page);
1603                 }
1604
1605                 unlock_extent(tree, start, end, GFP_ATOMIC);
1606
1607         } while (bvec >= bio->bi_io_vec);
1608
1609         bio_put(bio);
1610 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1611         return 0;
1612 #endif
1613 }
1614
1615 static struct bio *
1616 extent_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1617                  gfp_t gfp_flags)
1618 {
1619         struct bio *bio;
1620
1621         bio = bio_alloc(gfp_flags, nr_vecs);
1622
1623         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1624                 while (!bio && (nr_vecs /= 2))
1625                         bio = bio_alloc(gfp_flags, nr_vecs);
1626         }
1627
1628         if (bio) {
1629                 bio->bi_size = 0;
1630                 bio->bi_bdev = bdev;
1631                 bio->bi_sector = first_sector;
1632         }
1633         return bio;
1634 }
1635
1636 static int submit_one_bio(int rw, struct bio *bio, int mirror_num)
1637 {
1638         int ret = 0;
1639         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1640         struct page *page = bvec->bv_page;
1641         struct extent_io_tree *tree = bio->bi_private;
1642         struct rb_node *node;
1643         struct extent_state *state;
1644         u64 start;
1645         u64 end;
1646
1647         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1648         end = start + bvec->bv_len - 1;
1649
1650         spin_lock_irq(&tree->lock);
1651         node = __etree_search(tree, start, NULL, NULL);
1652         BUG_ON(!node);
1653         state = rb_entry(node, struct extent_state, rb_node);
1654         while(state->end < end) {
1655                 node = rb_next(node);
1656                 state = rb_entry(node, struct extent_state, rb_node);
1657         }
1658         BUG_ON(state->end != end);
1659         spin_unlock_irq(&tree->lock);
1660
1661         bio->bi_private = NULL;
1662
1663         bio_get(bio);
1664
1665         if (tree->ops && tree->ops->submit_bio_hook)
1666                 tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1667                                            mirror_num);
1668         else
1669                 submit_bio(rw, bio);
1670         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1671                 ret = -EOPNOTSUPP;
1672         bio_put(bio);
1673         return ret;
1674 }
1675
1676 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1677                               struct page *page, sector_t sector,
1678                               size_t size, unsigned long offset,
1679                               struct block_device *bdev,
1680                               struct bio **bio_ret,
1681                               unsigned long max_pages,
1682                               bio_end_io_t end_io_func,
1683                               int mirror_num)
1684 {
1685         int ret = 0;
1686         struct bio *bio;
1687         int nr;
1688
1689         if (bio_ret && *bio_ret) {
1690                 bio = *bio_ret;
1691                 if (bio->bi_sector + (bio->bi_size >> 9) != sector ||
1692                     (tree->ops && tree->ops->merge_bio_hook &&
1693                      tree->ops->merge_bio_hook(page, offset, size, bio)) ||
1694                     bio_add_page(bio, page, size, offset) < size) {
1695                         ret = submit_one_bio(rw, bio, mirror_num);
1696                         bio = NULL;
1697                 } else {
1698                         return 0;
1699                 }
1700         }
1701         nr = bio_get_nr_vecs(bdev);
1702         bio = extent_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1703         if (!bio) {
1704                 printk("failed to allocate bio nr %d\n", nr);
1705         }
1706
1707
1708         bio_add_page(bio, page, size, offset);
1709         bio->bi_end_io = end_io_func;
1710         bio->bi_private = tree;
1711
1712         if (bio_ret) {
1713                 *bio_ret = bio;
1714         } else {
1715                 ret = submit_one_bio(rw, bio, mirror_num);
1716         }
1717
1718         return ret;
1719 }
1720
1721 void set_page_extent_mapped(struct page *page)
1722 {
1723         if (!PagePrivate(page)) {
1724                 SetPagePrivate(page);
1725                 page_cache_get(page);
1726                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1727         }
1728 }
1729
1730 void set_page_extent_head(struct page *page, unsigned long len)
1731 {
1732         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1733 }
1734
1735 /*
1736  * basic readpage implementation.  Locked extent state structs are inserted
1737  * into the tree that are removed when the IO is done (by the end_io
1738  * handlers)
1739  */
1740 static int __extent_read_full_page(struct extent_io_tree *tree,
1741                                    struct page *page,
1742                                    get_extent_t *get_extent,
1743                                    struct bio **bio, int mirror_num)
1744 {
1745         struct inode *inode = page->mapping->host;
1746         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1747         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1748         u64 end;
1749         u64 cur = start;
1750         u64 extent_offset;
1751         u64 last_byte = i_size_read(inode);
1752         u64 block_start;
1753         u64 cur_end;
1754         sector_t sector;
1755         struct extent_map *em;
1756         struct block_device *bdev;
1757         int ret;
1758         int nr = 0;
1759         size_t page_offset = 0;
1760         size_t iosize;
1761         size_t blocksize = inode->i_sb->s_blocksize;
1762
1763         set_page_extent_mapped(page);
1764
1765         end = page_end;
1766         lock_extent(tree, start, end, GFP_NOFS);
1767
1768         while (cur <= end) {
1769                 if (cur >= last_byte) {
1770                         char *userpage;
1771                         iosize = PAGE_CACHE_SIZE - page_offset;
1772                         userpage = kmap_atomic(page, KM_USER0);
1773                         memset(userpage + page_offset, 0, iosize);
1774                         flush_dcache_page(page);
1775                         kunmap_atomic(userpage, KM_USER0);
1776                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1777                                             GFP_NOFS);
1778                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1779                         break;
1780                 }
1781                 em = get_extent(inode, page, page_offset, cur,
1782                                 end - cur + 1, 0);
1783                 if (IS_ERR(em) || !em) {
1784                         SetPageError(page);
1785                         unlock_extent(tree, cur, end, GFP_NOFS);
1786                         break;
1787                 }
1788                 extent_offset = cur - em->start;
1789                 if (extent_map_end(em) <= cur) {
1790 printk("bad mapping em [%Lu %Lu] cur %Lu\n", em->start, extent_map_end(em), cur);
1791                 }
1792                 BUG_ON(extent_map_end(em) <= cur);
1793                 if (end < cur) {
1794 printk("2bad mapping end %Lu cur %Lu\n", end, cur);
1795                 }
1796                 BUG_ON(end < cur);
1797
1798                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1799                 cur_end = min(extent_map_end(em) - 1, end);
1800                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1801                 sector = (em->block_start + extent_offset) >> 9;
1802                 bdev = em->bdev;
1803                 block_start = em->block_start;
1804                 free_extent_map(em);
1805                 em = NULL;
1806
1807                 /* we've found a hole, just zero and go on */
1808                 if (block_start == EXTENT_MAP_HOLE) {
1809                         char *userpage;
1810                         userpage = kmap_atomic(page, KM_USER0);
1811                         memset(userpage + page_offset, 0, iosize);
1812                         flush_dcache_page(page);
1813                         kunmap_atomic(userpage, KM_USER0);
1814
1815                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1816                                             GFP_NOFS);
1817                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1818                         cur = cur + iosize;
1819                         page_offset += iosize;
1820                         continue;
1821                 }
1822                 /* the get_extent function already copied into the page */
1823                 if (test_range_bit(tree, cur, cur_end, EXTENT_UPTODATE, 1)) {
1824                         check_page_uptodate(tree, page);
1825                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1826                         cur = cur + iosize;
1827                         page_offset += iosize;
1828                         continue;
1829                 }
1830                 /* we have an inline extent but it didn't get marked up
1831                  * to date.  Error out
1832                  */
1833                 if (block_start == EXTENT_MAP_INLINE) {
1834                         SetPageError(page);
1835                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1836                         cur = cur + iosize;
1837                         page_offset += iosize;
1838                         continue;
1839                 }
1840
1841                 ret = 0;
1842                 if (tree->ops && tree->ops->readpage_io_hook) {
1843                         ret = tree->ops->readpage_io_hook(page, cur,
1844                                                           cur + iosize - 1);
1845                 }
1846                 if (!ret) {
1847                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
1848                         pnr -= page->index;
1849                         ret = submit_extent_page(READ, tree, page,
1850                                          sector, iosize, page_offset,
1851                                          bdev, bio, pnr,
1852                                          end_bio_extent_readpage, mirror_num);
1853                         nr++;
1854                 }
1855                 if (ret)
1856                         SetPageError(page);
1857                 cur = cur + iosize;
1858                 page_offset += iosize;
1859         }
1860         if (!nr) {
1861                 if (!PageError(page))
1862                         SetPageUptodate(page);
1863                 unlock_page(page);
1864         }
1865         return 0;
1866 }
1867
1868 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
1869                             get_extent_t *get_extent)
1870 {
1871         struct bio *bio = NULL;
1872         int ret;
1873
1874         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0);
1875         if (bio)
1876                 submit_one_bio(READ, bio, 0);
1877         return ret;
1878 }
1879 EXPORT_SYMBOL(extent_read_full_page);
1880
1881 /*
1882  * the writepage semantics are similar to regular writepage.  extent
1883  * records are inserted to lock ranges in the tree, and as dirty areas
1884  * are found, they are marked writeback.  Then the lock bits are removed
1885  * and the end_io handler clears the writeback ranges
1886  */
1887 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
1888                               void *data)
1889 {
1890         struct inode *inode = page->mapping->host;
1891         struct extent_page_data *epd = data;
1892         struct extent_io_tree *tree = epd->tree;
1893         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1894         u64 delalloc_start;
1895         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1896         u64 end;
1897         u64 cur = start;
1898         u64 extent_offset;
1899         u64 last_byte = i_size_read(inode);
1900         u64 block_start;
1901         u64 iosize;
1902         u64 unlock_start;
1903         sector_t sector;
1904         struct extent_map *em;
1905         struct block_device *bdev;
1906         int ret;
1907         int nr = 0;
1908         size_t pg_offset = 0;
1909         size_t blocksize;
1910         loff_t i_size = i_size_read(inode);
1911         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
1912         u64 nr_delalloc;
1913         u64 delalloc_end;
1914
1915         WARN_ON(!PageLocked(page));
1916         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
1917         if (page->index > end_index ||
1918            (page->index == end_index && !pg_offset)) {
1919                 page->mapping->a_ops->invalidatepage(page, 0);
1920                 unlock_page(page);
1921                 return 0;
1922         }
1923
1924         if (page->index == end_index) {
1925                 char *userpage;
1926
1927                 userpage = kmap_atomic(page, KM_USER0);
1928                 memset(userpage + pg_offset, 0,
1929                        PAGE_CACHE_SIZE - pg_offset);
1930                 kunmap_atomic(userpage, KM_USER0);
1931                 flush_dcache_page(page);
1932         }
1933         pg_offset = 0;
1934
1935         set_page_extent_mapped(page);
1936
1937         delalloc_start = start;
1938         delalloc_end = 0;
1939         while(delalloc_end < page_end) {
1940                 nr_delalloc = find_lock_delalloc_range(tree, &delalloc_start,
1941                                                        &delalloc_end,
1942                                                        128 * 1024 * 1024);
1943                 if (nr_delalloc == 0) {
1944                         delalloc_start = delalloc_end + 1;
1945                         continue;
1946                 }
1947                 tree->ops->fill_delalloc(inode, delalloc_start,
1948                                          delalloc_end);
1949                 clear_extent_bit(tree, delalloc_start,
1950                                  delalloc_end,
1951                                  EXTENT_LOCKED | EXTENT_DELALLOC,
1952                                  1, 0, GFP_NOFS);
1953                 delalloc_start = delalloc_end + 1;
1954         }
1955         lock_extent(tree, start, page_end, GFP_NOFS);
1956         unlock_start = start;
1957
1958         if (tree->ops && tree->ops->writepage_start_hook) {
1959                 ret = tree->ops->writepage_start_hook(page, start, page_end);
1960                 if (ret == -EAGAIN) {
1961                         unlock_extent(tree, start, page_end, GFP_NOFS);
1962                         redirty_page_for_writepage(wbc, page);
1963                         unlock_page(page);
1964                         return 0;
1965                 }
1966         }
1967
1968         end = page_end;
1969         if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0)) {
1970                 printk("found delalloc bits after lock_extent\n");
1971         }
1972
1973         if (last_byte <= start) {
1974                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
1975                 unlock_extent(tree, start, page_end, GFP_NOFS);
1976                 if (tree->ops && tree->ops->writepage_end_io_hook)
1977                         tree->ops->writepage_end_io_hook(page, start,
1978                                                          page_end, NULL, 1);
1979                 unlock_start = page_end + 1;
1980                 goto done;
1981         }
1982
1983         set_extent_uptodate(tree, start, page_end, GFP_NOFS);
1984         blocksize = inode->i_sb->s_blocksize;
1985
1986         while (cur <= end) {
1987                 if (cur >= last_byte) {
1988                         clear_extent_dirty(tree, cur, page_end, GFP_NOFS);
1989                         unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
1990                         if (tree->ops && tree->ops->writepage_end_io_hook)
1991                                 tree->ops->writepage_end_io_hook(page, cur,
1992                                                          page_end, NULL, 1);
1993                         unlock_start = page_end + 1;
1994                         break;
1995                 }
1996                 em = epd->get_extent(inode, page, pg_offset, cur,
1997                                      end - cur + 1, 1);
1998                 if (IS_ERR(em) || !em) {
1999                         SetPageError(page);
2000                         break;
2001                 }
2002
2003                 extent_offset = cur - em->start;
2004                 BUG_ON(extent_map_end(em) <= cur);
2005                 BUG_ON(end < cur);
2006                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2007                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2008                 sector = (em->block_start + extent_offset) >> 9;
2009                 bdev = em->bdev;
2010                 block_start = em->block_start;
2011                 free_extent_map(em);
2012                 em = NULL;
2013
2014                 if (block_start == EXTENT_MAP_HOLE ||
2015                     block_start == EXTENT_MAP_INLINE) {
2016                         clear_extent_dirty(tree, cur,
2017                                            cur + iosize - 1, GFP_NOFS);
2018
2019                         unlock_extent(tree, unlock_start, cur + iosize -1,
2020                                       GFP_NOFS);
2021
2022                         if (tree->ops && tree->ops->writepage_end_io_hook)
2023                                 tree->ops->writepage_end_io_hook(page, cur,
2024                                                          cur + iosize - 1,
2025                                                          NULL, 1);
2026                         cur = cur + iosize;
2027                         pg_offset += iosize;
2028                         unlock_start = cur;
2029                         continue;
2030                 }
2031
2032                 /* leave this out until we have a page_mkwrite call */
2033                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2034                                    EXTENT_DIRTY, 0)) {
2035                         cur = cur + iosize;
2036                         pg_offset += iosize;
2037                         continue;
2038                 }
2039                 clear_extent_dirty(tree, cur, cur + iosize - 1, GFP_NOFS);
2040                 if (tree->ops && tree->ops->writepage_io_hook) {
2041                         ret = tree->ops->writepage_io_hook(page, cur,
2042                                                 cur + iosize - 1);
2043                 } else {
2044                         ret = 0;
2045                 }
2046                 if (ret) {
2047                         SetPageError(page);
2048                 } else {
2049                         unsigned long max_nr = end_index + 1;
2050
2051                         set_range_writeback(tree, cur, cur + iosize - 1);
2052                         if (!PageWriteback(page)) {
2053                                 printk("warning page %lu not writeback, "
2054                                        "cur %llu end %llu\n", page->index,
2055                                        (unsigned long long)cur,
2056                                        (unsigned long long)end);
2057                         }
2058
2059                         ret = submit_extent_page(WRITE, tree, page, sector,
2060                                                  iosize, pg_offset, bdev,
2061                                                  &epd->bio, max_nr,
2062                                                  end_bio_extent_writepage, 0);
2063                         if (ret)
2064                                 SetPageError(page);
2065                 }
2066                 cur = cur + iosize;
2067                 pg_offset += iosize;
2068                 nr++;
2069         }
2070 done:
2071         if (nr == 0) {
2072                 /* make sure the mapping tag for page dirty gets cleared */
2073                 set_page_writeback(page);
2074                 end_page_writeback(page);
2075         }
2076         if (unlock_start <= page_end)
2077                 unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
2078         unlock_page(page);
2079         return 0;
2080 }
2081
2082 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,22)
2083 /* Taken directly from 2.6.23 with a mod for a lockpage hook */
2084 typedef int (*writepage_t)(struct page *page, struct writeback_control *wbc,
2085                                 void *data);
2086 #endif
2087
2088 /**
2089  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2090  * @mapping: address space structure to write
2091  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2092  * @writepage: function called for each page
2093  * @data: data passed to writepage function
2094  *
2095  * If a page is already under I/O, write_cache_pages() skips it, even
2096  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2097  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2098  * and msync() need to guarantee that all the data which was dirty at the time
2099  * the call was made get new I/O started against them.  If wbc->sync_mode is
2100  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2101  * existing IO to complete.
2102  */
2103 int extent_write_cache_pages(struct extent_io_tree *tree,
2104                              struct address_space *mapping,
2105                              struct writeback_control *wbc,
2106                              writepage_t writepage, void *data)
2107 {
2108         struct backing_dev_info *bdi = mapping->backing_dev_info;
2109         int ret = 0;
2110         int done = 0;
2111         struct pagevec pvec;
2112         int nr_pages;
2113         pgoff_t index;
2114         pgoff_t end;            /* Inclusive */
2115         int scanned = 0;
2116         int range_whole = 0;
2117
2118         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2119                 wbc->encountered_congestion = 1;
2120                 return 0;
2121         }
2122
2123         pagevec_init(&pvec, 0);
2124         if (wbc->range_cyclic) {
2125                 index = mapping->writeback_index; /* Start from prev offset */
2126                 end = -1;
2127         } else {
2128                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2129                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2130                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2131                         range_whole = 1;
2132                 scanned = 1;
2133         }
2134 retry:
2135         while (!done && (index <= end) &&
2136                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2137                                               PAGECACHE_TAG_DIRTY,
2138                                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2139                 unsigned i;
2140
2141                 scanned = 1;
2142                 for (i = 0; i < nr_pages; i++) {
2143                         struct page *page = pvec.pages[i];
2144
2145                         /*
2146                          * At this point we hold neither mapping->tree_lock nor
2147                          * lock on the page itself: the page may be truncated or
2148                          * invalidated (changing page->mapping to NULL), or even
2149                          * swizzled back from swapper_space to tmpfs file
2150                          * mapping
2151                          */
2152                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2153                                 tree->ops->write_cache_pages_lock_hook(page);
2154                         else
2155                                 lock_page(page);
2156
2157                         if (unlikely(page->mapping != mapping)) {
2158                                 unlock_page(page);
2159                                 continue;
2160                         }
2161
2162                         if (!wbc->range_cyclic && page->index > end) {
2163                                 done = 1;
2164                                 unlock_page(page);
2165                                 continue;
2166                         }
2167
2168                         if (wbc->sync_mode != WB_SYNC_NONE)
2169                                 wait_on_page_writeback(page);
2170
2171                         if (PageWriteback(page) ||
2172                             !clear_page_dirty_for_io(page)) {
2173                                 unlock_page(page);
2174                                 continue;
2175                         }
2176
2177                         ret = (*writepage)(page, wbc, data);
2178
2179                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2180                                 unlock_page(page);
2181                                 ret = 0;
2182                         }
2183                         if (ret || (--(wbc->nr_to_write) <= 0))
2184                                 done = 1;
2185                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2186                                 wbc->encountered_congestion = 1;
2187                                 done = 1;
2188                         }
2189                 }
2190                 pagevec_release(&pvec);
2191                 cond_resched();
2192         }
2193         if (!scanned && !done) {
2194                 /*
2195                  * We hit the last page and there is more work to be done: wrap
2196                  * back to the start of the file
2197                  */
2198                 scanned = 1;
2199                 index = 0;
2200                 goto retry;
2201         }
2202         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2203                 mapping->writeback_index = index;
2204
2205         if (wbc->range_cont)
2206                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2207         return ret;
2208 }
2209 EXPORT_SYMBOL(extent_write_cache_pages);
2210
2211 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2212                           get_extent_t *get_extent,
2213                           struct writeback_control *wbc)
2214 {
2215         int ret;
2216         struct address_space *mapping = page->mapping;
2217         struct extent_page_data epd = {
2218                 .bio = NULL,
2219                 .tree = tree,
2220                 .get_extent = get_extent,
2221         };
2222         struct writeback_control wbc_writepages = {
2223                 .bdi            = wbc->bdi,
2224                 .sync_mode      = WB_SYNC_NONE,
2225                 .older_than_this = NULL,
2226                 .nr_to_write    = 64,
2227                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2228                 .range_end      = (loff_t)-1,
2229         };
2230
2231
2232         ret = __extent_writepage(page, wbc, &epd);
2233
2234         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2235                                  __extent_writepage, &epd);
2236         if (epd.bio) {
2237                 submit_one_bio(WRITE, epd.bio, 0);
2238         }
2239         return ret;
2240 }
2241 EXPORT_SYMBOL(extent_write_full_page);
2242
2243
2244 int extent_writepages(struct extent_io_tree *tree,
2245                       struct address_space *mapping,
2246                       get_extent_t *get_extent,
2247                       struct writeback_control *wbc)
2248 {
2249         int ret = 0;
2250         struct extent_page_data epd = {
2251                 .bio = NULL,
2252                 .tree = tree,
2253                 .get_extent = get_extent,
2254         };
2255
2256         ret = extent_write_cache_pages(tree, mapping, wbc,
2257                                        __extent_writepage, &epd);
2258         if (epd.bio) {
2259                 submit_one_bio(WRITE, epd.bio, 0);
2260         }
2261         return ret;
2262 }
2263 EXPORT_SYMBOL(extent_writepages);
2264
2265 int extent_readpages(struct extent_io_tree *tree,
2266                      struct address_space *mapping,
2267                      struct list_head *pages, unsigned nr_pages,
2268                      get_extent_t get_extent)
2269 {
2270         struct bio *bio = NULL;
2271         unsigned page_idx;
2272         struct pagevec pvec;
2273
2274         pagevec_init(&pvec, 0);
2275         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2276                 struct page *page = list_entry(pages->prev, struct page, lru);
2277
2278                 prefetchw(&page->flags);
2279                 list_del(&page->lru);
2280                 /*
2281                  * what we want to do here is call add_to_page_cache_lru,
2282                  * but that isn't exported, so we reproduce it here
2283                  */
2284                 if (!add_to_page_cache(page, mapping,
2285                                         page->index, GFP_KERNEL)) {
2286
2287                         /* open coding of lru_cache_add, also not exported */
2288                         page_cache_get(page);
2289                         if (!pagevec_add(&pvec, page))
2290                                 __pagevec_lru_add(&pvec);
2291                         __extent_read_full_page(tree, page, get_extent,
2292                                                 &bio, 0);
2293                 }
2294                 page_cache_release(page);
2295         }
2296         if (pagevec_count(&pvec))
2297                 __pagevec_lru_add(&pvec);
2298         BUG_ON(!list_empty(pages));
2299         if (bio)
2300                 submit_one_bio(READ, bio, 0);
2301         return 0;
2302 }
2303 EXPORT_SYMBOL(extent_readpages);
2304
2305 /*
2306  * basic invalidatepage code, this waits on any locked or writeback
2307  * ranges corresponding to the page, and then deletes any extent state
2308  * records from the tree
2309  */
2310 int extent_invalidatepage(struct extent_io_tree *tree,
2311                           struct page *page, unsigned long offset)
2312 {
2313         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2314         u64 end = start + PAGE_CACHE_SIZE - 1;
2315         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2316
2317         start += (offset + blocksize -1) & ~(blocksize - 1);
2318         if (start > end)
2319                 return 0;
2320
2321         lock_extent(tree, start, end, GFP_NOFS);
2322         wait_on_extent_writeback(tree, start, end);
2323         clear_extent_bit(tree, start, end,
2324                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC,
2325                          1, 1, GFP_NOFS);
2326         return 0;
2327 }
2328 EXPORT_SYMBOL(extent_invalidatepage);
2329
2330 /*
2331  * simple commit_write call, set_range_dirty is used to mark both
2332  * the pages and the extent records as dirty
2333  */
2334 int extent_commit_write(struct extent_io_tree *tree,
2335                         struct inode *inode, struct page *page,
2336                         unsigned from, unsigned to)
2337 {
2338         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2339
2340         set_page_extent_mapped(page);
2341         set_page_dirty(page);
2342
2343         if (pos > inode->i_size) {
2344                 i_size_write(inode, pos);
2345                 mark_inode_dirty(inode);
2346         }
2347         return 0;
2348 }
2349 EXPORT_SYMBOL(extent_commit_write);
2350
2351 int extent_prepare_write(struct extent_io_tree *tree,
2352                          struct inode *inode, struct page *page,
2353                          unsigned from, unsigned to, get_extent_t *get_extent)
2354 {
2355         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2356         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2357         u64 block_start;
2358         u64 orig_block_start;
2359         u64 block_end;
2360         u64 cur_end;
2361         struct extent_map *em;
2362         unsigned blocksize = 1 << inode->i_blkbits;
2363         size_t page_offset = 0;
2364         size_t block_off_start;
2365         size_t block_off_end;
2366         int err = 0;
2367         int iocount = 0;
2368         int ret = 0;
2369         int isnew;
2370
2371         set_page_extent_mapped(page);
2372
2373         block_start = (page_start + from) & ~((u64)blocksize - 1);
2374         block_end = (page_start + to - 1) | (blocksize - 1);
2375         orig_block_start = block_start;
2376
2377         lock_extent(tree, page_start, page_end, GFP_NOFS);
2378         while(block_start <= block_end) {
2379                 em = get_extent(inode, page, page_offset, block_start,
2380                                 block_end - block_start + 1, 1);
2381                 if (IS_ERR(em) || !em) {
2382                         goto err;
2383                 }
2384                 cur_end = min(block_end, extent_map_end(em) - 1);
2385                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2386                 block_off_end = block_off_start + blocksize;
2387                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2388
2389                 if (!PageUptodate(page) && isnew &&
2390                     (block_off_end > to || block_off_start < from)) {
2391                         void *kaddr;
2392
2393                         kaddr = kmap_atomic(page, KM_USER0);
2394                         if (block_off_end > to)
2395                                 memset(kaddr + to, 0, block_off_end - to);
2396                         if (block_off_start < from)
2397                                 memset(kaddr + block_off_start, 0,
2398                                        from - block_off_start);
2399                         flush_dcache_page(page);
2400                         kunmap_atomic(kaddr, KM_USER0);
2401                 }
2402                 if ((em->block_start != EXTENT_MAP_HOLE &&
2403                      em->block_start != EXTENT_MAP_INLINE) &&
2404                     !isnew && !PageUptodate(page) &&
2405                     (block_off_end > to || block_off_start < from) &&
2406                     !test_range_bit(tree, block_start, cur_end,
2407                                     EXTENT_UPTODATE, 1)) {
2408                         u64 sector;
2409                         u64 extent_offset = block_start - em->start;
2410                         size_t iosize;
2411                         sector = (em->block_start + extent_offset) >> 9;
2412                         iosize = (cur_end - block_start + blocksize) &
2413                                 ~((u64)blocksize - 1);
2414                         /*
2415                          * we've already got the extent locked, but we
2416                          * need to split the state such that our end_bio
2417                          * handler can clear the lock.
2418                          */
2419                         set_extent_bit(tree, block_start,
2420                                        block_start + iosize - 1,
2421                                        EXTENT_LOCKED, 0, NULL, GFP_NOFS);
2422                         ret = submit_extent_page(READ, tree, page,
2423                                          sector, iosize, page_offset, em->bdev,
2424                                          NULL, 1,
2425                                          end_bio_extent_preparewrite, 0);
2426                         iocount++;
2427                         block_start = block_start + iosize;
2428                 } else {
2429                         set_extent_uptodate(tree, block_start, cur_end,
2430                                             GFP_NOFS);
2431                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2432                         block_start = cur_end + 1;
2433                 }
2434                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2435                 free_extent_map(em);
2436         }
2437         if (iocount) {
2438                 wait_extent_bit(tree, orig_block_start,
2439                                 block_end, EXTENT_LOCKED);
2440         }
2441         check_page_uptodate(tree, page);
2442 err:
2443         /* FIXME, zero out newly allocated blocks on error */
2444         return err;
2445 }
2446 EXPORT_SYMBOL(extent_prepare_write);
2447
2448 /*
2449  * a helper for releasepage, this tests for areas of the page that
2450  * are locked or under IO and drops the related state bits if it is safe
2451  * to drop the page.
2452  */
2453 int try_release_extent_state(struct extent_map_tree *map,
2454                              struct extent_io_tree *tree, struct page *page,
2455                              gfp_t mask)
2456 {
2457         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2458         u64 end = start + PAGE_CACHE_SIZE - 1;
2459         int ret = 1;
2460
2461         if (test_range_bit(tree, start, end,
2462                            EXTENT_IOBITS | EXTENT_ORDERED, 0))
2463                 ret = 0;
2464         else {
2465                 if ((mask & GFP_NOFS) == GFP_NOFS)
2466                         mask = GFP_NOFS;
2467                 clear_extent_bit(tree, start, end, EXTENT_UPTODATE,
2468                                  1, 1, mask);
2469         }
2470         return ret;
2471 }
2472 EXPORT_SYMBOL(try_release_extent_state);
2473
2474 /*
2475  * a helper for releasepage.  As long as there are no locked extents
2476  * in the range corresponding to the page, both state records and extent
2477  * map records are removed
2478  */
2479 int try_release_extent_mapping(struct extent_map_tree *map,
2480                                struct extent_io_tree *tree, struct page *page,
2481                                gfp_t mask)
2482 {
2483         struct extent_map *em;
2484         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2485         u64 end = start + PAGE_CACHE_SIZE - 1;
2486
2487         if ((mask & __GFP_WAIT) &&
2488             page->mapping->host->i_size > 16 * 1024 * 1024) {
2489                 u64 len;
2490                 while (start <= end) {
2491                         len = end - start + 1;
2492                         spin_lock(&map->lock);
2493                         em = lookup_extent_mapping(map, start, len);
2494                         if (!em || IS_ERR(em)) {
2495                                 spin_unlock(&map->lock);
2496                                 break;
2497                         }
2498                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2499                             em->start != start) {
2500                                 spin_unlock(&map->lock);
2501                                 free_extent_map(em);
2502                                 break;
2503                         }
2504                         if (!test_range_bit(tree, em->start,
2505                                             extent_map_end(em) - 1,
2506                                             EXTENT_LOCKED, 0)) {
2507                                 remove_extent_mapping(map, em);
2508                                 /* once for the rb tree */
2509                                 free_extent_map(em);
2510                         }
2511                         start = extent_map_end(em);
2512                         spin_unlock(&map->lock);
2513
2514                         /* once for us */
2515                         free_extent_map(em);
2516                 }
2517         }
2518         return try_release_extent_state(map, tree, page, mask);
2519 }
2520 EXPORT_SYMBOL(try_release_extent_mapping);
2521
2522 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2523                 get_extent_t *get_extent)
2524 {
2525         struct inode *inode = mapping->host;
2526         u64 start = iblock << inode->i_blkbits;
2527         sector_t sector = 0;
2528         struct extent_map *em;
2529
2530         em = get_extent(inode, NULL, 0, start, (1 << inode->i_blkbits), 0);
2531         if (!em || IS_ERR(em))
2532                 return 0;
2533
2534         if (em->block_start == EXTENT_MAP_INLINE ||
2535             em->block_start == EXTENT_MAP_HOLE)
2536                 goto out;
2537
2538         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2539 out:
2540         free_extent_map(em);
2541         return sector;
2542 }
2543
2544 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2545                                               unsigned long i)
2546 {
2547         struct page *p;
2548         struct address_space *mapping;
2549
2550         if (i == 0)
2551                 return eb->first_page;
2552         i += eb->start >> PAGE_CACHE_SHIFT;
2553         mapping = eb->first_page->mapping;
2554         if (!mapping)
2555                 return NULL;
2556
2557         /*
2558          * extent_buffer_page is only called after pinning the page
2559          * by increasing the reference count.  So we know the page must
2560          * be in the radix tree.
2561          */
2562 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,26)
2563         rcu_read_lock();
2564 #else
2565         read_lock_irq(&mapping->tree_lock);
2566 #endif
2567         p = radix_tree_lookup(&mapping->page_tree, i);
2568
2569 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,26)
2570         rcu_read_unlock();
2571 #else
2572         read_unlock_irq(&mapping->tree_lock);
2573 #endif
2574         return p;
2575 }
2576
2577 static inline unsigned long num_extent_pages(u64 start, u64 len)
2578 {
2579         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
2580                 (start >> PAGE_CACHE_SHIFT);
2581 }
2582
2583 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
2584                                                    u64 start,
2585                                                    unsigned long len,
2586                                                    gfp_t mask)
2587 {
2588         struct extent_buffer *eb = NULL;
2589 #ifdef LEAK_DEBUG
2590         unsigned long flags;
2591 #endif
2592
2593         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
2594         eb->start = start;
2595         eb->len = len;
2596         mutex_init(&eb->mutex);
2597 #ifdef LEAK_DEBUG
2598         spin_lock_irqsave(&leak_lock, flags);
2599         list_add(&eb->leak_list, &buffers);
2600         spin_unlock_irqrestore(&leak_lock, flags);
2601 #endif
2602         atomic_set(&eb->refs, 1);
2603
2604         return eb;
2605 }
2606
2607 static void __free_extent_buffer(struct extent_buffer *eb)
2608 {
2609 #ifdef LEAK_DEBUG
2610         unsigned long flags;
2611         spin_lock_irqsave(&leak_lock, flags);
2612         list_del(&eb->leak_list);
2613         spin_unlock_irqrestore(&leak_lock, flags);
2614 #endif
2615         kmem_cache_free(extent_buffer_cache, eb);
2616 }
2617
2618 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
2619                                           u64 start, unsigned long len,
2620                                           struct page *page0,
2621                                           gfp_t mask)
2622 {
2623         unsigned long num_pages = num_extent_pages(start, len);
2624         unsigned long i;
2625         unsigned long index = start >> PAGE_CACHE_SHIFT;
2626         struct extent_buffer *eb;
2627         struct extent_buffer *exists = NULL;
2628         struct page *p;
2629         struct address_space *mapping = tree->mapping;
2630         int uptodate = 1;
2631
2632         spin_lock(&tree->buffer_lock);
2633         eb = buffer_search(tree, start);
2634         if (eb) {
2635                 atomic_inc(&eb->refs);
2636                 spin_unlock(&tree->buffer_lock);
2637                 mark_page_accessed(eb->first_page);
2638                 return eb;
2639         }
2640         spin_unlock(&tree->buffer_lock);
2641
2642         eb = __alloc_extent_buffer(tree, start, len, mask);
2643         if (!eb)
2644                 return NULL;
2645
2646         if (page0) {
2647                 eb->first_page = page0;
2648                 i = 1;
2649                 index++;
2650                 page_cache_get(page0);
2651                 mark_page_accessed(page0);
2652                 set_page_extent_mapped(page0);
2653                 set_page_extent_head(page0, len);
2654                 uptodate = PageUptodate(page0);
2655         } else {
2656                 i = 0;
2657         }
2658         for (; i < num_pages; i++, index++) {
2659                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
2660                 if (!p) {
2661                         WARN_ON(1);
2662                         goto free_eb;
2663                 }
2664                 set_page_extent_mapped(p);
2665                 mark_page_accessed(p);
2666                 if (i == 0) {
2667                         eb->first_page = p;
2668                         set_page_extent_head(p, len);
2669                 } else {
2670                         set_page_private(p, EXTENT_PAGE_PRIVATE);
2671                 }
2672                 if (!PageUptodate(p))
2673                         uptodate = 0;
2674                 unlock_page(p);
2675         }
2676         if (uptodate)
2677                 eb->flags |= EXTENT_UPTODATE;
2678         eb->flags |= EXTENT_BUFFER_FILLED;
2679
2680         spin_lock(&tree->buffer_lock);
2681         exists = buffer_tree_insert(tree, start, &eb->rb_node);
2682         if (exists) {
2683                 /* add one reference for the caller */
2684                 atomic_inc(&exists->refs);
2685                 spin_unlock(&tree->buffer_lock);
2686                 goto free_eb;
2687         }
2688         spin_unlock(&tree->buffer_lock);
2689
2690         /* add one reference for the tree */
2691         atomic_inc(&eb->refs);
2692         return eb;
2693
2694 free_eb:
2695         if (!atomic_dec_and_test(&eb->refs))
2696                 return exists;
2697         for (index = 1; index < i; index++)
2698                 page_cache_release(extent_buffer_page(eb, index));
2699         page_cache_release(extent_buffer_page(eb, 0));
2700         __free_extent_buffer(eb);
2701         return exists;
2702 }
2703 EXPORT_SYMBOL(alloc_extent_buffer);
2704
2705 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
2706                                          u64 start, unsigned long len,
2707                                           gfp_t mask)
2708 {
2709         struct extent_buffer *eb;
2710
2711         spin_lock(&tree->buffer_lock);
2712         eb = buffer_search(tree, start);
2713         if (eb)
2714                 atomic_inc(&eb->refs);
2715         spin_unlock(&tree->buffer_lock);
2716
2717         if (eb)
2718                 mark_page_accessed(eb->first_page);
2719
2720         return eb;
2721 }
2722 EXPORT_SYMBOL(find_extent_buffer);
2723
2724 void free_extent_buffer(struct extent_buffer *eb)
2725 {
2726         if (!eb)
2727                 return;
2728
2729         if (!atomic_dec_and_test(&eb->refs))
2730                 return;
2731
2732         WARN_ON(1);
2733 }
2734 EXPORT_SYMBOL(free_extent_buffer);
2735
2736 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
2737                               struct extent_buffer *eb)
2738 {
2739         int set;
2740         unsigned long i;
2741         unsigned long num_pages;
2742         struct page *page;
2743
2744         u64 start = eb->start;
2745         u64 end = start + eb->len - 1;
2746
2747         set = clear_extent_dirty(tree, start, end, GFP_NOFS);
2748         num_pages = num_extent_pages(eb->start, eb->len);
2749
2750         for (i = 0; i < num_pages; i++) {
2751                 page = extent_buffer_page(eb, i);
2752                 lock_page(page);
2753                 if (i == 0)
2754                         set_page_extent_head(page, eb->len);
2755                 else
2756                         set_page_private(page, EXTENT_PAGE_PRIVATE);
2757
2758                 /*
2759                  * if we're on the last page or the first page and the
2760                  * block isn't aligned on a page boundary, do extra checks
2761                  * to make sure we don't clean page that is partially dirty
2762                  */
2763                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
2764                     ((i == num_pages - 1) &&
2765                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
2766                         start = (u64)page->index << PAGE_CACHE_SHIFT;
2767                         end  = start + PAGE_CACHE_SIZE - 1;
2768                         if (test_range_bit(tree, start, end,
2769                                            EXTENT_DIRTY, 0)) {
2770                                 unlock_page(page);
2771                                 continue;
2772                         }
2773                 }
2774                 clear_page_dirty_for_io(page);
2775 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,26)
2776                 spin_lock_irq(&page->mapping->tree_lock);
2777 #else
2778                 read_lock_irq(&page->mapping->tree_lock);
2779 #endif
2780                 if (!PageDirty(page)) {
2781                         radix_tree_tag_clear(&page->mapping->page_tree,
2782                                                 page_index(page),
2783                                                 PAGECACHE_TAG_DIRTY);
2784                 }
2785 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,26)
2786                 spin_unlock_irq(&page->mapping->tree_lock);
2787 #else
2788                 read_unlock_irq(&page->mapping->tree_lock);
2789 #endif
2790                 unlock_page(page);
2791         }
2792         return 0;
2793 }
2794 EXPORT_SYMBOL(clear_extent_buffer_dirty);
2795
2796 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
2797                                     struct extent_buffer *eb)
2798 {
2799         return wait_on_extent_writeback(tree, eb->start,
2800                                         eb->start + eb->len - 1);
2801 }
2802 EXPORT_SYMBOL(wait_on_extent_buffer_writeback);
2803
2804 int set_extent_buffer_dirty(struct extent_io_tree *tree,
2805                              struct extent_buffer *eb)
2806 {
2807         unsigned long i;
2808         unsigned long num_pages;
2809
2810         num_pages = num_extent_pages(eb->start, eb->len);
2811         for (i = 0; i < num_pages; i++) {
2812                 struct page *page = extent_buffer_page(eb, i);
2813                 /* writepage may need to do something special for the
2814                  * first page, we have to make sure page->private is
2815                  * properly set.  releasepage may drop page->private
2816                  * on us if the page isn't already dirty.
2817                  */
2818                 lock_page(page);
2819                 if (i == 0) {
2820                         set_page_extent_head(page, eb->len);
2821                 } else if (PagePrivate(page) &&
2822                            page->private != EXTENT_PAGE_PRIVATE) {
2823                         set_page_extent_mapped(page);
2824                 }
2825                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
2826                 set_extent_dirty(tree, page_offset(page),
2827                                  page_offset(page) + PAGE_CACHE_SIZE -1,
2828                                  GFP_NOFS);
2829                 unlock_page(page);
2830         }
2831         return 0;
2832 }
2833 EXPORT_SYMBOL(set_extent_buffer_dirty);
2834
2835 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
2836                                 struct extent_buffer *eb)
2837 {
2838         unsigned long i;
2839         struct page *page;
2840         unsigned long num_pages;
2841
2842         num_pages = num_extent_pages(eb->start, eb->len);
2843         eb->flags &= ~EXTENT_UPTODATE;
2844
2845         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
2846                               GFP_NOFS);
2847         for (i = 0; i < num_pages; i++) {
2848                 page = extent_buffer_page(eb, i);
2849                 if (page)
2850                         ClearPageUptodate(page);
2851         }
2852         return 0;
2853 }
2854
2855 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
2856                                 struct extent_buffer *eb)
2857 {
2858         unsigned long i;
2859         struct page *page;
2860         unsigned long num_pages;
2861
2862         num_pages = num_extent_pages(eb->start, eb->len);
2863
2864         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
2865                             GFP_NOFS);
2866         for (i = 0; i < num_pages; i++) {
2867                 page = extent_buffer_page(eb, i);
2868                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
2869                     ((i == num_pages - 1) &&
2870                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
2871                         check_page_uptodate(tree, page);
2872                         continue;
2873                 }
2874                 SetPageUptodate(page);
2875         }
2876         return 0;
2877 }
2878 EXPORT_SYMBOL(set_extent_buffer_uptodate);
2879
2880 int extent_range_uptodate(struct extent_io_tree *tree,
2881                           u64 start, u64 end)
2882 {
2883         struct page *page;
2884         int ret;
2885         int pg_uptodate = 1;
2886         int uptodate;
2887         unsigned long index;
2888
2889         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1);
2890         if (ret)
2891                 return 1;
2892         while(start <= end) {
2893                 index = start >> PAGE_CACHE_SHIFT;
2894                 page = find_get_page(tree->mapping, index);
2895                 uptodate = PageUptodate(page);
2896                 page_cache_release(page);
2897                 if (!uptodate) {
2898                         pg_uptodate = 0;
2899                         break;
2900                 }
2901                 start += PAGE_CACHE_SIZE;
2902         }
2903         return pg_uptodate;
2904 }
2905
2906 int extent_buffer_uptodate(struct extent_io_tree *tree,
2907                            struct extent_buffer *eb)
2908 {
2909         int ret = 0;
2910         unsigned long num_pages;
2911         unsigned long i;
2912         struct page *page;
2913         int pg_uptodate = 1;
2914
2915         if (eb->flags & EXTENT_UPTODATE)
2916                 return 1;
2917
2918         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
2919                            EXTENT_UPTODATE, 1);
2920         if (ret)
2921                 return ret;
2922
2923         num_pages = num_extent_pages(eb->start, eb->len);
2924         for (i = 0; i < num_pages; i++) {
2925                 page = extent_buffer_page(eb, i);
2926                 if (!PageUptodate(page)) {
2927                         pg_uptodate = 0;
2928                         break;
2929                 }
2930         }
2931         return pg_uptodate;
2932 }
2933 EXPORT_SYMBOL(extent_buffer_uptodate);
2934
2935 int read_extent_buffer_pages(struct extent_io_tree *tree,
2936                              struct extent_buffer *eb,
2937                              u64 start, int wait,
2938                              get_extent_t *get_extent, int mirror_num)
2939 {
2940         unsigned long i;
2941         unsigned long start_i;
2942         struct page *page;
2943         int err;
2944         int ret = 0;
2945         int locked_pages = 0;
2946         int all_uptodate = 1;
2947         int inc_all_pages = 0;
2948         unsigned long num_pages;
2949         struct bio *bio = NULL;
2950
2951         if (eb->flags & EXTENT_UPTODATE)
2952                 return 0;
2953
2954         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
2955                            EXTENT_UPTODATE, 1)) {
2956                 return 0;
2957         }
2958
2959         if (start) {
2960                 WARN_ON(start < eb->start);
2961                 start_i = (start >> PAGE_CACHE_SHIFT) -
2962                         (eb->start >> PAGE_CACHE_SHIFT);
2963         } else {
2964                 start_i = 0;
2965         }
2966
2967         num_pages = num_extent_pages(eb->start, eb->len);
2968         for (i = start_i; i < num_pages; i++) {
2969                 page = extent_buffer_page(eb, i);
2970                 if (!wait) {
2971                         if (!trylock_page(page))
2972                                 goto unlock_exit;
2973                 } else {
2974                         lock_page(page);
2975                 }
2976                 locked_pages++;
2977                 if (!PageUptodate(page)) {
2978                         all_uptodate = 0;
2979                 }
2980         }
2981         if (all_uptodate) {
2982                 if (start_i == 0)
2983                         eb->flags |= EXTENT_UPTODATE;
2984                 if (ret) {
2985                         printk("all up to date but ret is %d\n", ret);
2986                 }
2987                 goto unlock_exit;
2988         }
2989
2990         for (i = start_i; i < num_pages; i++) {
2991                 page = extent_buffer_page(eb, i);
2992                 if (inc_all_pages)
2993                         page_cache_get(page);
2994                 if (!PageUptodate(page)) {
2995                         if (start_i == 0)
2996                                 inc_all_pages = 1;
2997                         ClearPageError(page);
2998                         err = __extent_read_full_page(tree, page,
2999                                                       get_extent, &bio,
3000                                                       mirror_num);
3001                         if (err) {
3002                                 ret = err;
3003                                 printk("err %d from __extent_read_full_page\n", ret);
3004                         }
3005                 } else {
3006                         unlock_page(page);
3007                 }
3008         }
3009
3010         if (bio)
3011                 submit_one_bio(READ, bio, mirror_num);
3012
3013         if (ret || !wait) {
3014                 if (ret)
3015                         printk("ret %d wait %d returning\n", ret, wait);
3016                 return ret;
3017         }
3018         for (i = start_i; i < num_pages; i++) {
3019                 page = extent_buffer_page(eb, i);
3020                 wait_on_page_locked(page);
3021                 if (!PageUptodate(page)) {
3022                         printk("page not uptodate after wait_on_page_locked\n");
3023                         ret = -EIO;
3024                 }
3025         }
3026         if (!ret)
3027                 eb->flags |= EXTENT_UPTODATE;
3028         return ret;
3029
3030 unlock_exit:
3031         i = start_i;
3032         while(locked_pages > 0) {
3033                 page = extent_buffer_page(eb, i);
3034                 i++;
3035                 unlock_page(page);
3036                 locked_pages--;
3037         }
3038         return ret;
3039 }
3040 EXPORT_SYMBOL(read_extent_buffer_pages);
3041
3042 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3043                         unsigned long start,
3044                         unsigned long len)
3045 {
3046         size_t cur;
3047         size_t offset;
3048         struct page *page;
3049         char *kaddr;
3050         char *dst = (char *)dstv;
3051         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3052         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3053
3054         WARN_ON(start > eb->len);
3055         WARN_ON(start + len > eb->start + eb->len);
3056
3057         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3058
3059         while(len > 0) {
3060                 page = extent_buffer_page(eb, i);
3061
3062                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3063                 kaddr = kmap_atomic(page, KM_USER1);
3064                 memcpy(dst, kaddr + offset, cur);
3065                 kunmap_atomic(kaddr, KM_USER1);
3066
3067                 dst += cur;
3068                 len -= cur;
3069                 offset = 0;
3070                 i++;
3071         }
3072 }
3073 EXPORT_SYMBOL(read_extent_buffer);
3074
3075 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3076                                unsigned long min_len, char **token, char **map,
3077                                unsigned long *map_start,
3078                                unsigned long *map_len, int km)
3079 {
3080         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3081         char *kaddr;
3082         struct page *p;
3083         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3084         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3085         unsigned long end_i = (start_offset + start + min_len - 1) >>
3086                 PAGE_CACHE_SHIFT;
3087
3088         if (i != end_i)
3089                 return -EINVAL;
3090
3091         if (i == 0) {
3092                 offset = start_offset;
3093                 *map_start = 0;
3094         } else {
3095                 offset = 0;
3096                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3097         }
3098         if (start + min_len > eb->len) {
3099 printk("bad mapping eb start %Lu len %lu, wanted %lu %lu\n", eb->start, eb->len, start, min_len);
3100                 WARN_ON(1);
3101         }
3102
3103         p = extent_buffer_page(eb, i);
3104         kaddr = kmap_atomic(p, km);
3105         *token = kaddr;
3106         *map = kaddr + offset;
3107         *map_len = PAGE_CACHE_SIZE - offset;
3108         return 0;
3109 }
3110 EXPORT_SYMBOL(map_private_extent_buffer);
3111
3112 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3113                       unsigned long min_len,
3114                       char **token, char **map,
3115                       unsigned long *map_start,
3116                       unsigned long *map_len, int km)
3117 {
3118         int err;
3119         int save = 0;
3120         if (eb->map_token) {
3121                 unmap_extent_buffer(eb, eb->map_token, km);
3122                 eb->map_token = NULL;
3123                 save = 1;
3124         }
3125         err = map_private_extent_buffer(eb, start, min_len, token, map,
3126                                        map_start, map_len, km);
3127         if (!err && save) {
3128                 eb->map_token = *token;
3129                 eb->kaddr = *map;
3130                 eb->map_start = *map_start;
3131                 eb->map_len = *map_len;
3132         }
3133         return err;
3134 }
3135 EXPORT_SYMBOL(map_extent_buffer);
3136
3137 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3138 {
3139         kunmap_atomic(token, km);
3140 }
3141 EXPORT_SYMBOL(unmap_extent_buffer);
3142
3143 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3144                           unsigned long start,
3145                           unsigned long len)
3146 {
3147         size_t cur;
3148         size_t offset;
3149         struct page *page;
3150         char *kaddr;
3151         char *ptr = (char *)ptrv;
3152         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3153         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3154         int ret = 0;
3155
3156         WARN_ON(start > eb->len);
3157         WARN_ON(start + len > eb->start + eb->len);
3158
3159         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3160
3161         while(len > 0) {
3162                 page = extent_buffer_page(eb, i);
3163
3164                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3165
3166                 kaddr = kmap_atomic(page, KM_USER0);
3167                 ret = memcmp(ptr, kaddr + offset, cur);
3168                 kunmap_atomic(kaddr, KM_USER0);
3169                 if (ret)
3170                         break;
3171
3172                 ptr += cur;
3173                 len -= cur;
3174                 offset = 0;
3175                 i++;
3176         }
3177         return ret;
3178 }
3179 EXPORT_SYMBOL(memcmp_extent_buffer);
3180
3181 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3182                          unsigned long start, unsigned long len)
3183 {
3184         size_t cur;
3185         size_t offset;
3186         struct page *page;
3187         char *kaddr;
3188         char *src = (char *)srcv;
3189         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3190         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3191
3192         WARN_ON(start > eb->len);
3193         WARN_ON(start + len > eb->start + eb->len);
3194
3195         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3196
3197         while(len > 0) {
3198                 page = extent_buffer_page(eb, i);
3199                 WARN_ON(!PageUptodate(page));
3200
3201                 cur = min(len, PAGE_CACHE_SIZE - offset);
3202                 kaddr = kmap_atomic(page, KM_USER1);
3203                 memcpy(kaddr + offset, src, cur);
3204                 kunmap_atomic(kaddr, KM_USER1);
3205
3206                 src += cur;
3207                 len -= cur;
3208                 offset = 0;
3209                 i++;
3210         }
3211 }
3212 EXPORT_SYMBOL(write_extent_buffer);
3213
3214 void memset_extent_buffer(struct extent_buffer *eb, char c,
3215                           unsigned long start, unsigned long len)
3216 {
3217         size_t cur;
3218         size_t offset;
3219         struct page *page;
3220         char *kaddr;
3221         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3222         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3223
3224         WARN_ON(start > eb->len);
3225         WARN_ON(start + len > eb->start + eb->len);
3226
3227         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3228
3229         while(len > 0) {
3230                 page = extent_buffer_page(eb, i);
3231                 WARN_ON(!PageUptodate(page));
3232
3233                 cur = min(len, PAGE_CACHE_SIZE - offset);
3234                 kaddr = kmap_atomic(page, KM_USER0);
3235                 memset(kaddr + offset, c, cur);
3236                 kunmap_atomic(kaddr, KM_USER0);
3237
3238                 len -= cur;
3239                 offset = 0;
3240                 i++;
3241         }
3242 }
3243 EXPORT_SYMBOL(memset_extent_buffer);
3244
3245 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3246                         unsigned long dst_offset, unsigned long src_offset,
3247                         unsigned long len)
3248 {
3249         u64 dst_len = dst->len;
3250         size_t cur;
3251         size_t offset;
3252         struct page *page;
3253         char *kaddr;
3254         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3255         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3256
3257         WARN_ON(src->len != dst_len);
3258
3259         offset = (start_offset + dst_offset) &
3260                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3261
3262         while(len > 0) {
3263                 page = extent_buffer_page(dst, i);
3264                 WARN_ON(!PageUptodate(page));
3265
3266                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3267
3268                 kaddr = kmap_atomic(page, KM_USER0);
3269                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3270                 kunmap_atomic(kaddr, KM_USER0);
3271
3272                 src_offset += cur;
3273                 len -= cur;
3274                 offset = 0;
3275                 i++;
3276         }
3277 }
3278 EXPORT_SYMBOL(copy_extent_buffer);
3279
3280 static void move_pages(struct page *dst_page, struct page *src_page,
3281                        unsigned long dst_off, unsigned long src_off,
3282                        unsigned long len)
3283 {
3284         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3285         if (dst_page == src_page) {
3286                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3287         } else {
3288                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3289                 char *p = dst_kaddr + dst_off + len;
3290                 char *s = src_kaddr + src_off + len;
3291
3292                 while (len--)
3293                         *--p = *--s;
3294
3295                 kunmap_atomic(src_kaddr, KM_USER1);
3296         }
3297         kunmap_atomic(dst_kaddr, KM_USER0);
3298 }
3299
3300 static void copy_pages(struct page *dst_page, struct page *src_page,
3301                        unsigned long dst_off, unsigned long src_off,
3302                        unsigned long len)
3303 {
3304         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3305         char *src_kaddr;
3306
3307         if (dst_page != src_page)
3308                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3309         else
3310                 src_kaddr = dst_kaddr;
3311
3312         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3313         kunmap_atomic(dst_kaddr, KM_USER0);
3314         if (dst_page != src_page)
3315                 kunmap_atomic(src_kaddr, KM_USER1);
3316 }
3317
3318 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3319                            unsigned long src_offset, unsigned long len)
3320 {
3321         size_t cur;
3322         size_t dst_off_in_page;
3323         size_t src_off_in_page;
3324         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3325         unsigned long dst_i;
3326         unsigned long src_i;
3327
3328         if (src_offset + len > dst->len) {
3329                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
3330                        src_offset, len, dst->len);
3331                 BUG_ON(1);
3332         }
3333         if (dst_offset + len > dst->len) {
3334                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
3335                        dst_offset, len, dst->len);
3336                 BUG_ON(1);
3337         }
3338
3339         while(len > 0) {
3340                 dst_off_in_page = (start_offset + dst_offset) &
3341                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3342                 src_off_in_page = (start_offset + src_offset) &
3343                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3344
3345                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3346                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3347
3348                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3349                                                src_off_in_page));
3350                 cur = min_t(unsigned long, cur,
3351                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3352
3353                 copy_pages(extent_buffer_page(dst, dst_i),
3354                            extent_buffer_page(dst, src_i),
3355                            dst_off_in_page, src_off_in_page, cur);
3356
3357                 src_offset += cur;
3358                 dst_offset += cur;
3359                 len -= cur;
3360         }
3361 }
3362 EXPORT_SYMBOL(memcpy_extent_buffer);
3363
3364 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3365                            unsigned long src_offset, unsigned long len)
3366 {
3367         size_t cur;
3368         size_t dst_off_in_page;
3369         size_t src_off_in_page;
3370         unsigned long dst_end = dst_offset + len - 1;
3371         unsigned long src_end = src_offset + len - 1;
3372         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3373         unsigned long dst_i;
3374         unsigned long src_i;
3375
3376         if (src_offset + len > dst->len) {
3377                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
3378                        src_offset, len, dst->len);
3379                 BUG_ON(1);
3380         }
3381         if (dst_offset + len > dst->len) {
3382                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
3383                        dst_offset, len, dst->len);
3384                 BUG_ON(1);
3385         }
3386         if (dst_offset < src_offset) {
3387                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3388                 return;
3389         }
3390         while(len > 0) {
3391                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3392                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3393
3394                 dst_off_in_page = (start_offset + dst_end) &
3395                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3396                 src_off_in_page = (start_offset + src_end) &
3397                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3398
3399                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3400                 cur = min(cur, dst_off_in_page + 1);
3401                 move_pages(extent_buffer_page(dst, dst_i),
3402                            extent_buffer_page(dst, src_i),
3403                            dst_off_in_page - cur + 1,
3404                            src_off_in_page - cur + 1, cur);
3405
3406                 dst_end -= cur;
3407                 src_end -= cur;
3408                 len -= cur;
3409         }
3410 }
3411 EXPORT_SYMBOL(memmove_extent_buffer);
3412
3413 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3414 {
3415         u64 start = page_offset(page);
3416         struct extent_buffer *eb;
3417         int ret = 1;
3418         unsigned long i;
3419         unsigned long num_pages;
3420
3421         spin_lock(&tree->buffer_lock);
3422         eb = buffer_search(tree, start);
3423         if (!eb)
3424                 goto out;
3425
3426         if (atomic_read(&eb->refs) > 1) {
3427                 ret = 0;
3428                 goto out;
3429         }
3430         /* at this point we can safely release the extent buffer */
3431         num_pages = num_extent_pages(eb->start, eb->len);
3432         for (i = 0; i < num_pages; i++)
3433                 page_cache_release(extent_buffer_page(eb, i));
3434         rb_erase(&eb->rb_node, &tree->buffer);
3435         __free_extent_buffer(eb);
3436 out:
3437         spin_unlock(&tree->buffer_lock);
3438         return ret;
3439 }
3440 EXPORT_SYMBOL(try_release_extent_buffer);