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
  18 /* temporary define until extent_map moves out of btrfs */
  19 struct kmem_cache *btrfs_cache_create(const char *name, size_t size,
  20                                        unsigned long extra_flags,
  21                                        void (*ctor)(void *, struct kmem_cache *,
  22                                                     unsigned long));
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26
  27 static LIST_HEAD(buffers);
  28 static LIST_HEAD(states);
  29 static spinlock_t leak_lock = SPIN_LOCK_UNLOCKED;
  30
  31 #define BUFFER_LRU_MAX 64
  32
  33 struct tree_entry {
  34         u64 start;
  35         u64 end;
  36         struct rb_node rb_node;
  37 };
  38
  39 struct extent_page_data {
  40         struct bio *bio;
  41         struct extent_io_tree *tree;
  42         get_extent_t *get_extent;
  43 };
  44
  45 int __init extent_io_init(void)
  46 {
  47         extent_state_cache = btrfs_cache_create("extent_state",
  48                                             sizeof(struct extent_state), 0,
  49                                             NULL);
  50         if (!extent_state_cache)
  51                 return -ENOMEM;
  52
  53         extent_buffer_cache = btrfs_cache_create("extent_buffers",
  54                                             sizeof(struct extent_buffer), 0,
  55                                             NULL);
  56         if (!extent_buffer_cache)
  57                 goto free_state_cache;
  58         return 0;
  59
  60 free_state_cache:
  61         kmem_cache_destroy(extent_state_cache);
  62         return -ENOMEM;
  63 }
  64
  65 void extent_io_exit(void)
  66 {
  67         struct extent_state *state;
  68         struct extent_buffer *eb;
  69
  70         while (!list_empty(&states)) {
  71                 state = list_entry(states.next, struct extent_state, leak_list);
  72                 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));
  73                 list_del(&state->leak_list);
  74                 kmem_cache_free(extent_state_cache, state);
  75
  76         }
  77
  78         while (!list_empty(&buffers)) {
  79                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  80                 printk("buffer leak start %Lu len %lu refs %d\n", eb->start, eb->len, atomic_read(&eb->refs));
  81                 list_del(&eb->leak_list);
  82                 kmem_cache_free(extent_buffer_cache, eb);
  83         }
  84         if (extent_state_cache)
  85                 kmem_cache_destroy(extent_state_cache);
  86         if (extent_buffer_cache)
  87                 kmem_cache_destroy(extent_buffer_cache);
  88 }
  89
  90 void extent_io_tree_init(struct extent_io_tree *tree,
  91                           struct address_space *mapping, gfp_t mask)
  92 {
  93         tree->state.rb_node = NULL;
  94         tree->ops = NULL;
  95         tree->dirty_bytes = 0;
  96         spin_lock_init(&tree->lock);
  97         spin_lock_init(&tree->lru_lock);
  98         tree->mapping = mapping;
  99         INIT_LIST_HEAD(&tree->buffer_lru);
 100         tree->lru_size = 0;
 101         tree->last = NULL;
 102 }
 103 EXPORT_SYMBOL(extent_io_tree_init);
 104
 105 void extent_io_tree_empty_lru(struct extent_io_tree *tree)
 106 {
 107         struct extent_buffer *eb;
 108         while(!list_empty(&tree->buffer_lru)) {
 109                 eb = list_entry(tree->buffer_lru.next, struct extent_buffer,
 110                                 lru);
 111                 list_del_init(&eb->lru);
 112                 free_extent_buffer(eb);
 113         }
 114 }
 115 EXPORT_SYMBOL(extent_io_tree_empty_lru);
 116
 117 struct extent_state *alloc_extent_state(gfp_t mask)
 118 {
 119         struct extent_state *state;
 120         unsigned long flags;
 121
 122         state = kmem_cache_alloc(extent_state_cache, mask);
 123         if (!state)
 124                 return state;
 125         state->state = 0;
 126         state->private = 0;
 127         state->tree = NULL;
 128         spin_lock_irqsave(&leak_lock, flags);
 129         list_add(&state->leak_list, &states);
 130         spin_unlock_irqrestore(&leak_lock, flags);
 131
 132         atomic_set(&state->refs, 1);
 133         init_waitqueue_head(&state->wq);
 134         return state;
 135 }
 136 EXPORT_SYMBOL(alloc_extent_state);
 137
 138 void free_extent_state(struct extent_state *state)
 139 {
 140         if (!state)
 141                 return;
 142         if (atomic_dec_and_test(&state->refs)) {
 143                 unsigned long flags;
 144                 WARN_ON(state->tree);
 145                 spin_lock_irqsave(&leak_lock, flags);
 146                 list_del(&state->leak_list);
 147                 spin_unlock_irqrestore(&leak_lock, flags);
 148                 kmem_cache_free(extent_state_cache, state);
 149         }
 150 }
 151 EXPORT_SYMBOL(free_extent_state);
 152
 153 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 154                                    struct rb_node *node)
 155 {
 156         struct rb_node ** p = &root->rb_node;
 157         struct rb_node * parent = NULL;
 158         struct tree_entry *entry;
 159
 160         while(*p) {
 161                 parent = *p;
 162                 entry = rb_entry(parent, struct tree_entry, rb_node);
 163
 164                 if (offset < entry->start)
 165                         p = &(*p)->rb_left;
 166                 else if (offset > entry->end)
 167                         p = &(*p)->rb_right;
 168                 else
 169                         return parent;
 170         }
 171
 172         entry = rb_entry(node, struct tree_entry, rb_node);
 173         rb_link_node(node, parent, p);
 174         rb_insert_color(node, root);
 175         return NULL;
 176 }
 177
 178 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 179                                      struct rb_node **prev_ret,
 180                                      struct rb_node **next_ret)
 181 {
 182         struct rb_root *root = &tree->state;
 183         struct rb_node * n = root->rb_node;
 184         struct rb_node *prev = NULL;
 185         struct rb_node *orig_prev = NULL;
 186         struct tree_entry *entry;
 187         struct tree_entry *prev_entry = NULL;
 188
 189         if (tree->last) {
 190                 struct extent_state *state;
 191                 state = tree->last;
 192                 if (state->start <= offset && offset <= state->end)
 193                         return &tree->last->rb_node;
 194         }
 195         while(n) {
 196                 entry = rb_entry(n, struct tree_entry, rb_node);
 197                 prev = n;
 198                 prev_entry = entry;
 199
 200                 if (offset < entry->start)
 201                         n = n->rb_left;
 202                 else if (offset > entry->end)
 203                         n = n->rb_right;
 204                 else {
 205                         tree->last = rb_entry(n, struct extent_state, rb_node);
 206                         return n;
 207                 }
 208         }
 209
 210         if (prev_ret) {
 211                 orig_prev = prev;
 212                 while(prev && offset > prev_entry->end) {
 213                         prev = rb_next(prev);
 214                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 215                 }
 216                 *prev_ret = prev;
 217                 prev = orig_prev;
 218         }
 219
 220         if (next_ret) {
 221                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 222                 while(prev && offset < prev_entry->start) {
 223                         prev = rb_prev(prev);
 224                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 225                 }
 226                 *next_ret = prev;
 227         }
 228         return NULL;
 229 }
 230
 231 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 232                                           u64 offset)
 233 {
 234         struct rb_node *prev = NULL;
 235         struct rb_node *ret;
 236
 237         ret = __etree_search(tree, offset, &prev, NULL);
 238         if (!ret) {
 239                 if (prev) {
 240                         tree->last = rb_entry(prev, struct extent_state,
 241                                               rb_node);
 242                 }
 243                 return prev;
 244         }
 245         return ret;
 246 }
 247
 248 /*
 249  * utility function to look for merge candidates inside a given range.
 250  * Any extents with matching state are merged together into a single
 251  * extent in the tree.  Extents with EXTENT_IO in their state field
 252  * are not merged because the end_io handlers need to be able to do
 253  * operations on them without sleeping (or doing allocations/splits).
 254  *
 255  * This should be called with the tree lock held.
 256  */
 257 static int merge_state(struct extent_io_tree *tree,
 258                        struct extent_state *state)
 259 {
 260         struct extent_state *other;
 261         struct rb_node *other_node;
 262
 263         if (state->state & EXTENT_IOBITS)
 264                 return 0;
 265
 266         other_node = rb_prev(&state->rb_node);
 267         if (other_node) {
 268                 other = rb_entry(other_node, struct extent_state, rb_node);
 269                 if (other->end == state->start - 1 &&
 270                     other->state == state->state) {
 271                         state->start = other->start;
 272                         other->tree = NULL;
 273                         if (tree->last == other)
 274                                 tree->last = state;
 275                         rb_erase(&other->rb_node, &tree->state);
 276                         free_extent_state(other);
 277                 }
 278         }
 279         other_node = rb_next(&state->rb_node);
 280         if (other_node) {
 281                 other = rb_entry(other_node, struct extent_state, rb_node);
 282                 if (other->start == state->end + 1 &&
 283                     other->state == state->state) {
 284                         other->start = state->start;
 285                         state->tree = NULL;
 286                         if (tree->last == state)
 287                                 tree->last = other;
 288                         rb_erase(&state->rb_node, &tree->state);
 289                         free_extent_state(state);
 290                 }
 291         }
 292         return 0;
 293 }
 294
 295 static void set_state_cb(struct extent_io_tree *tree,
 296                          struct extent_state *state,
 297                          unsigned long bits)
 298 {
 299         if (tree->ops && tree->ops->set_bit_hook) {
 300                 tree->ops->set_bit_hook(tree->mapping->host, state->start,
 301                                         state->end, state->state, bits);
 302         }
 303 }
 304
 305 static void clear_state_cb(struct extent_io_tree *tree,
 306                            struct extent_state *state,
 307                            unsigned long bits)
 308 {
 309         if (tree->ops && tree->ops->set_bit_hook) {
 310                 tree->ops->clear_bit_hook(tree->mapping->host, state->start,
 311                                           state->end, state->state, bits);
 312         }
 313 }
 314
 315 /*
 316  * insert an extent_state struct into the tree.  'bits' are set on the
 317  * struct before it is inserted.
 318  *
 319  * This may return -EEXIST if the extent is already there, in which case the
 320  * state struct is freed.
 321  *
 322  * The tree lock is not taken internally.  This is a utility function and
 323  * probably isn't what you want to call (see set/clear_extent_bit).
 324  */
 325 static int insert_state(struct extent_io_tree *tree,
 326                         struct extent_state *state, u64 start, u64 end,
 327                         int bits)
 328 {
 329         struct rb_node *node;
 330
 331         if (end < start) {
 332                 printk("end < start %Lu %Lu\n", end, start);
 333                 WARN_ON(1);
 334         }
 335         if (bits & EXTENT_DIRTY)
 336                 tree->dirty_bytes += end - start + 1;
 337         set_state_cb(tree, state, bits);
 338         state->state |= bits;
 339         state->start = start;
 340         state->end = end;
 341         node = tree_insert(&tree->state, end, &state->rb_node);
 342         if (node) {
 343                 struct extent_state *found;
 344                 found = rb_entry(node, struct extent_state, rb_node);
 345                 printk("found node %Lu %Lu on insert of %Lu %Lu\n", found->start, found->end, start, end);
 346                 free_extent_state(state);
 347                 return -EEXIST;
 348         }
 349         state->tree = tree;
 350         tree->last = state;
 351         merge_state(tree, state);
 352         return 0;
 353 }
 354
 355 /*
 356  * split a given extent state struct in two, inserting the preallocated
 357  * struct 'prealloc' as the newly created second half.  'split' indicates an
 358  * offset inside 'orig' where it should be split.
 359  *
 360  * Before calling,
 361  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 362  * are two extent state structs in the tree:
 363  * prealloc: [orig->start, split - 1]
 364  * orig: [ split, orig->end ]
 365  *
 366  * The tree locks are not taken by this function. They need to be held
 367  * by the caller.
 368  */
 369 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 370                        struct extent_state *prealloc, u64 split)
 371 {
 372         struct rb_node *node;
 373         prealloc->start = orig->start;
 374         prealloc->end = split - 1;
 375         prealloc->state = orig->state;
 376         orig->start = split;
 377
 378         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 379         if (node) {
 380                 struct extent_state *found;
 381                 found = rb_entry(node, struct extent_state, rb_node);
 382                 printk("found node %Lu %Lu on insert of %Lu %Lu\n", found->start, found->end, prealloc->start, prealloc->end);
 383                 free_extent_state(prealloc);
 384                 return -EEXIST;
 385         }
 386         prealloc->tree = tree;
 387         return 0;
 388 }
 389
 390 /*
 391  * utility function to clear some bits in an extent state struct.
 392  * it will optionally wake up any one waiting on this state (wake == 1), or
 393  * forcibly remove the state from the tree (delete == 1).
 394  *
 395  * If no bits are set on the state struct after clearing things, the
 396  * struct is freed and removed from the tree
 397  */
 398 static int clear_state_bit(struct extent_io_tree *tree,
 399                             struct extent_state *state, int bits, int wake,
 400                             int delete)
 401 {
 402         int ret = state->state & bits;
 403
 404         if ((bits & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 405                 u64 range = state->end - state->start + 1;
 406                 WARN_ON(range > tree->dirty_bytes);
 407                 tree->dirty_bytes -= range;
 408         }
 409         clear_state_cb(tree, state, bits);
 410         state->state &= ~bits;
 411         if (wake)
 412                 wake_up(&state->wq);
 413         if (delete || state->state == 0) {
 414                 if (state->tree) {
 415                         clear_state_cb(tree, state, state->state);
 416                         if (tree->last == state) {
 417                                 tree->last = extent_state_next(state);
 418                         }
 419                         rb_erase(&state->rb_node, &tree->state);
 420                         state->tree = NULL;
 421                         free_extent_state(state);
 422                 } else {
 423                         WARN_ON(1);
 424                 }
 425         } else {
 426                 merge_state(tree, state);
 427         }
 428         return ret;
 429 }
 430
 431 /*
 432  * clear some bits on a range in the tree.  This may require splitting
 433  * or inserting elements in the tree, so the gfp mask is used to
 434  * indicate which allocations or sleeping are allowed.
 435  *
 436  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 437  * the given range from the tree regardless of state (ie for truncate).
 438  *
 439  * the range [start, end] is inclusive.
 440  *
 441  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 442  * bits were already set, or zero if none of the bits were already set.
 443  */
 444 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 445                      int bits, int wake, int delete, gfp_t mask)
 446 {
 447         struct extent_state *state;
 448         struct extent_state *prealloc = NULL;
 449         struct rb_node *node;
 450         unsigned long flags;
 451         int err;
 452         int set = 0;
 453
 454 again:
 455         if (!prealloc && (mask & __GFP_WAIT)) {
 456                 prealloc = alloc_extent_state(mask);
 457                 if (!prealloc)
 458                         return -ENOMEM;
 459         }
 460
 461         spin_lock_irqsave(&tree->lock, flags);
 462         /*
 463          * this search will find the extents that end after
 464          * our range starts
 465          */
 466         node = tree_search(tree, start);
 467         if (!node)
 468                 goto out;
 469         state = rb_entry(node, struct extent_state, rb_node);
 470         if (state->start > end)
 471                 goto out;
 472         WARN_ON(state->end < start);
 473
 474         /*
 475          *     | ---- desired range ---- |
 476          *  | state | or
 477          *  | ------------- state -------------- |
 478          *
 479          * We need to split the extent we found, and may flip
 480          * bits on second half.
 481          *
 482          * If the extent we found extends past our range, we
 483          * just split and search again.  It'll get split again
 484          * the next time though.
 485          *
 486          * If the extent we found is inside our range, we clear
 487          * the desired bit on it.
 488          */
 489
 490         if (state->start < start) {
 491                 if (!prealloc)
 492                         prealloc = alloc_extent_state(GFP_ATOMIC);
 493                 err = split_state(tree, state, prealloc, start);
 494                 BUG_ON(err == -EEXIST);
 495                 prealloc = NULL;
 496                 if (err)
 497                         goto out;
 498                 if (state->end <= end) {
 499                         start = state->end + 1;
 500                         set |= clear_state_bit(tree, state, bits,
 501                                         wake, delete);
 502                 } else {
 503                         start = state->start;
 504                 }
 505                 goto search_again;
 506         }
 507         /*
 508          * | ---- desired range ---- |
 509          *                        | state |
 510          * We need to split the extent, and clear the bit
 511          * on the first half
 512          */
 513         if (state->start <= end && state->end > end) {
 514                 if (!prealloc)
 515                         prealloc = alloc_extent_state(GFP_ATOMIC);
 516                 err = split_state(tree, state, prealloc, end + 1);
 517                 BUG_ON(err == -EEXIST);
 518
 519                 if (wake)
 520                         wake_up(&state->wq);
 521                 set |= clear_state_bit(tree, prealloc, bits,
 522                                        wake, delete);
 523                 prealloc = NULL;
 524                 goto out;
 525         }
 526
 527         start = state->end + 1;
 528         set |= clear_state_bit(tree, state, bits, wake, delete);
 529         goto search_again;
 530
 531 out:
 532         spin_unlock_irqrestore(&tree->lock, flags);
 533         if (prealloc)
 534                 free_extent_state(prealloc);
 535
 536         return set;
 537
 538 search_again:
 539         if (start > end)
 540                 goto out;
 541         spin_unlock_irqrestore(&tree->lock, flags);
 542         if (mask & __GFP_WAIT)
 543                 cond_resched();
 544         goto again;
 545 }
 546 EXPORT_SYMBOL(clear_extent_bit);
 547
 548 static int wait_on_state(struct extent_io_tree *tree,
 549                          struct extent_state *state)
 550 {
 551         DEFINE_WAIT(wait);
 552         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 553         spin_unlock_irq(&tree->lock);
 554         schedule();
 555         spin_lock_irq(&tree->lock);
 556         finish_wait(&state->wq, &wait);
 557         return 0;
 558 }
 559
 560 /*
 561  * waits for one or more bits to clear on a range in the state tree.
 562  * The range [start, end] is inclusive.
 563  * The tree lock is taken by this function
 564  */
 565 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 566 {
 567         struct extent_state *state;
 568         struct rb_node *node;
 569
 570         spin_lock_irq(&tree->lock);
 571 again:
 572         while (1) {
 573                 /*
 574                  * this search will find all the extents that end after
 575                  * our range starts
 576                  */
 577                 node = tree_search(tree, start);
 578                 if (!node)
 579                         break;
 580
 581                 state = rb_entry(node, struct extent_state, rb_node);
 582
 583                 if (state->start > end)
 584                         goto out;
 585
 586                 if (state->state & bits) {
 587                         start = state->start;
 588                         atomic_inc(&state->refs);
 589                         wait_on_state(tree, state);
 590                         free_extent_state(state);
 591                         goto again;
 592                 }
 593                 start = state->end + 1;
 594
 595                 if (start > end)
 596                         break;
 597
 598                 if (need_resched()) {
 599                         spin_unlock_irq(&tree->lock);
 600                         cond_resched();
 601                         spin_lock_irq(&tree->lock);
 602                 }
 603         }
 604 out:
 605         spin_unlock_irq(&tree->lock);
 606         return 0;
 607 }
 608 EXPORT_SYMBOL(wait_extent_bit);
 609
 610 static void set_state_bits(struct extent_io_tree *tree,
 611                            struct extent_state *state,
 612                            int bits)
 613 {
 614         if ((bits & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 615                 u64 range = state->end - state->start + 1;
 616                 tree->dirty_bytes += range;
 617         }
 618         set_state_cb(tree, state, bits);
 619         state->state |= bits;
 620 }
 621
 622 /*
 623  * set some bits on a range in the tree.  This may require allocations
 624  * or sleeping, so the gfp mask is used to indicate what is allowed.
 625  *
 626  * If 'exclusive' == 1, this will fail with -EEXIST if some part of the
 627  * range already has the desired bits set.  The start of the existing
 628  * range is returned in failed_start in this case.
 629  *
 630  * [start, end] is inclusive
 631  * This takes the tree lock.
 632  */
 633 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits,
 634                    int exclusive, u64 *failed_start, gfp_t mask)
 635 {
 636         struct extent_state *state;
 637         struct extent_state *prealloc = NULL;
 638         struct rb_node *node;
 639         unsigned long flags;
 640         int err = 0;
 641         int set;
 642         u64 last_start;
 643         u64 last_end;
 644 again:
 645         if (!prealloc && (mask & __GFP_WAIT)) {
 646                 prealloc = alloc_extent_state(mask);
 647                 if (!prealloc)
 648                         return -ENOMEM;
 649         }
 650
 651         spin_lock_irqsave(&tree->lock, flags);
 652         /*
 653          * this search will find all the extents that end after
 654          * our range starts.
 655          */
 656         node = tree_search(tree, start);
 657         if (!node) {
 658                 err = insert_state(tree, prealloc, start, end, bits);
 659                 prealloc = NULL;
 660                 BUG_ON(err == -EEXIST);
 661                 goto out;
 662         }
 663
 664         state = rb_entry(node, struct extent_state, rb_node);
 665         last_start = state->start;
 666         last_end = state->end;
 667
 668         /*
 669          * | ---- desired range ---- |
 670          * | state |
 671          *
 672          * Just lock what we found and keep going
 673          */
 674         if (state->start == start && state->end <= end) {
 675                 set = state->state & bits;
 676                 if (set && exclusive) {
 677                         *failed_start = state->start;
 678                         err = -EEXIST;
 679                         goto out;
 680                 }
 681                 set_state_bits(tree, state, bits);
 682                 start = state->end + 1;
 683                 merge_state(tree, state);
 684                 goto search_again;
 685         }
 686
 687         /*
 688          *     | ---- desired range ---- |
 689          * | state |
 690          *   or
 691          * | ------------- state -------------- |
 692          *
 693          * We need to split the extent we found, and may flip bits on
 694          * second half.
 695          *
 696          * If the extent we found extends past our
 697          * range, we just split and search again.  It'll get split
 698          * again the next time though.
 699          *
 700          * If the extent we found is inside our range, we set the
 701          * desired bit on it.
 702          */
 703         if (state->start < start) {
 704                 set = state->state & bits;
 705                 if (exclusive && set) {
 706                         *failed_start = start;
 707                         err = -EEXIST;
 708                         goto out;
 709                 }
 710                 err = split_state(tree, state, prealloc, start);
 711                 BUG_ON(err == -EEXIST);
 712                 prealloc = NULL;
 713                 if (err)
 714                         goto out;
 715                 if (state->end <= end) {
 716                         set_state_bits(tree, state, bits);
 717                         start = state->end + 1;
 718                         merge_state(tree, state);
 719                 } else {
 720                         start = state->start;
 721                 }
 722                 goto search_again;
 723         }
 724         /*
 725          * | ---- desired range ---- |
 726          *     | state | or               | state |
 727          *
 728          * There's a hole, we need to insert something in it and
 729          * ignore the extent we found.
 730          */
 731         if (state->start > start) {
 732                 u64 this_end;
 733                 if (end < last_start)
 734                         this_end = end;
 735                 else
 736                         this_end = last_start -1;
 737                 err = insert_state(tree, prealloc, start, this_end,
 738                                    bits);
 739                 prealloc = NULL;
 740                 BUG_ON(err == -EEXIST);
 741                 if (err)
 742                         goto out;
 743                 start = this_end + 1;
 744                 goto search_again;
 745         }
 746         /*
 747          * | ---- desired range ---- |
 748          *                        | state |
 749          * We need to split the extent, and set the bit
 750          * on the first half
 751          */
 752         if (state->start <= end && state->end > end) {
 753                 set = state->state & bits;
 754                 if (exclusive && set) {
 755                         *failed_start = start;
 756                         err = -EEXIST;
 757                         goto out;
 758                 }
 759                 err = split_state(tree, state, prealloc, end + 1);
 760                 BUG_ON(err == -EEXIST);
 761
 762                 set_state_bits(tree, prealloc, bits);
 763                 merge_state(tree, prealloc);
 764                 prealloc = NULL;
 765                 goto out;
 766         }
 767
 768         goto search_again;
 769
 770 out:
 771         spin_unlock_irqrestore(&tree->lock, flags);
 772         if (prealloc)
 773                 free_extent_state(prealloc);
 774
 775         return err;
 776
 777 search_again:
 778         if (start > end)
 779                 goto out;
 780         spin_unlock_irqrestore(&tree->lock, flags);
 781         if (mask & __GFP_WAIT)
 782                 cond_resched();
 783         goto again;
 784 }
 785 EXPORT_SYMBOL(set_extent_bit);
 786
 787 /* wrappers around set/clear extent bit */
 788 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 789                      gfp_t mask)
 790 {
 791         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 792                               mask);
 793 }
 794 EXPORT_SYMBOL(set_extent_dirty);
 795
 796 int set_extent_ordered(struct extent_io_tree *tree, u64 start, u64 end,
 797                        gfp_t mask)
 798 {
 799         return set_extent_bit(tree, start, end, EXTENT_ORDERED, 0, NULL, mask);
 800 }
 801 EXPORT_SYMBOL(set_extent_ordered);
 802
 803 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 804                     int bits, gfp_t mask)
 805 {
 806         return set_extent_bit(tree, start, end, bits, 0, NULL,
 807                               mask);
 808 }
 809 EXPORT_SYMBOL(set_extent_bits);
 810
 811 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 812                       int bits, gfp_t mask)
 813 {
 814         return clear_extent_bit(tree, start, end, bits, 0, 0, mask);
 815 }
 816 EXPORT_SYMBOL(clear_extent_bits);
 817
 818 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
 819                      gfp_t mask)
 820 {
 821         return set_extent_bit(tree, start, end,
 822                               EXTENT_DELALLOC | EXTENT_DIRTY,
 823                               0, NULL, mask);
 824 }
 825 EXPORT_SYMBOL(set_extent_delalloc);
 826
 827 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 828                        gfp_t mask)
 829 {
 830         return clear_extent_bit(tree, start, end,
 831                                 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, mask);
 832 }
 833 EXPORT_SYMBOL(clear_extent_dirty);
 834
 835 int clear_extent_ordered(struct extent_io_tree *tree, u64 start, u64 end,
 836                          gfp_t mask)
 837 {
 838         return clear_extent_bit(tree, start, end, EXTENT_ORDERED, 1, 0, mask);
 839 }
 840 EXPORT_SYMBOL(clear_extent_ordered);
 841
 842 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 843                      gfp_t mask)
 844 {
 845         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 846                               mask);
 847 }
 848 EXPORT_SYMBOL(set_extent_new);
 849
 850 int clear_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 851                        gfp_t mask)
 852 {
 853         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0, mask);
 854 }
 855 EXPORT_SYMBOL(clear_extent_new);
 856
 857 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 858                         gfp_t mask)
 859 {
 860         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, NULL,
 861                               mask);
 862 }
 863 EXPORT_SYMBOL(set_extent_uptodate);
 864
 865 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 866                           gfp_t mask)
 867 {
 868         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0, mask);
 869 }
 870 EXPORT_SYMBOL(clear_extent_uptodate);
 871
 872 int set_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end,
 873                          gfp_t mask)
 874 {
 875         return set_extent_bit(tree, start, end, EXTENT_WRITEBACK,
 876                               0, NULL, mask);
 877 }
 878 EXPORT_SYMBOL(set_extent_writeback);
 879
 880 int clear_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end,
 881                            gfp_t mask)
 882 {
 883         return clear_extent_bit(tree, start, end, EXTENT_WRITEBACK, 1, 0, mask);
 884 }
 885 EXPORT_SYMBOL(clear_extent_writeback);
 886
 887 int wait_on_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 888 {
 889         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 890 }
 891 EXPORT_SYMBOL(wait_on_extent_writeback);
 892
 893 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
 894 {
 895         int err;
 896         u64 failed_start;
 897         while (1) {
 898                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED, 1,
 899                                      &failed_start, mask);
 900                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
 901                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
 902                         start = failed_start;
 903                 } else {
 904                         break;
 905                 }
 906                 WARN_ON(start > end);
 907         }
 908         return err;
 909 }
 910 EXPORT_SYMBOL(lock_extent);
 911
 912 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end,
 913                   gfp_t mask)
 914 {
 915         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, mask);
 916 }
 917 EXPORT_SYMBOL(unlock_extent);
 918
 919 /*
 920  * helper function to set pages and extents in the tree dirty
 921  */
 922 int set_range_dirty(struct extent_io_tree *tree, u64 start, u64 end)
 923 {
 924         unsigned long index = start >> PAGE_CACHE_SHIFT;
 925         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 926         struct page *page;
 927
 928         while (index <= end_index) {
 929                 page = find_get_page(tree->mapping, index);
 930                 BUG_ON(!page);
 931                 __set_page_dirty_nobuffers(page);
 932                 page_cache_release(page);
 933                 index++;
 934         }
 935         set_extent_dirty(tree, start, end, GFP_NOFS);
 936         return 0;
 937 }
 938 EXPORT_SYMBOL(set_range_dirty);
 939
 940 /*
 941  * helper function to set both pages and extents in the tree writeback
 942  */
 943 int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 944 {
 945         unsigned long index = start >> PAGE_CACHE_SHIFT;
 946         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 947         struct page *page;
 948
 949         while (index <= end_index) {
 950                 page = find_get_page(tree->mapping, index);
 951                 BUG_ON(!page);
 952                 set_page_writeback(page);
 953                 page_cache_release(page);
 954                 index++;
 955         }
 956         set_extent_writeback(tree, start, end, GFP_NOFS);
 957         return 0;
 958 }
 959 EXPORT_SYMBOL(set_range_writeback);
 960
 961 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
 962                           u64 *start_ret, u64 *end_ret, int bits)
 963 {
 964         struct rb_node *node;
 965         struct extent_state *state;
 966         int ret = 1;
 967
 968         spin_lock_irq(&tree->lock);
 969         /*
 970          * this search will find all the extents that end after
 971          * our range starts.
 972          */
 973         node = tree_search(tree, start);
 974         if (!node) {
 975                 goto out;
 976         }
 977
 978         while(1) {
 979                 state = rb_entry(node, struct extent_state, rb_node);
 980                 if (state->end >= start && (state->state & bits)) {
 981                         *start_ret = state->start;
 982                         *end_ret = state->end;
 983                         ret = 0;
 984                         break;
 985                 }
 986                 node = rb_next(node);
 987                 if (!node)
 988                         break;
 989         }
 990 out:
 991         spin_unlock_irq(&tree->lock);
 992         return ret;
 993 }
 994 EXPORT_SYMBOL(find_first_extent_bit);
 995
 996 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
 997                                                  u64 start, int bits)
 998 {
 999         struct rb_node *node;
1000         struct extent_state *state;
1001
1002         /*
1003          * this search will find all the extents that end after
1004          * our range starts.
1005          */
1006         node = tree_search(tree, start);
1007         if (!node) {
1008                 goto out;
1009         }
1010
1011         while(1) {
1012                 state = rb_entry(node, struct extent_state, rb_node);
1013                 if (state->end >= start && (state->state & bits)) {
1014                         return state;
1015                 }
1016                 node = rb_next(node);
1017                 if (!node)
1018                         break;
1019         }
1020 out:
1021         return NULL;
1022 }
1023 EXPORT_SYMBOL(find_first_extent_bit_state);
1024
1025 u64 find_lock_delalloc_range(struct extent_io_tree *tree,
1026                              u64 *start, u64 *end, u64 max_bytes)
1027 {
1028         struct rb_node *node;
1029         struct extent_state *state;
1030         u64 cur_start = *start;
1031         u64 found = 0;
1032         u64 total_bytes = 0;
1033
1034         spin_lock_irq(&tree->lock);
1035         /*
1036          * this search will find all the extents that end after
1037          * our range starts.
1038          */
1039 search_again:
1040         node = tree_search(tree, cur_start);
1041         if (!node) {
1042                 if (!found)
1043                         *end = (u64)-1;
1044                 goto out;
1045         }
1046
1047         while(1) {
1048                 state = rb_entry(node, struct extent_state, rb_node);
1049                 if (found && state->start != cur_start) {
1050                         goto out;
1051                 }
1052                 if (!(state->state & EXTENT_DELALLOC)) {
1053                         if (!found)
1054                                 *end = state->end;
1055                         goto out;
1056                 }
1057                 if (!found) {
1058                         struct extent_state *prev_state;
1059                         struct rb_node *prev_node = node;
1060                         while(1) {
1061                                 prev_node = rb_prev(prev_node);
1062                                 if (!prev_node)
1063                                         break;
1064                                 prev_state = rb_entry(prev_node,
1065                                                       struct extent_state,
1066                                                       rb_node);
1067                                 if (!(prev_state->state & EXTENT_DELALLOC))
1068                                         break;
1069                                 state = prev_state;
1070                                 node = prev_node;
1071                         }
1072                 }
1073                 if (state->state & EXTENT_LOCKED) {
1074                         DEFINE_WAIT(wait);
1075                         atomic_inc(&state->refs);
1076                         prepare_to_wait(&state->wq, &wait,
1077                                         TASK_UNINTERRUPTIBLE);
1078                         spin_unlock_irq(&tree->lock);
1079                         schedule();
1080                         spin_lock_irq(&tree->lock);
1081                         finish_wait(&state->wq, &wait);
1082                         free_extent_state(state);
1083                         goto search_again;
1084                 }
1085                 set_state_cb(tree, state, EXTENT_LOCKED);
1086                 state->state |= EXTENT_LOCKED;
1087                 if (!found)
1088                         *start = state->start;
1089                 found++;
1090                 *end = state->end;
1091                 cur_start = state->end + 1;
1092                 node = rb_next(node);
1093                 if (!node)
1094                         break;
1095                 total_bytes += state->end - state->start + 1;
1096                 if (total_bytes >= max_bytes)
1097                         break;
1098         }
1099 out:
1100         spin_unlock_irq(&tree->lock);
1101         return found;
1102 }
1103
1104 u64 count_range_bits(struct extent_io_tree *tree,
1105                      u64 *start, u64 search_end, u64 max_bytes,
1106                      unsigned long bits)
1107 {
1108         struct rb_node *node;
1109         struct extent_state *state;
1110         u64 cur_start = *start;
1111         u64 total_bytes = 0;
1112         int found = 0;
1113
1114         if (search_end <= cur_start) {
1115                 printk("search_end %Lu start %Lu\n", search_end, cur_start);
1116                 WARN_ON(1);
1117                 return 0;
1118         }
1119
1120         spin_lock_irq(&tree->lock);
1121         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1122                 total_bytes = tree->dirty_bytes;
1123                 goto out;
1124         }
1125         /*
1126          * this search will find all the extents that end after
1127          * our range starts.
1128          */
1129         node = tree_search(tree, cur_start);
1130         if (!node) {
1131                 goto out;
1132         }
1133
1134         while(1) {
1135                 state = rb_entry(node, struct extent_state, rb_node);
1136                 if (state->start > search_end)
1137                         break;
1138                 if (state->end >= cur_start && (state->state & bits)) {
1139                         total_bytes += min(search_end, state->end) + 1 -
1140                                        max(cur_start, state->start);
1141                         if (total_bytes >= max_bytes)
1142                                 break;
1143                         if (!found) {
1144                                 *start = state->start;
1145                                 found = 1;
1146                         }
1147                 }
1148                 node = rb_next(node);
1149                 if (!node)
1150                         break;
1151         }
1152 out:
1153         spin_unlock_irq(&tree->lock);
1154         return total_bytes;
1155 }
1156 /*
1157  * helper function to lock both pages and extents in the tree.
1158  * pages must be locked first.
1159  */
1160 int lock_range(struct extent_io_tree *tree, u64 start, u64 end)
1161 {
1162         unsigned long index = start >> PAGE_CACHE_SHIFT;
1163         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1164         struct page *page;
1165         int err;
1166
1167         while (index <= end_index) {
1168                 page = grab_cache_page(tree->mapping, index);
1169                 if (!page) {
1170                         err = -ENOMEM;
1171                         goto failed;
1172                 }
1173                 if (IS_ERR(page)) {
1174                         err = PTR_ERR(page);
1175                         goto failed;
1176                 }
1177                 index++;
1178         }
1179         lock_extent(tree, start, end, GFP_NOFS);
1180         return 0;
1181
1182 failed:
1183         /*
1184          * we failed above in getting the page at 'index', so we undo here
1185          * up to but not including the page at 'index'
1186          */
1187         end_index = index;
1188         index = start >> PAGE_CACHE_SHIFT;
1189         while (index < end_index) {
1190                 page = find_get_page(tree->mapping, index);
1191                 unlock_page(page);
1192                 page_cache_release(page);
1193                 index++;
1194         }
1195         return err;
1196 }
1197 EXPORT_SYMBOL(lock_range);
1198
1199 /*
1200  * helper function to unlock both pages and extents in the tree.
1201  */
1202 int unlock_range(struct extent_io_tree *tree, u64 start, u64 end)
1203 {
1204         unsigned long index = start >> PAGE_CACHE_SHIFT;
1205         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1206         struct page *page;
1207
1208         while (index <= end_index) {
1209                 page = find_get_page(tree->mapping, index);
1210                 unlock_page(page);
1211                 page_cache_release(page);
1212                 index++;
1213         }
1214         unlock_extent(tree, start, end, GFP_NOFS);
1215         return 0;
1216 }
1217 EXPORT_SYMBOL(unlock_range);
1218
1219 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1220 {
1221         struct rb_node *node;
1222         struct extent_state *state;
1223         int ret = 0;
1224
1225         spin_lock_irq(&tree->lock);
1226         /*
1227          * this search will find all the extents that end after
1228          * our range starts.
1229          */
1230         node = tree_search(tree, start);
1231         if (!node) {
1232                 ret = -ENOENT;
1233                 goto out;
1234         }
1235         state = rb_entry(node, struct extent_state, rb_node);
1236         if (state->start != start) {
1237                 ret = -ENOENT;
1238                 goto out;
1239         }
1240         state->private = private;
1241 out:
1242         spin_unlock_irq(&tree->lock);
1243         return ret;
1244 }
1245
1246 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1247 {
1248         struct rb_node *node;
1249         struct extent_state *state;
1250         int ret = 0;
1251
1252         spin_lock_irq(&tree->lock);
1253         /*
1254          * this search will find all the extents that end after
1255          * our range starts.
1256          */
1257         node = tree_search(tree, start);
1258         if (!node) {
1259                 ret = -ENOENT;
1260                 goto out;
1261         }
1262         state = rb_entry(node, struct extent_state, rb_node);
1263         if (state->start != start) {
1264                 ret = -ENOENT;
1265                 goto out;
1266         }
1267         *private = state->private;
1268 out:
1269         spin_unlock_irq(&tree->lock);
1270         return ret;
1271 }
1272
1273 /*
1274  * searches a range in the state tree for a given mask.
1275  * If 'filled' == 1, this returns 1 only if every extent in the tree
1276  * has the bits set.  Otherwise, 1 is returned if any bit in the
1277  * range is found set.
1278  */
1279 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1280                    int bits, int filled)
1281 {
1282         struct extent_state *state = NULL;
1283         struct rb_node *node;
1284         int bitset = 0;
1285         unsigned long flags;
1286
1287         spin_lock_irqsave(&tree->lock, flags);
1288         node = tree_search(tree, start);
1289         while (node && start <= end) {
1290                 state = rb_entry(node, struct extent_state, rb_node);
1291
1292                 if (filled && state->start > start) {
1293                         bitset = 0;
1294                         break;
1295                 }
1296
1297                 if (state->start > end)
1298                         break;
1299
1300                 if (state->state & bits) {
1301                         bitset = 1;
1302                         if (!filled)
1303                                 break;
1304                 } else if (filled) {
1305                         bitset = 0;
1306                         break;
1307                 }
1308                 start = state->end + 1;
1309                 if (start > end)
1310                         break;
1311                 node = rb_next(node);
1312                 if (!node) {
1313                         if (filled)
1314                                 bitset = 0;
1315                         break;
1316                 }
1317         }
1318         spin_unlock_irqrestore(&tree->lock, flags);
1319         return bitset;
1320 }
1321 EXPORT_SYMBOL(test_range_bit);
1322
1323 /*
1324  * helper function to set a given page up to date if all the
1325  * extents in the tree for that page are up to date
1326  */
1327 static int check_page_uptodate(struct extent_io_tree *tree,
1328                                struct page *page)
1329 {
1330         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1331         u64 end = start + PAGE_CACHE_SIZE - 1;
1332         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1))
1333                 SetPageUptodate(page);
1334         return 0;
1335 }
1336
1337 /*
1338  * helper function to unlock a page if all the extents in the tree
1339  * for that page are unlocked
1340  */
1341 static int check_page_locked(struct extent_io_tree *tree,
1342                              struct page *page)
1343 {
1344         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1345         u64 end = start + PAGE_CACHE_SIZE - 1;
1346         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0))
1347                 unlock_page(page);
1348         return 0;
1349 }
1350
1351 /*
1352  * helper function to end page writeback if all the extents
1353  * in the tree for that page are done with writeback
1354  */
1355 static int check_page_writeback(struct extent_io_tree *tree,
1356                              struct page *page)
1357 {
1358         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1359         u64 end = start + PAGE_CACHE_SIZE - 1;
1360         if (!test_range_bit(tree, start, end, EXTENT_WRITEBACK, 0))
1361                 end_page_writeback(page);
1362         return 0;
1363 }
1364
1365 /* lots and lots of room for performance fixes in the end_bio funcs */
1366
1367 /*
1368  * after a writepage IO is done, we need to:
1369  * clear the uptodate bits on error
1370  * clear the writeback bits in the extent tree for this IO
1371  * end_page_writeback if the page has no more pending IO
1372  *
1373  * Scheduling is not allowed, so the extent state tree is expected
1374  * to have one and only one object corresponding to this IO.
1375  */
1376 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1377 static void end_bio_extent_writepage(struct bio *bio, int err)
1378 #else
1379 static int end_bio_extent_writepage(struct bio *bio,
1380                                    unsigned int bytes_done, int err)
1381 #endif
1382 {
1383         int uptodate = err == 0;
1384         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1385         struct extent_state *state = bio->bi_private;
1386         struct extent_io_tree *tree = state->tree;
1387         struct rb_node *node;
1388         u64 start;
1389         u64 end;
1390         u64 cur;
1391         int whole_page;
1392         int ret;
1393         unsigned long flags;
1394
1395 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1396         if (bio->bi_size)
1397                 return 1;
1398 #endif
1399         do {
1400                 struct page *page = bvec->bv_page;
1401                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1402                          bvec->bv_offset;
1403                 end = start + bvec->bv_len - 1;
1404
1405                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1406                         whole_page = 1;
1407                 else
1408                         whole_page = 0;
1409
1410                 if (--bvec >= bio->bi_io_vec)
1411                         prefetchw(&bvec->bv_page->flags);
1412                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1413                         ret = tree->ops->writepage_end_io_hook(page, start,
1414                                                        end, state, uptodate);
1415                         if (ret)
1416                                 uptodate = 0;
1417                 }
1418
1419                 if (!uptodate && tree->ops &&
1420                     tree->ops->writepage_io_failed_hook) {
1421                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1422                                                          start, end, state);
1423                         if (ret == 0) {
1424                                 state = NULL;
1425                                 uptodate = (err == 0);
1426                                 continue;
1427                         }
1428                 }
1429
1430                 if (!uptodate) {
1431                         clear_extent_uptodate(tree, start, end, GFP_ATOMIC);
1432                         ClearPageUptodate(page);
1433                         SetPageError(page);
1434                 }
1435
1436                 /*
1437                  * bios can get merged in funny ways, and so we need to
1438                  * be careful with the state variable.  We know the
1439                  * state won't be merged with others because it has
1440                  * WRITEBACK set, but we can't be sure each biovec is
1441                  * sequential in the file.  So, if our cached state
1442                  * doesn't match the expected end, search the tree
1443                  * for the correct one.
1444                  */
1445
1446                 spin_lock_irqsave(&tree->lock, flags);
1447                 if (!state || state->end != end) {
1448                         state = NULL;
1449                         node = __etree_search(tree, start, NULL, NULL);
1450                         if (node) {
1451                                 state = rb_entry(node, struct extent_state,
1452                                                  rb_node);
1453                                 if (state->end != end ||
1454                                     !(state->state & EXTENT_WRITEBACK))
1455                                         state = NULL;
1456                         }
1457                         if (!state) {
1458                                 spin_unlock_irqrestore(&tree->lock, flags);
1459                                 clear_extent_writeback(tree, start,
1460                                                        end, GFP_ATOMIC);
1461                                 goto next_io;
1462                         }
1463                 }
1464                 cur = end;
1465                 while(1) {
1466                         struct extent_state *clear = state;
1467                         cur = state->start;
1468                         node = rb_prev(&state->rb_node);
1469                         if (node) {
1470                                 state = rb_entry(node,
1471                                                  struct extent_state,
1472                                                  rb_node);
1473                         } else {
1474                                 state = NULL;
1475                         }
1476
1477                         clear_state_bit(tree, clear, EXTENT_WRITEBACK,
1478                                         1, 0);
1479                         if (cur == start)
1480                                 break;
1481                         if (cur < start) {
1482                                 WARN_ON(1);
1483                                 break;
1484                         }
1485                         if (!node)
1486                                 break;
1487                 }
1488                 /* before releasing the lock, make sure the next state
1489                  * variable has the expected bits set and corresponds
1490                  * to the correct offsets in the file
1491                  */
1492                 if (state && (state->end + 1 != start ||
1493                     !(state->state & EXTENT_WRITEBACK))) {
1494                         state = NULL;
1495                 }
1496                 spin_unlock_irqrestore(&tree->lock, flags);
1497 next_io:
1498
1499                 if (whole_page)
1500                         end_page_writeback(page);
1501                 else
1502                         check_page_writeback(tree, page);
1503         } while (bvec >= bio->bi_io_vec);
1504         bio_put(bio);
1505 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1506         return 0;
1507 #endif
1508 }
1509
1510 /*
1511  * after a readpage IO is done, we need to:
1512  * clear the uptodate bits on error
1513  * set the uptodate bits if things worked
1514  * set the page up to date if all extents in the tree are uptodate
1515  * clear the lock bit in the extent tree
1516  * unlock the page if there are no other extents locked for it
1517  *
1518  * Scheduling is not allowed, so the extent state tree is expected
1519  * to have one and only one object corresponding to this IO.
1520  */
1521 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1522 static void end_bio_extent_readpage(struct bio *bio, int err)
1523 #else
1524 static int end_bio_extent_readpage(struct bio *bio,
1525                                    unsigned int bytes_done, int err)
1526 #endif
1527 {
1528         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1529         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1530         struct extent_state *state = bio->bi_private;
1531         struct extent_io_tree *tree = state->tree;
1532         struct rb_node *node;
1533         u64 start;
1534         u64 end;
1535         u64 cur;
1536         unsigned long flags;
1537         int whole_page;
1538         int ret;
1539
1540 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1541         if (bio->bi_size)
1542                 return 1;
1543 #endif
1544
1545         do {
1546                 struct page *page = bvec->bv_page;
1547                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1548                         bvec->bv_offset;
1549                 end = start + bvec->bv_len - 1;
1550
1551                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1552                         whole_page = 1;
1553                 else
1554                         whole_page = 0;
1555
1556                 if (--bvec >= bio->bi_io_vec)
1557                         prefetchw(&bvec->bv_page->flags);
1558
1559                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1560                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1561                                                               state);
1562                         if (ret)
1563                                 uptodate = 0;
1564                 }
1565                 if (!uptodate && tree->ops &&
1566                     tree->ops->readpage_io_failed_hook) {
1567                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1568                                                          start, end, state);
1569                         if (ret == 0) {
1570                                 state = NULL;
1571                                 uptodate =
1572                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1573                                 continue;
1574                         }
1575                 }
1576
1577                 spin_lock_irqsave(&tree->lock, flags);
1578                 if (!state || state->end != end) {
1579                         state = NULL;
1580                         node = __etree_search(tree, start, NULL, NULL);
1581                         if (node) {
1582                                 state = rb_entry(node, struct extent_state,
1583                                                  rb_node);
1584                                 if (state->end != end ||
1585                                     !(state->state & EXTENT_LOCKED))
1586                                         state = NULL;
1587                         }
1588                         if (!state) {
1589                                 spin_unlock_irqrestore(&tree->lock, flags);
1590                                 if (uptodate)
1591                                         set_extent_uptodate(tree, start, end,
1592                                                             GFP_ATOMIC);
1593                                 unlock_extent(tree, start, end, GFP_ATOMIC);
1594                                 goto next_io;
1595                         }
1596                 }
1597
1598                 cur = end;
1599                 while(1) {
1600                         struct extent_state *clear = state;
1601                         cur = state->start;
1602                         node = rb_prev(&state->rb_node);
1603                         if (node) {
1604                                 state = rb_entry(node,
1605                                          struct extent_state,
1606                                          rb_node);
1607                         } else {
1608                                 state = NULL;
1609                         }
1610                         if (uptodate) {
1611                                 set_state_cb(tree, clear, EXTENT_UPTODATE);
1612                                 clear->state |= EXTENT_UPTODATE;
1613                         }
1614                         clear_state_bit(tree, clear, EXTENT_LOCKED,
1615                                         1, 0);
1616                         if (cur == start)
1617                                 break;
1618                         if (cur < start) {
1619                                 WARN_ON(1);
1620                                 break;
1621                         }
1622                         if (!node)
1623                                 break;
1624                 }
1625                 /* before releasing the lock, make sure the next state
1626                  * variable has the expected bits set and corresponds
1627                  * to the correct offsets in the file
1628                  */
1629                 if (state && (state->end + 1 != start ||
1630                     !(state->state & EXTENT_LOCKED))) {
1631                         state = NULL;
1632                 }
1633                 spin_unlock_irqrestore(&tree->lock, flags);
1634 next_io:
1635                 if (whole_page) {
1636                         if (uptodate) {
1637                                 SetPageUptodate(page);
1638                         } else {
1639                                 ClearPageUptodate(page);
1640                                 SetPageError(page);
1641                         }
1642                         unlock_page(page);
1643                 } else {
1644                         if (uptodate) {
1645                                 check_page_uptodate(tree, page);
1646                         } else {
1647                                 ClearPageUptodate(page);
1648                                 SetPageError(page);
1649                         }
1650                         check_page_locked(tree, page);
1651                 }
1652         } while (bvec >= bio->bi_io_vec);
1653
1654         bio_put(bio);
1655 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1656         return 0;
1657 #endif
1658 }
1659
1660 /*
1661  * IO done from prepare_write is pretty simple, we just unlock
1662  * the structs in the extent tree when done, and set the uptodate bits
1663  * as appropriate.
1664  */
1665 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,23)
1666 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1667 #else
1668 static int end_bio_extent_preparewrite(struct bio *bio,
1669                                        unsigned int bytes_done, int err)
1670 #endif
1671 {
1672         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1673         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1674         struct extent_state *state = bio->bi_private;
1675         struct extent_io_tree *tree = state->tree;
1676         u64 start;
1677         u64 end;
1678
1679 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1680         if (bio->bi_size)
1681                 return 1;
1682 #endif
1683
1684         do {
1685                 struct page *page = bvec->bv_page;
1686                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1687                         bvec->bv_offset;
1688                 end = start + bvec->bv_len - 1;
1689
1690                 if (--bvec >= bio->bi_io_vec)
1691                         prefetchw(&bvec->bv_page->flags);
1692
1693                 if (uptodate) {
1694                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1695                 } else {
1696                         ClearPageUptodate(page);
1697                         SetPageError(page);
1698                 }
1699
1700                 unlock_extent(tree, start, end, GFP_ATOMIC);
1701
1702         } while (bvec >= bio->bi_io_vec);
1703
1704         bio_put(bio);
1705 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
1706         return 0;
1707 #endif
1708 }
1709
1710 static struct bio *
1711 extent_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1712                  gfp_t gfp_flags)
1713 {
1714         struct bio *bio;
1715
1716         bio = bio_alloc(gfp_flags, nr_vecs);
1717
1718         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1719                 while (!bio && (nr_vecs /= 2))
1720                         bio = bio_alloc(gfp_flags, nr_vecs);
1721         }
1722
1723         if (bio) {
1724                 bio->bi_size = 0;
1725                 bio->bi_bdev = bdev;
1726                 bio->bi_sector = first_sector;
1727         }
1728         return bio;
1729 }
1730
1731 static int submit_one_bio(int rw, struct bio *bio, int mirror_num)
1732 {
1733         int ret = 0;
1734         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1735         struct page *page = bvec->bv_page;
1736         struct extent_io_tree *tree = bio->bi_private;
1737         struct rb_node *node;
1738         struct extent_state *state;
1739         u64 start;
1740         u64 end;
1741
1742         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1743         end = start + bvec->bv_len - 1;
1744
1745         spin_lock_irq(&tree->lock);
1746         node = __etree_search(tree, start, NULL, NULL);
1747         BUG_ON(!node);
1748         state = rb_entry(node, struct extent_state, rb_node);
1749         while(state->end < end) {
1750                 node = rb_next(node);
1751                 state = rb_entry(node, struct extent_state, rb_node);
1752         }
1753         BUG_ON(state->end != end);
1754         spin_unlock_irq(&tree->lock);
1755
1756         bio->bi_private = state;
1757
1758         bio_get(bio);
1759
1760         if (tree->ops && tree->ops->submit_bio_hook)
1761                 tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1762                                            mirror_num);
1763         else
1764                 submit_bio(rw, bio);
1765         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1766                 ret = -EOPNOTSUPP;
1767         bio_put(bio);
1768         return ret;
1769 }
1770
1771 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1772                               struct page *page, sector_t sector,
1773                               size_t size, unsigned long offset,
1774                               struct block_device *bdev,
1775                               struct bio **bio_ret,
1776                               unsigned long max_pages,
1777                               bio_end_io_t end_io_func,
1778                               int mirror_num)
1779 {
1780         int ret = 0;
1781         struct bio *bio;
1782         int nr;
1783
1784         if (bio_ret && *bio_ret) {
1785                 bio = *bio_ret;
1786                 if (bio->bi_sector + (bio->bi_size >> 9) != sector ||
1787                     (tree->ops && tree->ops->merge_bio_hook &&
1788                      tree->ops->merge_bio_hook(page, offset, size, bio)) ||
1789                     bio_add_page(bio, page, size, offset) < size) {
1790                         ret = submit_one_bio(rw, bio, mirror_num);
1791                         bio = NULL;
1792                 } else {
1793                         return 0;
1794                 }
1795         }
1796         nr = bio_get_nr_vecs(bdev);
1797         bio = extent_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1798         if (!bio) {
1799                 printk("failed to allocate bio nr %d\n", nr);
1800         }
1801
1802
1803         bio_add_page(bio, page, size, offset);
1804         bio->bi_end_io = end_io_func;
1805         bio->bi_private = tree;
1806
1807         if (bio_ret) {
1808                 *bio_ret = bio;
1809         } else {
1810                 ret = submit_one_bio(rw, bio, mirror_num);
1811         }
1812
1813         return ret;
1814 }
1815
1816 void set_page_extent_mapped(struct page *page)
1817 {
1818         if (!PagePrivate(page)) {
1819                 SetPagePrivate(page);
1820                 WARN_ON(!page->mapping->a_ops->invalidatepage);
1821                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1822                 page_cache_get(page);
1823         }
1824 }
1825
1826 void set_page_extent_head(struct page *page, unsigned long len)
1827 {
1828         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1829 }
1830
1831 /*
1832  * basic readpage implementation.  Locked extent state structs are inserted
1833  * into the tree that are removed when the IO is done (by the end_io
1834  * handlers)
1835  */
1836 static int __extent_read_full_page(struct extent_io_tree *tree,
1837                                    struct page *page,
1838                                    get_extent_t *get_extent,
1839                                    struct bio **bio, int mirror_num)
1840 {
1841         struct inode *inode = page->mapping->host;
1842         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1843         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1844         u64 end;
1845         u64 cur = start;
1846         u64 extent_offset;
1847         u64 last_byte = i_size_read(inode);
1848         u64 block_start;
1849         u64 cur_end;
1850         sector_t sector;
1851         struct extent_map *em;
1852         struct block_device *bdev;
1853         int ret;
1854         int nr = 0;
1855         size_t page_offset = 0;
1856         size_t iosize;
1857         size_t blocksize = inode->i_sb->s_blocksize;
1858
1859         set_page_extent_mapped(page);
1860
1861         end = page_end;
1862         lock_extent(tree, start, end, GFP_NOFS);
1863
1864         while (cur <= end) {
1865                 if (cur >= last_byte) {
1866                         char *userpage;
1867                         iosize = PAGE_CACHE_SIZE - page_offset;
1868                         userpage = kmap_atomic(page, KM_USER0);
1869                         memset(userpage + page_offset, 0, iosize);
1870                         flush_dcache_page(page);
1871                         kunmap_atomic(userpage, KM_USER0);
1872                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1873                                             GFP_NOFS);
1874                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1875                         break;
1876                 }
1877                 em = get_extent(inode, page, page_offset, cur,
1878                                 end - cur + 1, 0);
1879                 if (IS_ERR(em) || !em) {
1880                         SetPageError(page);
1881                         unlock_extent(tree, cur, end, GFP_NOFS);
1882                         break;
1883                 }
1884                 extent_offset = cur - em->start;
1885                 if (extent_map_end(em) <= cur) {
1886 printk("bad mapping em [%Lu %Lu] cur %Lu\n", em->start, extent_map_end(em), cur);
1887                 }
1888                 BUG_ON(extent_map_end(em) <= cur);
1889                 if (end < cur) {
1890 printk("2bad mapping end %Lu cur %Lu\n", end, cur);
1891                 }
1892                 BUG_ON(end < cur);
1893
1894                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1895                 cur_end = min(extent_map_end(em) - 1, end);
1896                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1897                 sector = (em->block_start + extent_offset) >> 9;
1898                 bdev = em->bdev;
1899                 block_start = em->block_start;
1900                 free_extent_map(em);
1901                 em = NULL;
1902
1903                 /* we've found a hole, just zero and go on */
1904                 if (block_start == EXTENT_MAP_HOLE) {
1905                         char *userpage;
1906                         userpage = kmap_atomic(page, KM_USER0);
1907                         memset(userpage + page_offset, 0, iosize);
1908                         flush_dcache_page(page);
1909                         kunmap_atomic(userpage, KM_USER0);
1910
1911                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1912                                             GFP_NOFS);
1913                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1914                         cur = cur + iosize;
1915                         page_offset += iosize;
1916                         continue;
1917                 }
1918                 /* the get_extent function already copied into the page */
1919                 if (test_range_bit(tree, cur, cur_end, EXTENT_UPTODATE, 1)) {
1920                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1921                         cur = cur + iosize;
1922                         page_offset += iosize;
1923                         continue;
1924                 }
1925                 /* we have an inline extent but it didn't get marked up
1926                  * to date.  Error out
1927                  */
1928                 if (block_start == EXTENT_MAP_INLINE) {
1929                         SetPageError(page);
1930                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1931                         cur = cur + iosize;
1932                         page_offset += iosize;
1933                         continue;
1934                 }
1935
1936                 ret = 0;
1937                 if (tree->ops && tree->ops->readpage_io_hook) {
1938                         ret = tree->ops->readpage_io_hook(page, cur,
1939                                                           cur + iosize - 1);
1940                 }
1941                 if (!ret) {
1942                         unsigned long nr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
1943                         nr -= page->index;
1944                         ret = submit_extent_page(READ, tree, page,
1945                                          sector, iosize, page_offset,
1946                                          bdev, bio, nr,
1947                                          end_bio_extent_readpage, mirror_num);
1948                 }
1949                 if (ret)
1950                         SetPageError(page);
1951                 cur = cur + iosize;
1952                 page_offset += iosize;
1953                 nr++;
1954         }
1955         if (!nr) {
1956                 if (!PageError(page))
1957                         SetPageUptodate(page);
1958                 unlock_page(page);
1959         }
1960         return 0;
1961 }
1962
1963 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
1964                             get_extent_t *get_extent)
1965 {
1966         struct bio *bio = NULL;
1967         int ret;
1968
1969         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0);
1970         if (bio)
1971                 submit_one_bio(READ, bio, 0);
1972         return ret;
1973 }
1974 EXPORT_SYMBOL(extent_read_full_page);
1975
1976 /*
1977  * the writepage semantics are similar to regular writepage.  extent
1978  * records are inserted to lock ranges in the tree, and as dirty areas
1979  * are found, they are marked writeback.  Then the lock bits are removed
1980  * and the end_io handler clears the writeback ranges
1981  */
1982 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
1983                               void *data)
1984 {
1985         struct inode *inode = page->mapping->host;
1986         struct extent_page_data *epd = data;
1987         struct extent_io_tree *tree = epd->tree;
1988         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1989         u64 delalloc_start;
1990         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1991         u64 end;
1992         u64 cur = start;
1993         u64 extent_offset;
1994         u64 last_byte = i_size_read(inode);
1995         u64 block_start;
1996         u64 iosize;
1997         u64 unlock_start;
1998         sector_t sector;
1999         struct extent_map *em;
2000         struct block_device *bdev;
2001         int ret;
2002         int nr = 0;
2003         size_t pg_offset = 0;
2004         size_t blocksize;
2005         loff_t i_size = i_size_read(inode);
2006         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2007         u64 nr_delalloc;
2008         u64 delalloc_end;
2009
2010         WARN_ON(!PageLocked(page));
2011         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2012         if (page->index > end_index ||
2013            (page->index == end_index && !pg_offset)) {
2014                 page->mapping->a_ops->invalidatepage(page, 0);
2015                 unlock_page(page);
2016                 return 0;
2017         }
2018
2019         if (page->index == end_index) {
2020                 char *userpage;
2021
2022                 userpage = kmap_atomic(page, KM_USER0);
2023                 memset(userpage + pg_offset, 0,
2024                        PAGE_CACHE_SIZE - pg_offset);
2025                 kunmap_atomic(userpage, KM_USER0);
2026                 flush_dcache_page(page);
2027         }
2028         pg_offset = 0;
2029
2030         set_page_extent_mapped(page);
2031
2032         delalloc_start = start;
2033         delalloc_end = 0;
2034         while(delalloc_end < page_end) {
2035                 nr_delalloc = find_lock_delalloc_range(tree, &delalloc_start,
2036                                                        &delalloc_end,
2037                                                        128 * 1024 * 1024);
2038                 if (nr_delalloc == 0) {
2039                         delalloc_start = delalloc_end + 1;
2040                         continue;
2041                 }
2042                 tree->ops->fill_delalloc(inode, delalloc_start,
2043                                          delalloc_end);
2044                 clear_extent_bit(tree, delalloc_start,
2045                                  delalloc_end,
2046                                  EXTENT_LOCKED | EXTENT_DELALLOC,
2047                                  1, 0, GFP_NOFS);
2048                 delalloc_start = delalloc_end + 1;
2049         }
2050         lock_extent(tree, start, page_end, GFP_NOFS);
2051         unlock_start = start;
2052
2053         if (tree->ops && tree->ops->writepage_start_hook) {
2054                 ret = tree->ops->writepage_start_hook(page, start, page_end);
2055                 if (ret == -EAGAIN) {
2056                         unlock_extent(tree, start, page_end, GFP_NOFS);
2057                         redirty_page_for_writepage(wbc, page);
2058                         unlock_page(page);
2059                         return 0;
2060                 }
2061         }
2062
2063         end = page_end;
2064         if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0)) {
2065                 printk("found delalloc bits after lock_extent\n");
2066         }
2067
2068         if (last_byte <= start) {
2069                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
2070                 unlock_extent(tree, start, page_end, GFP_NOFS);
2071                 if (tree->ops && tree->ops->writepage_end_io_hook)
2072                         tree->ops->writepage_end_io_hook(page, start,
2073                                                          page_end, NULL, 1);
2074                 unlock_start = page_end + 1;
2075                 goto done;
2076         }
2077
2078         set_extent_uptodate(tree, start, page_end, GFP_NOFS);
2079         blocksize = inode->i_sb->s_blocksize;
2080
2081         while (cur <= end) {
2082                 if (cur >= last_byte) {
2083                         clear_extent_dirty(tree, cur, page_end, GFP_NOFS);
2084                         unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
2085                         if (tree->ops && tree->ops->writepage_end_io_hook)
2086                                 tree->ops->writepage_end_io_hook(page, cur,
2087                                                          page_end, NULL, 1);
2088                         unlock_start = page_end + 1;
2089                         break;
2090                 }
2091                 em = epd->get_extent(inode, page, pg_offset, cur,
2092                                      end - cur + 1, 1);
2093                 if (IS_ERR(em) || !em) {
2094                         SetPageError(page);
2095                         break;
2096                 }
2097
2098                 extent_offset = cur - em->start;
2099                 BUG_ON(extent_map_end(em) <= cur);
2100                 BUG_ON(end < cur);
2101                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2102                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2103                 sector = (em->block_start + extent_offset) >> 9;
2104                 bdev = em->bdev;
2105                 block_start = em->block_start;
2106                 free_extent_map(em);
2107                 em = NULL;
2108
2109                 if (block_start == EXTENT_MAP_HOLE ||
2110                     block_start == EXTENT_MAP_INLINE) {
2111                         clear_extent_dirty(tree, cur,
2112                                            cur + iosize - 1, GFP_NOFS);
2113
2114                         unlock_extent(tree, unlock_start, cur + iosize -1,
2115                                       GFP_NOFS);
2116
2117                         if (tree->ops && tree->ops->writepage_end_io_hook)
2118                                 tree->ops->writepage_end_io_hook(page, cur,
2119                                                          cur + iosize - 1,
2120                                                          NULL, 1);
2121                         cur = cur + iosize;
2122                         pg_offset += iosize;
2123                         unlock_start = cur;
2124                         continue;
2125                 }
2126
2127                 /* leave this out until we have a page_mkwrite call */
2128                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2129                                    EXTENT_DIRTY, 0)) {
2130                         cur = cur + iosize;
2131                         pg_offset += iosize;
2132                         continue;
2133                 }
2134                 clear_extent_dirty(tree, cur, cur + iosize - 1, GFP_NOFS);
2135                 if (tree->ops && tree->ops->writepage_io_hook) {
2136                         ret = tree->ops->writepage_io_hook(page, cur,
2137                                                 cur + iosize - 1);
2138                 } else {
2139                         ret = 0;
2140                 }
2141                 if (ret) {
2142                         SetPageError(page);
2143                 } else {
2144                         unsigned long max_nr = end_index + 1;
2145
2146                         set_range_writeback(tree, cur, cur + iosize - 1);
2147                         if (!PageWriteback(page)) {
2148                                 printk("warning page %lu not writeback, "
2149                                        "cur %llu end %llu\n", page->index,
2150                                        (unsigned long long)cur,
2151                                        (unsigned long long)end);
2152                         }
2153
2154                         ret = submit_extent_page(WRITE, tree, page, sector,
2155                                                  iosize, pg_offset, bdev,
2156                                                  &epd->bio, max_nr,
2157                                                  end_bio_extent_writepage, 0);
2158                         if (ret)
2159                                 SetPageError(page);
2160                 }
2161                 cur = cur + iosize;
2162                 pg_offset += iosize;
2163                 nr++;
2164         }
2165 done:
2166         if (nr == 0) {
2167                 /* make sure the mapping tag for page dirty gets cleared */
2168                 set_page_writeback(page);
2169                 end_page_writeback(page);
2170         }
2171         if (unlock_start <= page_end)
2172                 unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
2173         unlock_page(page);
2174         return 0;
2175 }
2176
2177 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,22)
2178 /* Taken directly from 2.6.23 for 2.6.18 back port */
2179 typedef int (*writepage_t)(struct page *page, struct writeback_control *wbc,
2180                                 void *data);
2181
2182 /**
2183  * write_cache_pages - walk the list of dirty pages of the given address space
2184  * and write all of them.
2185  * @mapping: address space structure to write
2186  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2187  * @writepage: function called for each page
2188  * @data: data passed to writepage function
2189  *
2190  * If a page is already under I/O, write_cache_pages() skips it, even
2191  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2192  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2193  * and msync() need to guarantee that all the data which was dirty at the time
2194  * the call was made get new I/O started against them.  If wbc->sync_mode is
2195  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2196  * existing IO to complete.
2197  */
2198 static int write_cache_pages(struct address_space *mapping,
2199                       struct writeback_control *wbc, writepage_t writepage,
2200                       void *data)
2201 {
2202         struct backing_dev_info *bdi = mapping->backing_dev_info;
2203         int ret = 0;
2204         int done = 0;
2205         struct pagevec pvec;
2206         int nr_pages;
2207         pgoff_t index;
2208         pgoff_t end;            /* Inclusive */
2209         int scanned = 0;
2210         int range_whole = 0;
2211
2212         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2213                 wbc->encountered_congestion = 1;
2214                 return 0;
2215         }
2216
2217         pagevec_init(&pvec, 0);
2218         if (wbc->range_cyclic) {
2219                 index = mapping->writeback_index; /* Start from prev offset */
2220                 end = -1;
2221         } else {
2222                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2223                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2224                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2225                         range_whole = 1;
2226                 scanned = 1;
2227         }
2228 retry:
2229         while (!done && (index <= end) &&
2230                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2231                                               PAGECACHE_TAG_DIRTY,
2232                                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2233                 unsigned i;
2234
2235                 scanned = 1;
2236                 for (i = 0; i < nr_pages; i++) {
2237                         struct page *page = pvec.pages[i];
2238
2239                         /*
2240                          * At this point we hold neither mapping->tree_lock nor
2241                          * lock on the page itself: the page may be truncated or
2242                          * invalidated (changing page->mapping to NULL), or even
2243                          * swizzled back from swapper_space to tmpfs file
2244                          * mapping
2245                          */
2246                         lock_page(page);
2247
2248                         if (unlikely(page->mapping != mapping)) {
2249                                 unlock_page(page);
2250                                 continue;
2251                         }
2252
2253                         if (!wbc->range_cyclic && page->index > end) {
2254                                 done = 1;
2255                                 unlock_page(page);
2256                                 continue;
2257                         }
2258
2259                         if (wbc->sync_mode != WB_SYNC_NONE)
2260                                 wait_on_page_writeback(page);
2261
2262                         if (PageWriteback(page) ||
2263                             !clear_page_dirty_for_io(page)) {
2264                                 unlock_page(page);
2265                                 continue;
2266                         }
2267
2268                         ret = (*writepage)(page, wbc, data);
2269
2270                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2271                                 unlock_page(page);
2272                                 ret = 0;
2273                         }
2274                         if (ret || (--(wbc->nr_to_write) <= 0))
2275                                 done = 1;
2276                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2277                                 wbc->encountered_congestion = 1;
2278                                 done = 1;
2279                         }
2280                 }
2281                 pagevec_release(&pvec);
2282                 cond_resched();
2283         }
2284         if (!scanned && !done) {
2285                 /*
2286                  * We hit the last page and there is more work to be done: wrap
2287                  * back to the start of the file
2288                  */
2289                 scanned = 1;
2290                 index = 0;
2291                 goto retry;
2292         }
2293         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2294                 mapping->writeback_index = index;
2295         return ret;
2296 }
2297 #endif
2298
2299 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2300                           get_extent_t *get_extent,
2301                           struct writeback_control *wbc)
2302 {
2303         int ret;
2304         struct address_space *mapping = page->mapping;
2305         struct extent_page_data epd = {
2306                 .bio = NULL,
2307                 .tree = tree,
2308                 .get_extent = get_extent,
2309         };
2310         struct writeback_control wbc_writepages = {
2311                 .bdi            = wbc->bdi,
2312                 .sync_mode      = WB_SYNC_NONE,
2313                 .older_than_this = NULL,
2314                 .nr_to_write    = 64,
2315                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2316                 .range_end      = (loff_t)-1,
2317         };
2318
2319
2320         ret = __extent_writepage(page, wbc, &epd);
2321
2322         write_cache_pages(mapping, &wbc_writepages, __extent_writepage, &epd);
2323         if (epd.bio) {
2324                 submit_one_bio(WRITE, epd.bio, 0);
2325         }
2326         return ret;
2327 }
2328 EXPORT_SYMBOL(extent_write_full_page);
2329
2330
2331 int extent_writepages(struct extent_io_tree *tree,
2332                       struct address_space *mapping,
2333                       get_extent_t *get_extent,
2334                       struct writeback_control *wbc)
2335 {
2336         int ret = 0;
2337         struct extent_page_data epd = {
2338                 .bio = NULL,
2339                 .tree = tree,
2340                 .get_extent = get_extent,
2341         };
2342
2343         ret = write_cache_pages(mapping, wbc, __extent_writepage, &epd);
2344         if (epd.bio) {
2345                 submit_one_bio(WRITE, epd.bio, 0);
2346         }
2347         return ret;
2348 }
2349 EXPORT_SYMBOL(extent_writepages);
2350
2351 int extent_readpages(struct extent_io_tree *tree,
2352                      struct address_space *mapping,
2353                      struct list_head *pages, unsigned nr_pages,
2354                      get_extent_t get_extent)
2355 {
2356         struct bio *bio = NULL;
2357         unsigned page_idx;
2358         struct pagevec pvec;
2359
2360         pagevec_init(&pvec, 0);
2361         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2362                 struct page *page = list_entry(pages->prev, struct page, lru);
2363
2364                 prefetchw(&page->flags);
2365                 list_del(&page->lru);
2366                 /*
2367                  * what we want to do here is call add_to_page_cache_lru,
2368                  * but that isn't exported, so we reproduce it here
2369                  */
2370                 if (!add_to_page_cache(page, mapping,
2371                                         page->index, GFP_KERNEL)) {
2372
2373                         /* open coding of lru_cache_add, also not exported */
2374                         page_cache_get(page);
2375                         if (!pagevec_add(&pvec, page))
2376                                 __pagevec_lru_add(&pvec);
2377                         __extent_read_full_page(tree, page, get_extent,
2378                                                 &bio, 0);
2379                 }
2380                 page_cache_release(page);
2381         }
2382         if (pagevec_count(&pvec))
2383                 __pagevec_lru_add(&pvec);
2384         BUG_ON(!list_empty(pages));
2385         if (bio)
2386                 submit_one_bio(READ, bio, 0);
2387         return 0;
2388 }
2389 EXPORT_SYMBOL(extent_readpages);
2390
2391 /*
2392  * basic invalidatepage code, this waits on any locked or writeback
2393  * ranges corresponding to the page, and then deletes any extent state
2394  * records from the tree
2395  */
2396 int extent_invalidatepage(struct extent_io_tree *tree,
2397                           struct page *page, unsigned long offset)
2398 {
2399         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2400         u64 end = start + PAGE_CACHE_SIZE - 1;
2401         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2402
2403         start += (offset + blocksize -1) & ~(blocksize - 1);
2404         if (start > end)
2405                 return 0;
2406
2407         lock_extent(tree, start, end, GFP_NOFS);
2408         wait_on_extent_writeback(tree, start, end);
2409         clear_extent_bit(tree, start, end,
2410                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC,
2411                          1, 1, GFP_NOFS);
2412         return 0;
2413 }
2414 EXPORT_SYMBOL(extent_invalidatepage);
2415
2416 /*
2417  * simple commit_write call, set_range_dirty is used to mark both
2418  * the pages and the extent records as dirty
2419  */
2420 int extent_commit_write(struct extent_io_tree *tree,
2421                         struct inode *inode, struct page *page,
2422                         unsigned from, unsigned to)
2423 {
2424         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2425
2426         set_page_extent_mapped(page);
2427         set_page_dirty(page);
2428
2429         if (pos > inode->i_size) {
2430                 i_size_write(inode, pos);
2431                 mark_inode_dirty(inode);
2432         }
2433         return 0;
2434 }
2435 EXPORT_SYMBOL(extent_commit_write);
2436
2437 int extent_prepare_write(struct extent_io_tree *tree,
2438                          struct inode *inode, struct page *page,
2439                          unsigned from, unsigned to, get_extent_t *get_extent)
2440 {
2441         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2442         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2443         u64 block_start;
2444         u64 orig_block_start;
2445         u64 block_end;
2446         u64 cur_end;
2447         struct extent_map *em;
2448         unsigned blocksize = 1 << inode->i_blkbits;
2449         size_t page_offset = 0;
2450         size_t block_off_start;
2451         size_t block_off_end;
2452         int err = 0;
2453         int iocount = 0;
2454         int ret = 0;
2455         int isnew;
2456
2457         set_page_extent_mapped(page);
2458
2459         block_start = (page_start + from) & ~((u64)blocksize - 1);
2460         block_end = (page_start + to - 1) | (blocksize - 1);
2461         orig_block_start = block_start;
2462
2463         lock_extent(tree, page_start, page_end, GFP_NOFS);
2464         while(block_start <= block_end) {
2465                 em = get_extent(inode, page, page_offset, block_start,
2466                                 block_end - block_start + 1, 1);
2467                 if (IS_ERR(em) || !em) {
2468                         goto err;
2469                 }
2470                 cur_end = min(block_end, extent_map_end(em) - 1);
2471                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2472                 block_off_end = block_off_start + blocksize;
2473                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2474
2475                 if (!PageUptodate(page) && isnew &&
2476                     (block_off_end > to || block_off_start < from)) {
2477                         void *kaddr;
2478
2479                         kaddr = kmap_atomic(page, KM_USER0);
2480                         if (block_off_end > to)
2481                                 memset(kaddr + to, 0, block_off_end - to);
2482                         if (block_off_start < from)
2483                                 memset(kaddr + block_off_start, 0,
2484                                        from - block_off_start);
2485                         flush_dcache_page(page);
2486                         kunmap_atomic(kaddr, KM_USER0);
2487                 }
2488                 if ((em->block_start != EXTENT_MAP_HOLE &&
2489                      em->block_start != EXTENT_MAP_INLINE) &&
2490                     !isnew && !PageUptodate(page) &&
2491                     (block_off_end > to || block_off_start < from) &&
2492                     !test_range_bit(tree, block_start, cur_end,
2493                                     EXTENT_UPTODATE, 1)) {
2494                         u64 sector;
2495                         u64 extent_offset = block_start - em->start;
2496                         size_t iosize;
2497                         sector = (em->block_start + extent_offset) >> 9;
2498                         iosize = (cur_end - block_start + blocksize) &
2499                                 ~((u64)blocksize - 1);
2500                         /*
2501                          * we've already got the extent locked, but we
2502                          * need to split the state such that our end_bio
2503                          * handler can clear the lock.
2504                          */
2505                         set_extent_bit(tree, block_start,
2506                                        block_start + iosize - 1,
2507                                        EXTENT_LOCKED, 0, NULL, GFP_NOFS);
2508                         ret = submit_extent_page(READ, tree, page,
2509                                          sector, iosize, page_offset, em->bdev,
2510                                          NULL, 1,
2511                                          end_bio_extent_preparewrite, 0);
2512                         iocount++;
2513                         block_start = block_start + iosize;
2514                 } else {
2515                         set_extent_uptodate(tree, block_start, cur_end,
2516                                             GFP_NOFS);
2517                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2518                         block_start = cur_end + 1;
2519                 }
2520                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2521                 free_extent_map(em);
2522         }
2523         if (iocount) {
2524                 wait_extent_bit(tree, orig_block_start,
2525                                 block_end, EXTENT_LOCKED);
2526         }
2527         check_page_uptodate(tree, page);
2528 err:
2529         /* FIXME, zero out newly allocated blocks on error */
2530         return err;
2531 }
2532 EXPORT_SYMBOL(extent_prepare_write);
2533
2534 /*
2535  * a helper for releasepage, this tests for areas of the page that
2536  * are locked or under IO and drops the related state bits if it is safe
2537  * to drop the page.
2538  */
2539 int try_release_extent_state(struct extent_map_tree *map,
2540                              struct extent_io_tree *tree, struct page *page,
2541                              gfp_t mask)
2542 {
2543         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2544         u64 end = start + PAGE_CACHE_SIZE - 1;
2545         int ret = 1;
2546
2547         if (test_range_bit(tree, start, end,
2548                            EXTENT_IOBITS | EXTENT_ORDERED, 0))
2549                 ret = 0;
2550         else {
2551                 if ((mask & GFP_NOFS) == GFP_NOFS)
2552                         mask = GFP_NOFS;
2553                 clear_extent_bit(tree, start, end, EXTENT_UPTODATE,
2554                                  1, 1, mask);
2555         }
2556         return ret;
2557 }
2558 EXPORT_SYMBOL(try_release_extent_state);
2559
2560 /*
2561  * a helper for releasepage.  As long as there are no locked extents
2562  * in the range corresponding to the page, both state records and extent
2563  * map records are removed
2564  */
2565 int try_release_extent_mapping(struct extent_map_tree *map,
2566                                struct extent_io_tree *tree, struct page *page,
2567                                gfp_t mask)
2568 {
2569         struct extent_map *em;
2570         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2571         u64 end = start + PAGE_CACHE_SIZE - 1;
2572
2573         if ((mask & __GFP_WAIT) &&
2574             page->mapping->host->i_size > 16 * 1024 * 1024) {
2575                 u64 len;
2576                 while (start <= end) {
2577                         len = end - start + 1;
2578                         spin_lock(&map->lock);
2579                         em = lookup_extent_mapping(map, start, len);
2580                         if (!em || IS_ERR(em)) {
2581                                 spin_unlock(&map->lock);
2582                                 break;
2583                         }
2584                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2585                             em->start != start) {
2586                                 spin_unlock(&map->lock);
2587                                 free_extent_map(em);
2588                                 break;
2589                         }
2590                         if (!test_range_bit(tree, em->start,
2591                                             extent_map_end(em) - 1,
2592                                             EXTENT_LOCKED, 0)) {
2593                                 remove_extent_mapping(map, em);
2594                                 /* once for the rb tree */
2595                                 free_extent_map(em);
2596                         }
2597                         start = extent_map_end(em);
2598                         spin_unlock(&map->lock);
2599
2600                         /* once for us */
2601                         free_extent_map(em);
2602                 }
2603         }
2604         return try_release_extent_state(map, tree, page, mask);
2605 }
2606 EXPORT_SYMBOL(try_release_extent_mapping);
2607
2608 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2609                 get_extent_t *get_extent)
2610 {
2611         struct inode *inode = mapping->host;
2612         u64 start = iblock << inode->i_blkbits;
2613         sector_t sector = 0;
2614         struct extent_map *em;
2615
2616         em = get_extent(inode, NULL, 0, start, (1 << inode->i_blkbits), 0);
2617         if (!em || IS_ERR(em))
2618                 return 0;
2619
2620         if (em->block_start == EXTENT_MAP_INLINE ||
2621             em->block_start == EXTENT_MAP_HOLE)
2622                 goto out;
2623
2624         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2625 out:
2626         free_extent_map(em);
2627         return sector;
2628 }
2629
2630 static int add_lru(struct extent_io_tree *tree, struct extent_buffer *eb)
2631 {
2632         if (list_empty(&eb->lru)) {
2633                 extent_buffer_get(eb);
2634                 list_add(&eb->lru, &tree->buffer_lru);
2635                 tree->lru_size++;
2636                 if (tree->lru_size >= BUFFER_LRU_MAX) {
2637                         struct extent_buffer *rm;
2638                         rm = list_entry(tree->buffer_lru.prev,
2639                                         struct extent_buffer, lru);
2640                         tree->lru_size--;
2641                         list_del_init(&rm->lru);
2642                         free_extent_buffer(rm);
2643                 }
2644         } else
2645                 list_move(&eb->lru, &tree->buffer_lru);
2646         return 0;
2647 }
2648 static struct extent_buffer *find_lru(struct extent_io_tree *tree,
2649                                       u64 start, unsigned long len)
2650 {
2651         struct list_head *lru = &tree->buffer_lru;
2652         struct list_head *cur = lru->next;
2653         struct extent_buffer *eb;
2654
2655         if (list_empty(lru))
2656                 return NULL;
2657
2658         do {
2659                 eb = list_entry(cur, struct extent_buffer, lru);
2660                 if (eb->start == start && eb->len == len) {
2661                         extent_buffer_get(eb);
2662                         return eb;
2663                 }
2664                 cur = cur->next;
2665         } while (cur != lru);
2666         return NULL;
2667 }
2668
2669 static inline unsigned long num_extent_pages(u64 start, u64 len)
2670 {
2671         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
2672                 (start >> PAGE_CACHE_SHIFT);
2673 }
2674
2675 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2676                                               unsigned long i)
2677 {
2678         struct page *p;
2679         struct address_space *mapping;
2680
2681         if (i == 0)
2682                 return eb->first_page;
2683         i += eb->start >> PAGE_CACHE_SHIFT;
2684         mapping = eb->first_page->mapping;
2685         read_lock_irq(&mapping->tree_lock);
2686         p = radix_tree_lookup(&mapping->page_tree, i);
2687         read_unlock_irq(&mapping->tree_lock);
2688         return p;
2689 }
2690
2691 int release_extent_buffer_tail_pages(struct extent_buffer *eb)
2692 {
2693         unsigned long num_pages = num_extent_pages(eb->start, eb->len);
2694         struct page *page;
2695         unsigned long i;
2696
2697         if (num_pages == 1)
2698                 return 0;
2699         for (i = 1; i < num_pages; i++) {
2700                 page = extent_buffer_page(eb, i);
2701                 page_cache_release(page);
2702         }
2703         return 0;
2704 }
2705
2706
2707 int invalidate_extent_lru(struct extent_io_tree *tree, u64 start,
2708                           unsigned long len)
2709 {
2710         struct list_head *lru = &tree->buffer_lru;
2711         struct list_head *cur = lru->next;
2712         struct extent_buffer *eb;
2713         int found = 0;
2714
2715         spin_lock(&tree->lru_lock);
2716         if (list_empty(lru))
2717                 goto out;
2718
2719         do {
2720                 eb = list_entry(cur, struct extent_buffer, lru);
2721                 if (eb->start <= start && eb->start + eb->len > start) {
2722                         eb->flags &= ~EXTENT_UPTODATE;
2723                 }
2724                 cur = cur->next;
2725         } while (cur != lru);
2726 out:
2727         spin_unlock(&tree->lru_lock);
2728         return found;
2729 }
2730
2731 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
2732                                                    u64 start,
2733                                                    unsigned long len,
2734                                                    gfp_t mask)
2735 {
2736         struct extent_buffer *eb = NULL;
2737         unsigned long flags;
2738
2739         spin_lock(&tree->lru_lock);
2740         eb = find_lru(tree, start, len);
2741         spin_unlock(&tree->lru_lock);
2742         if (eb) {
2743                 return eb;
2744         }
2745
2746         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
2747         INIT_LIST_HEAD(&eb->lru);
2748         eb->start = start;
2749         eb->len = len;
2750         spin_lock_irqsave(&leak_lock, flags);
2751         list_add(&eb->leak_list, &buffers);
2752         spin_unlock_irqrestore(&leak_lock, flags);
2753         atomic_set(&eb->refs, 1);
2754
2755         return eb;
2756 }
2757
2758 static void __free_extent_buffer(struct extent_buffer *eb)
2759 {
2760         unsigned long flags;
2761         spin_lock_irqsave(&leak_lock, flags);
2762         list_del(&eb->leak_list);
2763         spin_unlock_irqrestore(&leak_lock, flags);
2764         kmem_cache_free(extent_buffer_cache, eb);
2765 }
2766
2767 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
2768                                           u64 start, unsigned long len,
2769                                           struct page *page0,
2770                                           gfp_t mask)
2771 {
2772         unsigned long num_pages = num_extent_pages(start, len);
2773         unsigned long i;
2774         unsigned long index = start >> PAGE_CACHE_SHIFT;
2775         struct extent_buffer *eb;
2776         struct page *p;
2777         struct address_space *mapping = tree->mapping;
2778         int uptodate = 1;
2779
2780         eb = __alloc_extent_buffer(tree, start, len, mask);
2781         if (!eb)
2782                 return NULL;
2783
2784         if (eb->flags & EXTENT_BUFFER_FILLED)
2785                 goto lru_add;
2786
2787         if (page0) {
2788                 eb->first_page = page0;
2789                 i = 1;
2790                 index++;
2791                 page_cache_get(page0);
2792                 mark_page_accessed(page0);
2793                 set_page_extent_mapped(page0);
2794                 set_page_extent_head(page0, len);
2795                 uptodate = PageUptodate(page0);
2796         } else {
2797                 i = 0;
2798         }
2799         for (; i < num_pages; i++, index++) {
2800                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
2801                 if (!p) {
2802                         WARN_ON(1);
2803                         goto fail;
2804                 }
2805                 set_page_extent_mapped(p);
2806                 mark_page_accessed(p);
2807                 if (i == 0) {
2808                         eb->first_page = p;
2809                         set_page_extent_head(p, len);
2810                 } else {
2811                         set_page_private(p, EXTENT_PAGE_PRIVATE);
2812                 }
2813                 if (!PageUptodate(p))
2814                         uptodate = 0;
2815                 unlock_page(p);
2816         }
2817         if (uptodate)
2818                 eb->flags |= EXTENT_UPTODATE;
2819         eb->flags |= EXTENT_BUFFER_FILLED;
2820
2821 lru_add:
2822         spin_lock(&tree->lru_lock);
2823         add_lru(tree, eb);
2824         spin_unlock(&tree->lru_lock);
2825         return eb;
2826
2827 fail:
2828         spin_lock(&tree->lru_lock);
2829         list_del_init(&eb->lru);
2830         spin_unlock(&tree->lru_lock);
2831         if (!atomic_dec_and_test(&eb->refs))
2832                 return NULL;
2833         for (index = 1; index < i; index++) {
2834                 page_cache_release(extent_buffer_page(eb, index));
2835         }
2836         if (i > 0)
2837                 page_cache_release(extent_buffer_page(eb, 0));
2838         __free_extent_buffer(eb);
2839         return NULL;
2840 }
2841 EXPORT_SYMBOL(alloc_extent_buffer);
2842
2843 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
2844                                          u64 start, unsigned long len,
2845                                           gfp_t mask)
2846 {
2847         unsigned long num_pages = num_extent_pages(start, len);
2848         unsigned long i;
2849         unsigned long index = start >> PAGE_CACHE_SHIFT;
2850         struct extent_buffer *eb;
2851         struct page *p;
2852         struct address_space *mapping = tree->mapping;
2853         int uptodate = 1;
2854
2855         eb = __alloc_extent_buffer(tree, start, len, mask);
2856         if (!eb)
2857                 return NULL;
2858
2859         if (eb->flags & EXTENT_BUFFER_FILLED)
2860                 goto lru_add;
2861
2862         for (i = 0; i < num_pages; i++, index++) {
2863                 p = find_get_page(mapping, index);
2864                 if (!p) {
2865                         goto fail;
2866                 }
2867                 if (TestSetPageLocked(p)) {
2868                         page_cache_release(p);
2869                         goto fail;
2870                 }
2871
2872                 set_page_extent_mapped(p);
2873                 mark_page_accessed(p);
2874
2875                 if (i == 0) {
2876                         eb->first_page = p;
2877                         set_page_extent_head(p, len);
2878                 } else {
2879                         set_page_private(p, EXTENT_PAGE_PRIVATE);
2880                 }
2881
2882                 if (!PageUptodate(p))
2883                         uptodate = 0;
2884                 unlock_page(p);
2885         }
2886         if (uptodate)
2887                 eb->flags |= EXTENT_UPTODATE;
2888         eb->flags |= EXTENT_BUFFER_FILLED;
2889
2890 lru_add:
2891         spin_lock(&tree->lru_lock);
2892         add_lru(tree, eb);
2893         spin_unlock(&tree->lru_lock);
2894         return eb;
2895 fail:
2896         spin_lock(&tree->lru_lock);
2897         list_del_init(&eb->lru);
2898         spin_unlock(&tree->lru_lock);
2899         if (!atomic_dec_and_test(&eb->refs))
2900                 return NULL;
2901         for (index = 1; index < i; index++) {
2902                 page_cache_release(extent_buffer_page(eb, index));
2903         }
2904         if (i > 0)
2905                 page_cache_release(extent_buffer_page(eb, 0));
2906         __free_extent_buffer(eb);
2907         return NULL;
2908 }
2909 EXPORT_SYMBOL(find_extent_buffer);
2910
2911 void free_extent_buffer(struct extent_buffer *eb)
2912 {
2913         unsigned long i;
2914         unsigned long num_pages;
2915
2916         if (!eb)
2917                 return;
2918
2919         if (!atomic_dec_and_test(&eb->refs))
2920                 return;
2921
2922         WARN_ON(!list_empty(&eb->lru));
2923         num_pages = num_extent_pages(eb->start, eb->len);
2924
2925         for (i = 1; i < num_pages; i++) {
2926                 page_cache_release(extent_buffer_page(eb, i));
2927         }
2928         page_cache_release(extent_buffer_page(eb, 0));
2929         __free_extent_buffer(eb);
2930 }
2931 EXPORT_SYMBOL(free_extent_buffer);
2932
2933 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
2934                               struct extent_buffer *eb)
2935 {
2936         int set;
2937         unsigned long i;
2938         unsigned long num_pages;
2939         struct page *page;
2940
2941         u64 start = eb->start;
2942         u64 end = start + eb->len - 1;
2943
2944         set = clear_extent_dirty(tree, start, end, GFP_NOFS);
2945         num_pages = num_extent_pages(eb->start, eb->len);
2946
2947         for (i = 0; i < num_pages; i++) {
2948                 page = extent_buffer_page(eb, i);
2949                 if (i == 0)
2950                         set_page_extent_head(page, eb->len);
2951                 else
2952                         set_page_private(page, EXTENT_PAGE_PRIVATE);
2953
2954                 /*
2955                  * if we're on the last page or the first page and the
2956                  * block isn't aligned on a page boundary, do extra checks
2957                  * to make sure we don't clean page that is partially dirty
2958                  */
2959                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
2960                     ((i == num_pages - 1) &&
2961                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
2962                         start = (u64)page->index << PAGE_CACHE_SHIFT;
2963                         end  = start + PAGE_CACHE_SIZE - 1;
2964                         if (test_range_bit(tree, start, end,
2965                                            EXTENT_DIRTY, 0)) {
2966                                 continue;
2967                         }
2968                 }
2969                 clear_page_dirty_for_io(page);
2970                 read_lock_irq(&page->mapping->tree_lock);
2971                 if (!PageDirty(page)) {
2972                         radix_tree_tag_clear(&page->mapping->page_tree,
2973                                                 page_index(page),
2974                                                 PAGECACHE_TAG_DIRTY);
2975                 }
2976                 read_unlock_irq(&page->mapping->tree_lock);
2977         }
2978         return 0;
2979 }
2980 EXPORT_SYMBOL(clear_extent_buffer_dirty);
2981
2982 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
2983                                     struct extent_buffer *eb)
2984 {
2985         return wait_on_extent_writeback(tree, eb->start,
2986                                         eb->start + eb->len - 1);
2987 }
2988 EXPORT_SYMBOL(wait_on_extent_buffer_writeback);
2989
2990 int set_extent_buffer_dirty(struct extent_io_tree *tree,
2991                              struct extent_buffer *eb)
2992 {
2993         unsigned long i;
2994         unsigned long num_pages;
2995
2996         num_pages = num_extent_pages(eb->start, eb->len);
2997         for (i = 0; i < num_pages; i++) {
2998                 struct page *page = extent_buffer_page(eb, i);
2999                 /* writepage may need to do something special for the
3000                  * first page, we have to make sure page->private is
3001                  * properly set.  releasepage may drop page->private
3002                  * on us if the page isn't already dirty.
3003                  */
3004                 if (i == 0) {
3005                         set_page_extent_head(page, eb->len);
3006                 } else if (PagePrivate(page) &&
3007                            page->private != EXTENT_PAGE_PRIVATE) {
3008                         set_page_extent_mapped(page);
3009                 }
3010                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3011         }
3012         return set_extent_dirty(tree, eb->start,
3013                                 eb->start + eb->len - 1, GFP_NOFS);
3014 }
3015 EXPORT_SYMBOL(set_extent_buffer_dirty);
3016
3017 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3018                                 struct extent_buffer *eb)
3019 {
3020         unsigned long i;
3021         struct page *page;
3022         unsigned long num_pages;
3023
3024         num_pages = num_extent_pages(eb->start, eb->len);
3025         eb->flags &= ~EXTENT_UPTODATE;
3026
3027         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3028                               GFP_NOFS);
3029         for (i = 0; i < num_pages; i++) {
3030                 page = extent_buffer_page(eb, i);
3031                 ClearPageUptodate(page);
3032         }
3033         return 0;
3034 }
3035
3036 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3037                                 struct extent_buffer *eb)
3038 {
3039         unsigned long i;
3040         struct page *page;
3041         unsigned long num_pages;
3042
3043         num_pages = num_extent_pages(eb->start, eb->len);
3044
3045         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3046                             GFP_NOFS);
3047         for (i = 0; i < num_pages; i++) {
3048                 page = extent_buffer_page(eb, i);
3049                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3050                     ((i == num_pages - 1) &&
3051                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3052                         check_page_uptodate(tree, page);
3053                         continue;
3054                 }
3055                 SetPageUptodate(page);
3056         }
3057         return 0;
3058 }
3059 EXPORT_SYMBOL(set_extent_buffer_uptodate);
3060
3061 int extent_range_uptodate(struct extent_io_tree *tree,
3062                           u64 start, u64 end)
3063 {
3064         struct page *page;
3065         int ret;
3066         int pg_uptodate = 1;
3067         int uptodate;
3068         unsigned long index;
3069
3070         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1);
3071         if (ret)
3072                 return 1;
3073         while(start <= end) {
3074                 index = start >> PAGE_CACHE_SHIFT;
3075                 page = find_get_page(tree->mapping, index);
3076                 uptodate = PageUptodate(page);
3077                 page_cache_release(page);
3078                 if (!uptodate) {
3079                         pg_uptodate = 0;
3080                         break;
3081                 }
3082                 start += PAGE_CACHE_SIZE;
3083         }
3084         return pg_uptodate;
3085 }
3086
3087 int extent_buffer_uptodate(struct extent_io_tree *tree,
3088                            struct extent_buffer *eb)
3089 {
3090         int ret = 0;
3091         unsigned long num_pages;
3092         unsigned long i;
3093         struct page *page;
3094         int pg_uptodate = 1;
3095
3096         if (eb->flags & EXTENT_UPTODATE)
3097                 return 1;
3098
3099         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3100                            EXTENT_UPTODATE, 1);
3101         if (ret)
3102                 return ret;
3103
3104         num_pages = num_extent_pages(eb->start, eb->len);
3105         for (i = 0; i < num_pages; i++) {
3106                 page = extent_buffer_page(eb, i);
3107                 if (!PageUptodate(page)) {
3108                         pg_uptodate = 0;
3109                         break;
3110                 }
3111         }
3112         return pg_uptodate;
3113 }
3114 EXPORT_SYMBOL(extent_buffer_uptodate);
3115
3116 int read_extent_buffer_pages(struct extent_io_tree *tree,
3117                              struct extent_buffer *eb,
3118                              u64 start, int wait,
3119                              get_extent_t *get_extent, int mirror_num)
3120 {
3121         unsigned long i;
3122         unsigned long start_i;
3123         struct page *page;
3124         int err;
3125         int ret = 0;
3126         int locked_pages = 0;
3127         int all_uptodate = 1;
3128         int inc_all_pages = 0;
3129         unsigned long num_pages;
3130         struct bio *bio = NULL;
3131
3132         if (eb->flags & EXTENT_UPTODATE)
3133                 return 0;
3134
3135         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3136                            EXTENT_UPTODATE, 1)) {
3137                 return 0;
3138         }
3139
3140         if (start) {
3141                 WARN_ON(start < eb->start);
3142                 start_i = (start >> PAGE_CACHE_SHIFT) -
3143                         (eb->start >> PAGE_CACHE_SHIFT);
3144         } else {
3145                 start_i = 0;
3146         }
3147
3148         num_pages = num_extent_pages(eb->start, eb->len);
3149         for (i = start_i; i < num_pages; i++) {
3150                 page = extent_buffer_page(eb, i);
3151                 if (!wait) {
3152                         if (TestSetPageLocked(page))
3153                                 goto unlock_exit;
3154                 } else {
3155                         lock_page(page);
3156                 }
3157                 locked_pages++;
3158                 if (!PageUptodate(page)) {
3159                         all_uptodate = 0;
3160                 }
3161         }
3162         if (all_uptodate) {
3163                 if (start_i == 0)
3164                         eb->flags |= EXTENT_UPTODATE;
3165                 goto unlock_exit;
3166         }
3167
3168         for (i = start_i; i < num_pages; i++) {
3169                 page = extent_buffer_page(eb, i);
3170                 if (inc_all_pages)
3171                         page_cache_get(page);
3172                 if (!PageUptodate(page)) {
3173                         if (start_i == 0)
3174                                 inc_all_pages = 1;
3175                         ClearPageError(page);
3176                         err = __extent_read_full_page(tree, page,
3177                                                       get_extent, &bio,
3178                                                       mirror_num);
3179                         if (err) {
3180                                 ret = err;
3181                         }
3182                 } else {
3183                         unlock_page(page);
3184                 }
3185         }
3186
3187         if (bio)
3188                 submit_one_bio(READ, bio, mirror_num);
3189
3190         if (ret || !wait) {
3191                 return ret;
3192         }
3193         for (i = start_i; i < num_pages; i++) {
3194                 page = extent_buffer_page(eb, i);
3195                 wait_on_page_locked(page);
3196                 if (!PageUptodate(page)) {
3197                         ret = -EIO;
3198                 }
3199         }
3200         if (!ret)
3201                 eb->flags |= EXTENT_UPTODATE;
3202         return ret;
3203
3204 unlock_exit:
3205         i = start_i;
3206         while(locked_pages > 0) {
3207                 page = extent_buffer_page(eb, i);
3208                 i++;
3209                 unlock_page(page);
3210                 locked_pages--;
3211         }
3212         return ret;
3213 }
3214 EXPORT_SYMBOL(read_extent_buffer_pages);
3215
3216 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3217                         unsigned long start,
3218                         unsigned long len)
3219 {
3220         size_t cur;
3221         size_t offset;
3222         struct page *page;
3223         char *kaddr;
3224         char *dst = (char *)dstv;
3225         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3226         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3227
3228         WARN_ON(start > eb->len);
3229         WARN_ON(start + len > eb->start + eb->len);
3230
3231         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3232
3233         while(len > 0) {
3234                 page = extent_buffer_page(eb, i);
3235
3236                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3237                 kaddr = kmap_atomic(page, KM_USER1);
3238                 memcpy(dst, kaddr + offset, cur);
3239                 kunmap_atomic(kaddr, KM_USER1);
3240
3241                 dst += cur;
3242                 len -= cur;
3243                 offset = 0;
3244                 i++;
3245         }
3246 }
3247 EXPORT_SYMBOL(read_extent_buffer);
3248
3249 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3250                                unsigned long min_len, char **token, char **map,
3251                                unsigned long *map_start,
3252                                unsigned long *map_len, int km)
3253 {
3254         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3255         char *kaddr;
3256         struct page *p;
3257         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3258         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3259         unsigned long end_i = (start_offset + start + min_len - 1) >>
3260                 PAGE_CACHE_SHIFT;
3261
3262         if (i != end_i)
3263                 return -EINVAL;
3264
3265         if (i == 0) {
3266                 offset = start_offset;
3267                 *map_start = 0;
3268         } else {
3269                 offset = 0;
3270                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3271         }
3272         if (start + min_len > eb->len) {
3273 printk("bad mapping eb start %Lu len %lu, wanted %lu %lu\n", eb->start, eb->len, start, min_len);
3274                 WARN_ON(1);
3275         }
3276
3277         p = extent_buffer_page(eb, i);
3278         kaddr = kmap_atomic(p, km);
3279         *token = kaddr;
3280         *map = kaddr + offset;
3281         *map_len = PAGE_CACHE_SIZE - offset;
3282         return 0;
3283 }
3284 EXPORT_SYMBOL(map_private_extent_buffer);
3285
3286 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3287                       unsigned long min_len,
3288                       char **token, char **map,
3289                       unsigned long *map_start,
3290                       unsigned long *map_len, int km)
3291 {
3292         int err;
3293         int save = 0;
3294         if (eb->map_token) {
3295                 unmap_extent_buffer(eb, eb->map_token, km);
3296                 eb->map_token = NULL;
3297                 save = 1;
3298         }
3299         err = map_private_extent_buffer(eb, start, min_len, token, map,
3300                                        map_start, map_len, km);
3301         if (!err && save) {
3302                 eb->map_token = *token;
3303                 eb->kaddr = *map;
3304                 eb->map_start = *map_start;
3305                 eb->map_len = *map_len;
3306         }
3307         return err;
3308 }
3309 EXPORT_SYMBOL(map_extent_buffer);
3310
3311 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3312 {
3313         kunmap_atomic(token, km);
3314 }
3315 EXPORT_SYMBOL(unmap_extent_buffer);
3316
3317 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3318                           unsigned long start,
3319                           unsigned long len)
3320 {
3321         size_t cur;
3322         size_t offset;
3323         struct page *page;
3324         char *kaddr;
3325         char *ptr = (char *)ptrv;
3326         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3327         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3328         int ret = 0;
3329
3330         WARN_ON(start > eb->len);
3331         WARN_ON(start + len > eb->start + eb->len);
3332
3333         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3334
3335         while(len > 0) {
3336                 page = extent_buffer_page(eb, i);
3337
3338                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3339
3340                 kaddr = kmap_atomic(page, KM_USER0);
3341                 ret = memcmp(ptr, kaddr + offset, cur);
3342                 kunmap_atomic(kaddr, KM_USER0);
3343                 if (ret)
3344                         break;
3345
3346                 ptr += cur;
3347                 len -= cur;
3348                 offset = 0;
3349                 i++;
3350         }
3351         return ret;
3352 }
3353 EXPORT_SYMBOL(memcmp_extent_buffer);
3354
3355 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3356                          unsigned long start, unsigned long len)
3357 {
3358         size_t cur;
3359         size_t offset;
3360         struct page *page;
3361         char *kaddr;
3362         char *src = (char *)srcv;
3363         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3364         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3365
3366         WARN_ON(start > eb->len);
3367         WARN_ON(start + len > eb->start + eb->len);
3368
3369         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3370
3371         while(len > 0) {
3372                 page = extent_buffer_page(eb, i);
3373                 WARN_ON(!PageUptodate(page));
3374
3375                 cur = min(len, PAGE_CACHE_SIZE - offset);
3376                 kaddr = kmap_atomic(page, KM_USER1);
3377                 memcpy(kaddr + offset, src, cur);
3378                 kunmap_atomic(kaddr, KM_USER1);
3379
3380                 src += cur;
3381                 len -= cur;
3382                 offset = 0;
3383                 i++;
3384         }
3385 }
3386 EXPORT_SYMBOL(write_extent_buffer);
3387
3388 void memset_extent_buffer(struct extent_buffer *eb, char c,
3389                           unsigned long start, unsigned long len)
3390 {
3391         size_t cur;
3392         size_t offset;
3393         struct page *page;
3394         char *kaddr;
3395         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3396         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3397
3398         WARN_ON(start > eb->len);
3399         WARN_ON(start + len > eb->start + eb->len);
3400
3401         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3402
3403         while(len > 0) {
3404                 page = extent_buffer_page(eb, i);
3405                 WARN_ON(!PageUptodate(page));
3406
3407                 cur = min(len, PAGE_CACHE_SIZE - offset);
3408                 kaddr = kmap_atomic(page, KM_USER0);
3409                 memset(kaddr + offset, c, cur);
3410                 kunmap_atomic(kaddr, KM_USER0);
3411
3412                 len -= cur;
3413                 offset = 0;
3414                 i++;
3415         }
3416 }
3417 EXPORT_SYMBOL(memset_extent_buffer);
3418
3419 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3420                         unsigned long dst_offset, unsigned long src_offset,
3421                         unsigned long len)
3422 {
3423         u64 dst_len = dst->len;
3424         size_t cur;
3425         size_t offset;
3426         struct page *page;
3427         char *kaddr;
3428         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3429         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3430
3431         WARN_ON(src->len != dst_len);
3432
3433         offset = (start_offset + dst_offset) &
3434                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3435
3436         while(len > 0) {
3437                 page = extent_buffer_page(dst, i);
3438                 WARN_ON(!PageUptodate(page));
3439
3440                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3441
3442                 kaddr = kmap_atomic(page, KM_USER0);
3443                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3444                 kunmap_atomic(kaddr, KM_USER0);
3445
3446                 src_offset += cur;
3447                 len -= cur;
3448                 offset = 0;
3449                 i++;
3450         }
3451 }
3452 EXPORT_SYMBOL(copy_extent_buffer);
3453
3454 static void move_pages(struct page *dst_page, struct page *src_page,
3455                        unsigned long dst_off, unsigned long src_off,
3456                        unsigned long len)
3457 {
3458         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3459         if (dst_page == src_page) {
3460                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3461         } else {
3462                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3463                 char *p = dst_kaddr + dst_off + len;
3464                 char *s = src_kaddr + src_off + len;
3465
3466                 while (len--)
3467                         *--p = *--s;
3468
3469                 kunmap_atomic(src_kaddr, KM_USER1);
3470         }
3471         kunmap_atomic(dst_kaddr, KM_USER0);
3472 }
3473
3474 static void copy_pages(struct page *dst_page, struct page *src_page,
3475                        unsigned long dst_off, unsigned long src_off,
3476                        unsigned long len)
3477 {
3478         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3479         char *src_kaddr;
3480
3481         if (dst_page != src_page)
3482                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3483         else
3484                 src_kaddr = dst_kaddr;
3485
3486         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3487         kunmap_atomic(dst_kaddr, KM_USER0);
3488         if (dst_page != src_page)
3489                 kunmap_atomic(src_kaddr, KM_USER1);
3490 }
3491
3492 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3493                            unsigned long src_offset, unsigned long len)
3494 {
3495         size_t cur;
3496         size_t dst_off_in_page;
3497         size_t src_off_in_page;
3498         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3499         unsigned long dst_i;
3500         unsigned long src_i;
3501
3502         if (src_offset + len > dst->len) {
3503                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
3504                        src_offset, len, dst->len);
3505                 BUG_ON(1);
3506         }
3507         if (dst_offset + len > dst->len) {
3508                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
3509                        dst_offset, len, dst->len);
3510                 BUG_ON(1);
3511         }
3512
3513         while(len > 0) {
3514                 dst_off_in_page = (start_offset + dst_offset) &
3515                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3516                 src_off_in_page = (start_offset + src_offset) &
3517                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3518
3519                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3520                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3521
3522                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3523                                                src_off_in_page));
3524                 cur = min_t(unsigned long, cur,
3525                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3526
3527                 copy_pages(extent_buffer_page(dst, dst_i),
3528                            extent_buffer_page(dst, src_i),
3529                            dst_off_in_page, src_off_in_page, cur);
3530
3531                 src_offset += cur;
3532                 dst_offset += cur;
3533                 len -= cur;
3534         }
3535 }
3536 EXPORT_SYMBOL(memcpy_extent_buffer);
3537
3538 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3539                            unsigned long src_offset, unsigned long len)
3540 {
3541         size_t cur;
3542         size_t dst_off_in_page;
3543         size_t src_off_in_page;
3544         unsigned long dst_end = dst_offset + len - 1;
3545         unsigned long src_end = src_offset + len - 1;
3546         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3547         unsigned long dst_i;
3548         unsigned long src_i;
3549
3550         if (src_offset + len > dst->len) {
3551                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
3552                        src_offset, len, dst->len);
3553                 BUG_ON(1);
3554         }
3555         if (dst_offset + len > dst->len) {
3556                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
3557                        dst_offset, len, dst->len);
3558                 BUG_ON(1);
3559         }
3560         if (dst_offset < src_offset) {
3561                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3562                 return;
3563         }
3564         while(len > 0) {
3565                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3566                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3567
3568                 dst_off_in_page = (start_offset + dst_end) &
3569                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3570                 src_off_in_page = (start_offset + src_end) &
3571                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3572
3573                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3574                 cur = min(cur, dst_off_in_page + 1);
3575                 move_pages(extent_buffer_page(dst, dst_i),
3576                            extent_buffer_page(dst, src_i),
3577                            dst_off_in_page - cur + 1,
3578                            src_off_in_page - cur + 1, cur);
3579
3580                 dst_end -= cur;
3581                 src_end -= cur;
3582                 len -= cur;
3583         }
3584 }
3585 EXPORT_SYMBOL(memmove_extent_buffer);