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