Pileus Git - ~andy/linux/blob - fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/gfp.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/page-flags.h>
   8 #include <linux/module.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/swap.h>
  12 #include <linux/version.h>
  13 #include <linux/writeback.h>
  14 #include <linux/pagevec.h>
  15 #include "extent_io.h"
  16 #include "extent_map.h"
  17
  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         if (tree->ops && tree->ops->submit_bio_hook)
1709                 tree->ops->submit_bio_hook(page->mapping->host, rw, bio);
1710         else
1711                 submit_bio(rw, bio);
1712         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1713                 ret = -EOPNOTSUPP;
1714         bio_put(bio);
1715         return ret;
1716 }
1717
1718 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1719                               struct page *page, sector_t sector,
1720                               size_t size, unsigned long offset,
1721                               struct block_device *bdev,
1722                               struct bio **bio_ret,
1723                               unsigned long max_pages,
1724                               bio_end_io_t end_io_func)
1725 {
1726         int ret = 0;
1727         struct bio *bio;
1728         int nr;
1729
1730         if (bio_ret && *bio_ret) {
1731                 bio = *bio_ret;
1732                 if (bio->bi_sector + (bio->bi_size >> 9) != sector ||
1733                     (tree->ops && tree->ops->merge_bio_hook &&
1734                      tree->ops->merge_bio_hook(page, offset, size, bio)) ||
1735                     bio_add_page(bio, page, size, offset) < size) {
1736                         ret = submit_one_bio(rw, bio);
1737                         bio = NULL;
1738                 } else {
1739                         return 0;
1740                 }
1741         }
1742         nr = bio_get_nr_vecs(bdev);
1743         bio = extent_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1744         if (!bio) {
1745                 printk("failed to allocate bio nr %d\n", nr);
1746         }
1747
1748
1749         bio_add_page(bio, page, size, offset);
1750         bio->bi_end_io = end_io_func;
1751         bio->bi_private = tree;
1752
1753         if (bio_ret) {
1754                 *bio_ret = bio;
1755         } else {
1756                 ret = submit_one_bio(rw, bio);
1757         }
1758
1759         return ret;
1760 }
1761
1762 void set_page_extent_mapped(struct page *page)
1763 {
1764         if (!PagePrivate(page)) {
1765                 SetPagePrivate(page);
1766                 WARN_ON(!page->mapping->a_ops->invalidatepage);
1767                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1768                 page_cache_get(page);
1769         }
1770 }
1771
1772 void set_page_extent_head(struct page *page, unsigned long len)
1773 {
1774         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1775 }
1776
1777 /*
1778  * basic readpage implementation.  Locked extent state structs are inserted
1779  * into the tree that are removed when the IO is done (by the end_io
1780  * handlers)
1781  */
1782 static int __extent_read_full_page(struct extent_io_tree *tree,
1783                                    struct page *page,
1784                                    get_extent_t *get_extent,
1785                                    struct bio **bio)
1786 {
1787         struct inode *inode = page->mapping->host;
1788         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1789         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1790         u64 end;
1791         u64 cur = start;
1792         u64 extent_offset;
1793         u64 last_byte = i_size_read(inode);
1794         u64 block_start;
1795         u64 cur_end;
1796         sector_t sector;
1797         struct extent_map *em;
1798         struct block_device *bdev;
1799         int ret;
1800         int nr = 0;
1801         size_t page_offset = 0;
1802         size_t iosize;
1803         size_t blocksize = inode->i_sb->s_blocksize;
1804
1805         set_page_extent_mapped(page);
1806
1807         end = page_end;
1808         lock_extent(tree, start, end, GFP_NOFS);
1809
1810         while (cur <= end) {
1811                 if (cur >= last_byte) {
1812                         char *userpage;
1813                         iosize = PAGE_CACHE_SIZE - page_offset;
1814                         userpage = kmap_atomic(page, KM_USER0);
1815                         memset(userpage + page_offset, 0, iosize);
1816                         flush_dcache_page(page);
1817                         kunmap_atomic(userpage, KM_USER0);
1818                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1819                                             GFP_NOFS);
1820                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1821                         break;
1822                 }
1823                 em = get_extent(inode, page, page_offset, cur,
1824                                 end - cur + 1, 0);
1825                 if (IS_ERR(em) || !em) {
1826                         SetPageError(page);
1827                         unlock_extent(tree, cur, end, GFP_NOFS);
1828                         break;
1829                 }
1830
1831                 extent_offset = cur - em->start;
1832                 BUG_ON(extent_map_end(em) <= cur);
1833                 BUG_ON(end < cur);
1834
1835                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1836                 cur_end = min(extent_map_end(em) - 1, end);
1837                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1838                 sector = (em->block_start + extent_offset) >> 9;
1839                 bdev = em->bdev;
1840                 block_start = em->block_start;
1841                 free_extent_map(em);
1842                 em = NULL;
1843
1844                 /* we've found a hole, just zero and go on */
1845                 if (block_start == EXTENT_MAP_HOLE) {
1846                         char *userpage;
1847                         userpage = kmap_atomic(page, KM_USER0);
1848                         memset(userpage + page_offset, 0, iosize);
1849                         flush_dcache_page(page);
1850                         kunmap_atomic(userpage, KM_USER0);
1851
1852                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1853                                             GFP_NOFS);
1854                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1855                         cur = cur + iosize;
1856                         page_offset += iosize;
1857                         continue;
1858                 }
1859                 /* the get_extent function already copied into the page */
1860                 if (test_range_bit(tree, cur, cur_end, EXTENT_UPTODATE, 1)) {
1861                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1862                         cur = cur + iosize;
1863                         page_offset += iosize;
1864                         continue;
1865                 }
1866                 /* we have an inline extent but it didn't get marked up
1867                  * to date.  Error out
1868                  */
1869                 if (block_start == EXTENT_MAP_INLINE) {
1870                         SetPageError(page);
1871                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1872                         cur = cur + iosize;
1873                         page_offset += iosize;
1874                         continue;
1875                 }
1876
1877                 ret = 0;
1878                 if (tree->ops && tree->ops->readpage_io_hook) {
1879                         ret = tree->ops->readpage_io_hook(page, cur,
1880                                                           cur + iosize - 1);
1881                 }
1882                 if (!ret) {
1883                         unsigned long nr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
1884                         nr -= page->index;
1885                         ret = submit_extent_page(READ, tree, page,
1886                                          sector, iosize, page_offset,
1887                                          bdev, bio, nr,
1888                                          end_bio_extent_readpage);
1889                 }
1890                 if (ret)
1891                         SetPageError(page);
1892                 cur = cur + iosize;
1893                 page_offset += iosize;
1894                 nr++;
1895         }
1896         if (!nr) {
1897                 if (!PageError(page))
1898                         SetPageUptodate(page);
1899                 unlock_page(page);
1900         }
1901         return 0;
1902 }
1903
1904 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
1905                             get_extent_t *get_extent)
1906 {
1907         struct bio *bio = NULL;
1908         int ret;
1909
1910         ret = __extent_read_full_page(tree, page, get_extent, &bio);
1911         if (bio)
1912                 submit_one_bio(READ, bio);
1913         return ret;
1914 }
1915 EXPORT_SYMBOL(extent_read_full_page);
1916
1917 /*
1918  * the writepage semantics are similar to regular writepage.  extent
1919  * records are inserted to lock ranges in the tree, and as dirty areas
1920  * are found, they are marked writeback.  Then the lock bits are removed
1921  * and the end_io handler clears the writeback ranges
1922  */
1923 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
1924                               void *data)
1925 {
1926         struct inode *inode = page->mapping->host;
1927         struct extent_page_data *epd = data;
1928         struct extent_io_tree *tree = epd->tree;
1929         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1930         u64 delalloc_start;
1931         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1932         u64 end;
1933         u64 cur = start;
1934         u64 extent_offset;
1935         u64 last_byte = i_size_read(inode);
1936         u64 block_start;
1937         u64 iosize;
1938         sector_t sector;
1939         struct extent_map *em;
1940         struct block_device *bdev;
1941         int ret;
1942         int nr = 0;
1943         size_t page_offset = 0;
1944         size_t blocksize;
1945         loff_t i_size = i_size_read(inode);
1946         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
1947         u64 nr_delalloc;
1948         u64 delalloc_end;
1949
1950         WARN_ON(!PageLocked(page));
1951         if (page->index > end_index) {
1952                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
1953                 unlock_page(page);
1954                 return 0;
1955         }
1956
1957         if (page->index == end_index) {
1958                 char *userpage;
1959
1960                 size_t offset = i_size & (PAGE_CACHE_SIZE - 1);
1961
1962                 userpage = kmap_atomic(page, KM_USER0);
1963                 memset(userpage + offset, 0, PAGE_CACHE_SIZE - offset);
1964                 flush_dcache_page(page);
1965                 kunmap_atomic(userpage, KM_USER0);
1966         }
1967
1968         set_page_extent_mapped(page);
1969
1970         delalloc_start = start;
1971         delalloc_end = 0;
1972         while(delalloc_end < page_end) {
1973                 nr_delalloc = find_lock_delalloc_range(tree, &delalloc_start,
1974                                                        &delalloc_end,
1975                                                        128 * 1024 * 1024);
1976                 if (nr_delalloc == 0) {
1977                         delalloc_start = delalloc_end + 1;
1978                         continue;
1979                 }
1980                 tree->ops->fill_delalloc(inode, delalloc_start,
1981                                          delalloc_end);
1982                 clear_extent_bit(tree, delalloc_start,
1983                                  delalloc_end,
1984                                  EXTENT_LOCKED | EXTENT_DELALLOC,
1985                                  1, 0, GFP_NOFS);
1986                 delalloc_start = delalloc_end + 1;
1987         }
1988         lock_extent(tree, start, page_end, GFP_NOFS);
1989
1990         end = page_end;
1991         if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0)) {
1992                 printk("found delalloc bits after lock_extent\n");
1993         }
1994
1995         if (last_byte <= start) {
1996                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
1997                 goto done;
1998         }
1999
2000         set_extent_uptodate(tree, start, page_end, GFP_NOFS);
2001         blocksize = inode->i_sb->s_blocksize;
2002
2003         while (cur <= end) {
2004                 if (cur >= last_byte) {
2005                         clear_extent_dirty(tree, cur, page_end, GFP_NOFS);
2006                         break;
2007                 }
2008                 em = epd->get_extent(inode, page, page_offset, cur,
2009                                      end - cur + 1, 1);
2010                 if (IS_ERR(em) || !em) {
2011                         SetPageError(page);
2012                         break;
2013                 }
2014
2015                 extent_offset = cur - em->start;
2016                 BUG_ON(extent_map_end(em) <= cur);
2017                 BUG_ON(end < cur);
2018                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2019                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2020                 sector = (em->block_start + extent_offset) >> 9;
2021                 bdev = em->bdev;
2022                 block_start = em->block_start;
2023                 free_extent_map(em);
2024                 em = NULL;
2025
2026                 if (block_start == EXTENT_MAP_HOLE ||
2027                     block_start == EXTENT_MAP_INLINE) {
2028                         clear_extent_dirty(tree, cur,
2029                                            cur + iosize - 1, GFP_NOFS);
2030                         cur = cur + iosize;
2031                         page_offset += iosize;
2032                         continue;
2033                 }
2034
2035                 /* leave this out until we have a page_mkwrite call */
2036                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2037                                    EXTENT_DIRTY, 0)) {
2038                         cur = cur + iosize;
2039                         page_offset += iosize;
2040                         continue;
2041                 }
2042                 clear_extent_dirty(tree, cur, cur + iosize - 1, GFP_NOFS);
2043                 if (tree->ops && tree->ops->writepage_io_hook) {
2044                         ret = tree->ops->writepage_io_hook(page, cur,
2045                                                 cur + iosize - 1);
2046                 } else {
2047                         ret = 0;
2048                 }
2049                 if (ret)
2050                         SetPageError(page);
2051                 else {
2052                         unsigned long max_nr = end_index + 1;
2053                         set_range_writeback(tree, cur, cur + iosize - 1);
2054                         if (!PageWriteback(page)) {
2055                                 printk("warning page %lu not writeback, "
2056                                        "cur %llu end %llu\n", page->index,
2057                                        (unsigned long long)cur,
2058                                        (unsigned long long)end);
2059                         }
2060
2061                         ret = submit_extent_page(WRITE, tree, page, sector,
2062                                                  iosize, page_offset, bdev,
2063                                                  &epd->bio, max_nr,
2064                                                  end_bio_extent_writepage);
2065                         if (ret)
2066                                 SetPageError(page);
2067                 }
2068                 cur = cur + iosize;
2069                 page_offset += iosize;
2070                 nr++;
2071         }
2072 done:
2073         if (nr == 0) {
2074                 /* make sure the mapping tag for page dirty gets cleared */
2075                 set_page_writeback(page);
2076                 end_page_writeback(page);
2077         }
2078         unlock_extent(tree, start, page_end, GFP_NOFS);
2079         unlock_page(page);
2080         return 0;
2081 }
2082
2083 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,18)
2084
2085 /* Taken directly from 2.6.23 for 2.6.18 back port */
2086 typedef int (*writepage_t)(struct page *page, struct writeback_control *wbc,
2087                                 void *data);
2088
2089 /**
2090  * write_cache_pages - walk the list of dirty pages of the given address space
2091  * and write all of them.
2092  * @mapping: address space structure to write
2093  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2094  * @writepage: function called for each page
2095  * @data: data passed to writepage function
2096  *
2097  * If a page is already under I/O, write_cache_pages() skips it, even
2098  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2099  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2100  * and msync() need to guarantee that all the data which was dirty at the time
2101  * the call was made get new I/O started against them.  If wbc->sync_mode is
2102  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2103  * existing IO to complete.
2104  */
2105 static int write_cache_pages(struct address_space *mapping,
2106                       struct writeback_control *wbc, writepage_t writepage,
2107                       void *data)
2108 {
2109         struct backing_dev_info *bdi = mapping->backing_dev_info;
2110         int ret = 0;
2111         int done = 0;
2112         struct pagevec pvec;
2113         int nr_pages;
2114         pgoff_t index;
2115         pgoff_t end;            /* Inclusive */
2116         int scanned = 0;
2117         int range_whole = 0;
2118
2119         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2120                 wbc->encountered_congestion = 1;
2121                 return 0;
2122         }
2123
2124         pagevec_init(&pvec, 0);
2125         if (wbc->range_cyclic) {
2126                 index = mapping->writeback_index; /* Start from prev offset */
2127                 end = -1;
2128         } else {
2129                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2130                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2131                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2132                         range_whole = 1;
2133                 scanned = 1;
2134         }
2135 retry:
2136         while (!done && (index <= end) &&
2137                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2138                                               PAGECACHE_TAG_DIRTY,
2139                                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2140                 unsigned i;
2141
2142                 scanned = 1;
2143                 for (i = 0; i < nr_pages; i++) {
2144                         struct page *page = pvec.pages[i];
2145
2146                         /*
2147                          * At this point we hold neither mapping->tree_lock nor
2148                          * lock on the page itself: the page may be truncated or
2149                          * invalidated (changing page->mapping to NULL), or even
2150                          * swizzled back from swapper_space to tmpfs file
2151                          * mapping
2152                          */
2153                         lock_page(page);
2154
2155                         if (unlikely(page->mapping != mapping)) {
2156                                 unlock_page(page);
2157                                 continue;
2158                         }
2159
2160                         if (!wbc->range_cyclic && page->index > end) {
2161                                 done = 1;
2162                                 unlock_page(page);
2163                                 continue;
2164                         }
2165
2166                         if (wbc->sync_mode != WB_SYNC_NONE)
2167                                 wait_on_page_writeback(page);
2168
2169                         if (PageWriteback(page) ||
2170                             !clear_page_dirty_for_io(page)) {
2171                                 unlock_page(page);
2172                                 continue;
2173                         }
2174
2175                         ret = (*writepage)(page, wbc, data);
2176
2177                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2178                                 unlock_page(page);
2179                                 ret = 0;
2180                         }
2181                         if (ret || (--(wbc->nr_to_write) <= 0))
2182                                 done = 1;
2183                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2184                                 wbc->encountered_congestion = 1;
2185                                 done = 1;
2186                         }
2187                 }
2188                 pagevec_release(&pvec);
2189                 cond_resched();
2190         }
2191         if (!scanned && !done) {
2192                 /*
2193                  * We hit the last page and there is more work to be done: wrap
2194                  * back to the start of the file
2195                  */
2196                 scanned = 1;
2197                 index = 0;
2198                 goto retry;
2199         }
2200         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2201                 mapping->writeback_index = index;
2202         return ret;
2203 }
2204 #endif
2205
2206 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2207                           get_extent_t *get_extent,
2208                           struct writeback_control *wbc)
2209 {
2210         int ret;
2211         struct address_space *mapping = page->mapping;
2212         struct extent_page_data epd = {
2213                 .bio = NULL,
2214                 .tree = tree,
2215                 .get_extent = get_extent,
2216         };
2217         struct writeback_control wbc_writepages = {
2218                 .bdi            = wbc->bdi,
2219                 .sync_mode      = WB_SYNC_NONE,
2220                 .older_than_this = NULL,
2221                 .nr_to_write    = 64,
2222                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2223                 .range_end      = (loff_t)-1,
2224         };
2225
2226
2227         ret = __extent_writepage(page, wbc, &epd);
2228
2229         write_cache_pages(mapping, &wbc_writepages, __extent_writepage, &epd);
2230         if (epd.bio) {
2231                 submit_one_bio(WRITE, epd.bio);
2232         }
2233         return ret;
2234 }
2235 EXPORT_SYMBOL(extent_write_full_page);
2236
2237
2238 int extent_writepages(struct extent_io_tree *tree,
2239                       struct address_space *mapping,
2240                       get_extent_t *get_extent,
2241                       struct writeback_control *wbc)
2242 {
2243         int ret = 0;
2244         struct extent_page_data epd = {
2245                 .bio = NULL,
2246                 .tree = tree,
2247                 .get_extent = get_extent,
2248         };
2249
2250         ret = write_cache_pages(mapping, wbc, __extent_writepage, &epd);
2251         if (epd.bio) {
2252                 submit_one_bio(WRITE, epd.bio);
2253         }
2254         return ret;
2255 }
2256 EXPORT_SYMBOL(extent_writepages);
2257
2258 int extent_readpages(struct extent_io_tree *tree,
2259                      struct address_space *mapping,
2260                      struct list_head *pages, unsigned nr_pages,
2261                      get_extent_t get_extent)
2262 {
2263         struct bio *bio = NULL;
2264         unsigned page_idx;
2265         struct pagevec pvec;
2266
2267         pagevec_init(&pvec, 0);
2268         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2269                 struct page *page = list_entry(pages->prev, struct page, lru);
2270
2271                 prefetchw(&page->flags);
2272                 list_del(&page->lru);
2273                 /*
2274                  * what we want to do here is call add_to_page_cache_lru,
2275                  * but that isn't exported, so we reproduce it here
2276                  */
2277                 if (!add_to_page_cache(page, mapping,
2278                                         page->index, GFP_KERNEL)) {
2279
2280                         /* open coding of lru_cache_add, also not exported */
2281                         page_cache_get(page);
2282                         if (!pagevec_add(&pvec, page))
2283                                 __pagevec_lru_add(&pvec);
2284                         __extent_read_full_page(tree, page, get_extent, &bio);
2285                 }
2286                 page_cache_release(page);
2287         }
2288         if (pagevec_count(&pvec))
2289                 __pagevec_lru_add(&pvec);
2290         BUG_ON(!list_empty(pages));
2291         if (bio)
2292                 submit_one_bio(READ, bio);
2293         return 0;
2294 }
2295 EXPORT_SYMBOL(extent_readpages);
2296
2297 /*
2298  * basic invalidatepage code, this waits on any locked or writeback
2299  * ranges corresponding to the page, and then deletes any extent state
2300  * records from the tree
2301  */
2302 int extent_invalidatepage(struct extent_io_tree *tree,
2303                           struct page *page, unsigned long offset)
2304 {
2305         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2306         u64 end = start + PAGE_CACHE_SIZE - 1;
2307         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2308
2309         start += (offset + blocksize -1) & ~(blocksize - 1);
2310         if (start > end)
2311                 return 0;
2312
2313         lock_extent(tree, start, end, GFP_NOFS);
2314         wait_on_extent_writeback(tree, start, end);
2315         clear_extent_bit(tree, start, end,
2316                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC,
2317                          1, 1, GFP_NOFS);
2318         return 0;
2319 }
2320 EXPORT_SYMBOL(extent_invalidatepage);
2321
2322 /*
2323  * simple commit_write call, set_range_dirty is used to mark both
2324  * the pages and the extent records as dirty
2325  */
2326 int extent_commit_write(struct extent_io_tree *tree,
2327                         struct inode *inode, struct page *page,
2328                         unsigned from, unsigned to)
2329 {
2330         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2331
2332         set_page_extent_mapped(page);
2333         set_page_dirty(page);
2334
2335         if (pos > inode->i_size) {
2336                 i_size_write(inode, pos);
2337                 mark_inode_dirty(inode);
2338         }
2339         return 0;
2340 }
2341 EXPORT_SYMBOL(extent_commit_write);
2342
2343 int extent_prepare_write(struct extent_io_tree *tree,
2344                          struct inode *inode, struct page *page,
2345                          unsigned from, unsigned to, get_extent_t *get_extent)
2346 {
2347         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2348         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2349         u64 block_start;
2350         u64 orig_block_start;
2351         u64 block_end;
2352         u64 cur_end;
2353         struct extent_map *em;
2354         unsigned blocksize = 1 << inode->i_blkbits;
2355         size_t page_offset = 0;
2356         size_t block_off_start;
2357         size_t block_off_end;
2358         int err = 0;
2359         int iocount = 0;
2360         int ret = 0;
2361         int isnew;
2362
2363         set_page_extent_mapped(page);
2364
2365         block_start = (page_start + from) & ~((u64)blocksize - 1);
2366         block_end = (page_start + to - 1) | (blocksize - 1);
2367         orig_block_start = block_start;
2368
2369         lock_extent(tree, page_start, page_end, GFP_NOFS);
2370         while(block_start <= block_end) {
2371                 em = get_extent(inode, page, page_offset, block_start,
2372                                 block_end - block_start + 1, 1);
2373                 if (IS_ERR(em) || !em) {
2374                         goto err;
2375                 }
2376                 cur_end = min(block_end, extent_map_end(em) - 1);
2377                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2378                 block_off_end = block_off_start + blocksize;
2379                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2380
2381                 if (!PageUptodate(page) && isnew &&
2382                     (block_off_end > to || block_off_start < from)) {
2383                         void *kaddr;
2384
2385                         kaddr = kmap_atomic(page, KM_USER0);
2386                         if (block_off_end > to)
2387                                 memset(kaddr + to, 0, block_off_end - to);
2388                         if (block_off_start < from)
2389                                 memset(kaddr + block_off_start, 0,
2390                                        from - block_off_start);
2391                         flush_dcache_page(page);
2392                         kunmap_atomic(kaddr, KM_USER0);
2393                 }
2394                 if ((em->block_start != EXTENT_MAP_HOLE &&
2395                      em->block_start != EXTENT_MAP_INLINE) &&
2396                     !isnew && !PageUptodate(page) &&
2397                     (block_off_end > to || block_off_start < from) &&
2398                     !test_range_bit(tree, block_start, cur_end,
2399                                     EXTENT_UPTODATE, 1)) {
2400                         u64 sector;
2401                         u64 extent_offset = block_start - em->start;
2402                         size_t iosize;
2403                         sector = (em->block_start + extent_offset) >> 9;
2404                         iosize = (cur_end - block_start + blocksize) &
2405                                 ~((u64)blocksize - 1);
2406                         /*
2407                          * we've already got the extent locked, but we
2408                          * need to split the state such that our end_bio
2409                          * handler can clear the lock.
2410                          */
2411                         set_extent_bit(tree, block_start,
2412                                        block_start + iosize - 1,
2413                                        EXTENT_LOCKED, 0, NULL, GFP_NOFS);
2414                         ret = submit_extent_page(READ, tree, page,
2415                                          sector, iosize, page_offset, em->bdev,
2416                                          NULL, 1,
2417                                          end_bio_extent_preparewrite);
2418                         iocount++;
2419                         block_start = block_start + iosize;
2420                 } else {
2421                         set_extent_uptodate(tree, block_start, cur_end,
2422                                             GFP_NOFS);
2423                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2424                         block_start = cur_end + 1;
2425                 }
2426                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2427                 free_extent_map(em);
2428         }
2429         if (iocount) {
2430                 wait_extent_bit(tree, orig_block_start,
2431                                 block_end, EXTENT_LOCKED);
2432         }
2433         check_page_uptodate(tree, page);
2434 err:
2435         /* FIXME, zero out newly allocated blocks on error */
2436         return err;
2437 }
2438 EXPORT_SYMBOL(extent_prepare_write);
2439
2440 /*
2441  * a helper for releasepage.  As long as there are no locked extents
2442  * in the range corresponding to the page, both state records and extent
2443  * map records are removed
2444  */
2445 int try_release_extent_mapping(struct extent_map_tree *map,
2446                                struct extent_io_tree *tree, struct page *page,
2447                                gfp_t mask)
2448 {
2449         struct extent_map *em;
2450         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2451         u64 end = start + PAGE_CACHE_SIZE - 1;
2452         u64 orig_start = start;
2453         int ret = 1;
2454         if ((mask & __GFP_WAIT) &&
2455             page->mapping->host->i_size > 16 * 1024 * 1024) {
2456                 u64 len;
2457                 while (start <= end) {
2458                         len = end - start + 1;
2459                         spin_lock(&map->lock);
2460                         em = lookup_extent_mapping(map, start, len);
2461                         if (!em || IS_ERR(em)) {
2462                                 spin_unlock(&map->lock);
2463                                 break;
2464                         }
2465                         if (em->start != start) {
2466                                 spin_unlock(&map->lock);
2467                                 free_extent_map(em);
2468                                 break;
2469                         }
2470                         if (!test_range_bit(tree, em->start,
2471                                             extent_map_end(em) - 1,
2472                                             EXTENT_LOCKED, 0)) {
2473                                 remove_extent_mapping(map, em);
2474                                 /* once for the rb tree */
2475                                 free_extent_map(em);
2476                         }
2477                         start = extent_map_end(em);
2478                         spin_unlock(&map->lock);
2479
2480                         /* once for us */
2481                         free_extent_map(em);
2482                 }
2483         }
2484         if (test_range_bit(tree, orig_start, end, EXTENT_IOBITS, 0))
2485                 ret = 0;
2486         else {
2487                 if ((mask & GFP_NOFS) == GFP_NOFS)
2488                         mask = GFP_NOFS;
2489                 clear_extent_bit(tree, orig_start, end, EXTENT_UPTODATE,
2490                                  1, 1, mask);
2491         }
2492         return ret;
2493 }
2494 EXPORT_SYMBOL(try_release_extent_mapping);
2495
2496 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2497                 get_extent_t *get_extent)
2498 {
2499         struct inode *inode = mapping->host;
2500         u64 start = iblock << inode->i_blkbits;
2501         sector_t sector = 0;
2502         struct extent_map *em;
2503
2504         em = get_extent(inode, NULL, 0, start, (1 << inode->i_blkbits), 0);
2505         if (!em || IS_ERR(em))
2506                 return 0;
2507
2508         if (em->block_start == EXTENT_MAP_INLINE ||
2509             em->block_start == EXTENT_MAP_HOLE)
2510                 goto out;
2511
2512         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2513 out:
2514         free_extent_map(em);
2515         return sector;
2516 }
2517
2518 static int add_lru(struct extent_io_tree *tree, struct extent_buffer *eb)
2519 {
2520         if (list_empty(&eb->lru)) {
2521                 extent_buffer_get(eb);
2522                 list_add(&eb->lru, &tree->buffer_lru);
2523                 tree->lru_size++;
2524                 if (tree->lru_size >= BUFFER_LRU_MAX) {
2525                         struct extent_buffer *rm;
2526                         rm = list_entry(tree->buffer_lru.prev,
2527                                         struct extent_buffer, lru);
2528                         tree->lru_size--;
2529                         list_del_init(&rm->lru);
2530                         free_extent_buffer(rm);
2531                 }
2532         } else
2533                 list_move(&eb->lru, &tree->buffer_lru);
2534         return 0;
2535 }
2536 static struct extent_buffer *find_lru(struct extent_io_tree *tree,
2537                                       u64 start, unsigned long len)
2538 {
2539         struct list_head *lru = &tree->buffer_lru;
2540         struct list_head *cur = lru->next;
2541         struct extent_buffer *eb;
2542
2543         if (list_empty(lru))
2544                 return NULL;
2545
2546         do {
2547                 eb = list_entry(cur, struct extent_buffer, lru);
2548                 if (eb->start == start && eb->len == len) {
2549                         extent_buffer_get(eb);
2550                         return eb;
2551                 }
2552                 cur = cur->next;
2553         } while (cur != lru);
2554         return NULL;
2555 }
2556
2557 static inline unsigned long num_extent_pages(u64 start, u64 len)
2558 {
2559         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
2560                 (start >> PAGE_CACHE_SHIFT);
2561 }
2562
2563 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2564                                               unsigned long i)
2565 {
2566         struct page *p;
2567         struct address_space *mapping;
2568
2569         if (i == 0)
2570                 return eb->first_page;
2571         i += eb->start >> PAGE_CACHE_SHIFT;
2572         mapping = eb->first_page->mapping;
2573         read_lock_irq(&mapping->tree_lock);
2574         p = radix_tree_lookup(&mapping->page_tree, i);
2575         read_unlock_irq(&mapping->tree_lock);
2576         return p;
2577 }
2578
2579 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
2580                                                    u64 start,
2581                                                    unsigned long len,
2582                                                    gfp_t mask)
2583 {
2584         struct extent_buffer *eb = NULL;
2585
2586         spin_lock(&tree->lru_lock);
2587         eb = find_lru(tree, start, len);
2588         spin_unlock(&tree->lru_lock);
2589         if (eb) {
2590                 return eb;
2591         }
2592
2593         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
2594         INIT_LIST_HEAD(&eb->lru);
2595         eb->start = start;
2596         eb->len = len;
2597         atomic_set(&eb->refs, 1);
2598
2599         return eb;
2600 }
2601
2602 static void __free_extent_buffer(struct extent_buffer *eb)
2603 {
2604         kmem_cache_free(extent_buffer_cache, eb);
2605 }
2606
2607 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
2608                                           u64 start, unsigned long len,
2609                                           struct page *page0,
2610                                           gfp_t mask)
2611 {
2612         unsigned long num_pages = num_extent_pages(start, len);
2613         unsigned long i;
2614         unsigned long index = start >> PAGE_CACHE_SHIFT;
2615         struct extent_buffer *eb;
2616         struct page *p;
2617         struct address_space *mapping = tree->mapping;
2618         int uptodate = 1;
2619
2620         eb = __alloc_extent_buffer(tree, start, len, mask);
2621         if (!eb || IS_ERR(eb))
2622                 return NULL;
2623
2624         if (eb->flags & EXTENT_BUFFER_FILLED)
2625                 goto lru_add;
2626
2627         if (page0) {
2628                 eb->first_page = page0;
2629                 i = 1;
2630                 index++;
2631                 page_cache_get(page0);
2632                 mark_page_accessed(page0);
2633                 set_page_extent_mapped(page0);
2634                 WARN_ON(!PageUptodate(page0));
2635                 set_page_extent_head(page0, len);
2636         } else {
2637                 i = 0;
2638         }
2639         for (; i < num_pages; i++, index++) {
2640                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
2641                 if (!p) {
2642                         WARN_ON(1);
2643                         goto fail;
2644                 }
2645                 set_page_extent_mapped(p);
2646                 mark_page_accessed(p);
2647                 if (i == 0) {
2648                         eb->first_page = p;
2649                         set_page_extent_head(p, len);
2650                 } else {
2651                         set_page_private(p, EXTENT_PAGE_PRIVATE);
2652                 }
2653                 if (!PageUptodate(p))
2654                         uptodate = 0;
2655                 unlock_page(p);
2656         }
2657         if (uptodate)
2658                 eb->flags |= EXTENT_UPTODATE;
2659         eb->flags |= EXTENT_BUFFER_FILLED;
2660
2661 lru_add:
2662         spin_lock(&tree->lru_lock);
2663         add_lru(tree, eb);
2664         spin_unlock(&tree->lru_lock);
2665         return eb;
2666
2667 fail:
2668         spin_lock(&tree->lru_lock);
2669         list_del_init(&eb->lru);
2670         spin_unlock(&tree->lru_lock);
2671         if (!atomic_dec_and_test(&eb->refs))
2672                 return NULL;
2673         for (index = 1; index < i; index++) {
2674                 page_cache_release(extent_buffer_page(eb, index));
2675         }
2676         if (i > 0)
2677                 page_cache_release(extent_buffer_page(eb, 0));
2678         __free_extent_buffer(eb);
2679         return NULL;
2680 }
2681 EXPORT_SYMBOL(alloc_extent_buffer);
2682
2683 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
2684                                          u64 start, unsigned long len,
2685                                           gfp_t mask)
2686 {
2687         unsigned long num_pages = num_extent_pages(start, len);
2688         unsigned long i;
2689         unsigned long index = start >> PAGE_CACHE_SHIFT;
2690         struct extent_buffer *eb;
2691         struct page *p;
2692         struct address_space *mapping = tree->mapping;
2693         int uptodate = 1;
2694
2695         eb = __alloc_extent_buffer(tree, start, len, mask);
2696         if (!eb || IS_ERR(eb))
2697                 return NULL;
2698
2699         if (eb->flags & EXTENT_BUFFER_FILLED)
2700                 goto lru_add;
2701
2702         for (i = 0; i < num_pages; i++, index++) {
2703                 p = find_lock_page(mapping, index);
2704                 if (!p) {
2705                         goto fail;
2706                 }
2707                 set_page_extent_mapped(p);
2708                 mark_page_accessed(p);
2709
2710                 if (i == 0) {
2711                         eb->first_page = p;
2712                         set_page_extent_head(p, len);
2713                 } else {
2714                         set_page_private(p, EXTENT_PAGE_PRIVATE);
2715                 }
2716
2717                 if (!PageUptodate(p))
2718                         uptodate = 0;
2719                 unlock_page(p);
2720         }
2721         if (uptodate)
2722                 eb->flags |= EXTENT_UPTODATE;
2723         eb->flags |= EXTENT_BUFFER_FILLED;
2724
2725 lru_add:
2726         spin_lock(&tree->lru_lock);
2727         add_lru(tree, eb);
2728         spin_unlock(&tree->lru_lock);
2729         return eb;
2730 fail:
2731         spin_lock(&tree->lru_lock);
2732         list_del_init(&eb->lru);
2733         spin_unlock(&tree->lru_lock);
2734         if (!atomic_dec_and_test(&eb->refs))
2735                 return NULL;
2736         for (index = 1; index < i; index++) {
2737                 page_cache_release(extent_buffer_page(eb, index));
2738         }
2739         if (i > 0)
2740                 page_cache_release(extent_buffer_page(eb, 0));
2741         __free_extent_buffer(eb);
2742         return NULL;
2743 }
2744 EXPORT_SYMBOL(find_extent_buffer);
2745
2746 void free_extent_buffer(struct extent_buffer *eb)
2747 {
2748         unsigned long i;
2749         unsigned long num_pages;
2750
2751         if (!eb)
2752                 return;
2753
2754         if (!atomic_dec_and_test(&eb->refs))
2755                 return;
2756
2757         WARN_ON(!list_empty(&eb->lru));
2758         num_pages = num_extent_pages(eb->start, eb->len);
2759
2760         for (i = 1; i < num_pages; i++) {
2761                 page_cache_release(extent_buffer_page(eb, i));
2762         }
2763         page_cache_release(extent_buffer_page(eb, 0));
2764         __free_extent_buffer(eb);
2765 }
2766 EXPORT_SYMBOL(free_extent_buffer);
2767
2768 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
2769                               struct extent_buffer *eb)
2770 {
2771         int set;
2772         unsigned long i;
2773         unsigned long num_pages;
2774         struct page *page;
2775
2776         u64 start = eb->start;
2777         u64 end = start + eb->len - 1;
2778
2779         set = clear_extent_dirty(tree, start, end, GFP_NOFS);
2780         num_pages = num_extent_pages(eb->start, eb->len);
2781
2782         for (i = 0; i < num_pages; i++) {
2783                 page = extent_buffer_page(eb, i);
2784                 lock_page(page);
2785                 if (i == 0)
2786                         set_page_extent_head(page, eb->len);
2787                 else
2788                         set_page_private(page, EXTENT_PAGE_PRIVATE);
2789
2790                 /*
2791                  * if we're on the last page or the first page and the
2792                  * block isn't aligned on a page boundary, do extra checks
2793                  * to make sure we don't clean page that is partially dirty
2794                  */
2795                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
2796                     ((i == num_pages - 1) &&
2797                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
2798                         start = (u64)page->index << PAGE_CACHE_SHIFT;
2799                         end  = start + PAGE_CACHE_SIZE - 1;
2800                         if (test_range_bit(tree, start, end,
2801                                            EXTENT_DIRTY, 0)) {
2802                                 unlock_page(page);
2803                                 continue;
2804                         }
2805                 }
2806                 clear_page_dirty_for_io(page);
2807                 read_lock_irq(&page->mapping->tree_lock);
2808                 if (!PageDirty(page)) {
2809                         radix_tree_tag_clear(&page->mapping->page_tree,
2810                                                 page_index(page),
2811                                                 PAGECACHE_TAG_DIRTY);
2812                 }
2813                 read_unlock_irq(&page->mapping->tree_lock);
2814                 unlock_page(page);
2815         }
2816         return 0;
2817 }
2818 EXPORT_SYMBOL(clear_extent_buffer_dirty);
2819
2820 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
2821                                     struct extent_buffer *eb)
2822 {
2823         return wait_on_extent_writeback(tree, eb->start,
2824                                         eb->start + eb->len - 1);
2825 }
2826 EXPORT_SYMBOL(wait_on_extent_buffer_writeback);
2827
2828 int set_extent_buffer_dirty(struct extent_io_tree *tree,
2829                              struct extent_buffer *eb)
2830 {
2831         unsigned long i;
2832         unsigned long num_pages;
2833
2834         num_pages = num_extent_pages(eb->start, eb->len);
2835         for (i = 0; i < num_pages; i++) {
2836                 struct page *page = extent_buffer_page(eb, i);
2837                 /* writepage may need to do something special for the
2838                  * first page, we have to make sure page->private is
2839                  * properly set.  releasepage may drop page->private
2840                  * on us if the page isn't already dirty.
2841                  */
2842                 if (i == 0) {
2843                         lock_page(page);
2844                         set_page_extent_head(page, eb->len);
2845                 } else if (PagePrivate(page) &&
2846                            page->private != EXTENT_PAGE_PRIVATE) {
2847                         lock_page(page);
2848                         set_page_extent_mapped(page);
2849                         unlock_page(page);
2850                 }
2851                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
2852                 if (i == 0)
2853                         unlock_page(page);
2854         }
2855         return set_extent_dirty(tree, eb->start,
2856                                 eb->start + eb->len - 1, GFP_NOFS);
2857 }
2858 EXPORT_SYMBOL(set_extent_buffer_dirty);
2859
2860 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
2861                                 struct extent_buffer *eb)
2862 {
2863         unsigned long i;
2864         struct page *page;
2865         unsigned long num_pages;
2866
2867         num_pages = num_extent_pages(eb->start, eb->len);
2868
2869         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
2870                             GFP_NOFS);
2871         for (i = 0; i < num_pages; i++) {
2872                 page = extent_buffer_page(eb, i);
2873                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
2874                     ((i == num_pages - 1) &&
2875                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
2876                         check_page_uptodate(tree, page);
2877                         continue;
2878                 }
2879                 SetPageUptodate(page);
2880         }
2881         return 0;
2882 }
2883 EXPORT_SYMBOL(set_extent_buffer_uptodate);
2884
2885 int extent_buffer_uptodate(struct extent_io_tree *tree,
2886                              struct extent_buffer *eb)
2887 {
2888         if (eb->flags & EXTENT_UPTODATE)
2889                 return 1;
2890         return test_range_bit(tree, eb->start, eb->start + eb->len - 1,
2891                            EXTENT_UPTODATE, 1);
2892 }
2893 EXPORT_SYMBOL(extent_buffer_uptodate);
2894
2895 int read_extent_buffer_pages(struct extent_io_tree *tree,
2896                              struct extent_buffer *eb,
2897                              u64 start, int wait,
2898                              get_extent_t *get_extent)
2899 {
2900         unsigned long i;
2901         unsigned long start_i;
2902         struct page *page;
2903         int err;
2904         int ret = 0;
2905         unsigned long num_pages;
2906         struct bio *bio = NULL;
2907
2908
2909         if (eb->flags & EXTENT_UPTODATE)
2910                 return 0;
2911
2912         if (0 && test_range_bit(tree, eb->start, eb->start + eb->len - 1,
2913                            EXTENT_UPTODATE, 1)) {
2914                 return 0;
2915         }
2916
2917         if (start) {
2918                 WARN_ON(start < eb->start);
2919                 start_i = (start >> PAGE_CACHE_SHIFT) -
2920                         (eb->start >> PAGE_CACHE_SHIFT);
2921         } else {
2922                 start_i = 0;
2923         }
2924
2925         num_pages = num_extent_pages(eb->start, eb->len);
2926         for (i = start_i; i < num_pages; i++) {
2927                 page = extent_buffer_page(eb, i);
2928                 if (PageUptodate(page)) {
2929                         continue;
2930                 }
2931                 if (!wait) {
2932                         if (TestSetPageLocked(page)) {
2933                                 continue;
2934                         }
2935                 } else {
2936                         lock_page(page);
2937                 }
2938                 if (!PageUptodate(page)) {
2939                         err = __extent_read_full_page(tree, page,
2940                                                       get_extent, &bio);
2941                         if (err) {
2942                                 ret = err;
2943                         }
2944                 } else {
2945                         unlock_page(page);
2946                 }
2947         }
2948
2949         if (bio)
2950                 submit_one_bio(READ, bio);
2951
2952         if (ret || !wait) {
2953                 return ret;
2954         }
2955         for (i = start_i; i < num_pages; i++) {
2956                 page = extent_buffer_page(eb, i);
2957                 wait_on_page_locked(page);
2958                 if (!PageUptodate(page)) {
2959                         ret = -EIO;
2960                 }
2961         }
2962         if (!ret)
2963                 eb->flags |= EXTENT_UPTODATE;
2964         return ret;
2965 }
2966 EXPORT_SYMBOL(read_extent_buffer_pages);
2967
2968 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
2969                         unsigned long start,
2970                         unsigned long len)
2971 {
2972         size_t cur;
2973         size_t offset;
2974         struct page *page;
2975         char *kaddr;
2976         char *dst = (char *)dstv;
2977         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2978         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2979         unsigned long num_pages = num_extent_pages(eb->start, eb->len);
2980
2981         WARN_ON(start > eb->len);
2982         WARN_ON(start + len > eb->start + eb->len);
2983
2984         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
2985
2986         while(len > 0) {
2987                 page = extent_buffer_page(eb, i);
2988                 if (!PageUptodate(page)) {
2989                         printk("page %lu not up to date i %lu, total %lu, len %lu\n", page->index, i, num_pages, eb->len);
2990                         WARN_ON(1);
2991                 }
2992                 WARN_ON(!PageUptodate(page));
2993
2994                 cur = min(len, (PAGE_CACHE_SIZE - offset));
2995                 kaddr = kmap_atomic(page, KM_USER1);
2996                 memcpy(dst, kaddr + offset, cur);
2997                 kunmap_atomic(kaddr, KM_USER1);
2998
2999                 dst += cur;
3000                 len -= cur;
3001                 offset = 0;
3002                 i++;
3003         }
3004 }
3005 EXPORT_SYMBOL(read_extent_buffer);
3006
3007 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3008                                unsigned long min_len, char **token, char **map,
3009                                unsigned long *map_start,
3010                                unsigned long *map_len, int km)
3011 {
3012         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3013         char *kaddr;
3014         struct page *p;
3015         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3016         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3017         unsigned long end_i = (start_offset + start + min_len - 1) >>
3018                 PAGE_CACHE_SHIFT;
3019
3020         if (i != end_i)
3021                 return -EINVAL;
3022
3023         if (i == 0) {
3024                 offset = start_offset;
3025                 *map_start = 0;
3026         } else {
3027                 offset = 0;
3028                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3029         }
3030         if (start + min_len > eb->len) {
3031 printk("bad mapping eb start %Lu len %lu, wanted %lu %lu\n", eb->start, eb->len, start, min_len);
3032                 WARN_ON(1);
3033         }
3034
3035         p = extent_buffer_page(eb, i);
3036         WARN_ON(!PageUptodate(p));
3037         kaddr = kmap_atomic(p, km);
3038         *token = kaddr;
3039         *map = kaddr + offset;
3040         *map_len = PAGE_CACHE_SIZE - offset;
3041         return 0;
3042 }
3043 EXPORT_SYMBOL(map_private_extent_buffer);
3044
3045 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3046                       unsigned long min_len,
3047                       char **token, char **map,
3048                       unsigned long *map_start,
3049                       unsigned long *map_len, int km)
3050 {
3051         int err;
3052         int save = 0;
3053         if (eb->map_token) {
3054                 unmap_extent_buffer(eb, eb->map_token, km);
3055                 eb->map_token = NULL;
3056                 save = 1;
3057         }
3058         err = map_private_extent_buffer(eb, start, min_len, token, map,
3059                                        map_start, map_len, km);
3060         if (!err && save) {
3061                 eb->map_token = *token;
3062                 eb->kaddr = *map;
3063                 eb->map_start = *map_start;
3064                 eb->map_len = *map_len;
3065         }
3066         return err;
3067 }
3068 EXPORT_SYMBOL(map_extent_buffer);
3069
3070 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3071 {
3072         kunmap_atomic(token, km);
3073 }
3074 EXPORT_SYMBOL(unmap_extent_buffer);
3075
3076 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3077                           unsigned long start,
3078                           unsigned long len)
3079 {
3080         size_t cur;
3081         size_t offset;
3082         struct page *page;
3083         char *kaddr;
3084         char *ptr = (char *)ptrv;
3085         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3086         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3087         int ret = 0;
3088
3089         WARN_ON(start > eb->len);
3090         WARN_ON(start + len > eb->start + eb->len);
3091
3092         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3093
3094         while(len > 0) {
3095                 page = extent_buffer_page(eb, i);
3096                 WARN_ON(!PageUptodate(page));
3097
3098                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3099
3100                 kaddr = kmap_atomic(page, KM_USER0);
3101                 ret = memcmp(ptr, kaddr + offset, cur);
3102                 kunmap_atomic(kaddr, KM_USER0);
3103                 if (ret)
3104                         break;
3105
3106                 ptr += cur;
3107                 len -= cur;
3108                 offset = 0;
3109                 i++;
3110         }
3111         return ret;
3112 }
3113 EXPORT_SYMBOL(memcmp_extent_buffer);
3114
3115 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3116                          unsigned long start, unsigned long len)
3117 {
3118         size_t cur;
3119         size_t offset;
3120         struct page *page;
3121         char *kaddr;
3122         char *src = (char *)srcv;
3123         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3124         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3125
3126         WARN_ON(start > eb->len);
3127         WARN_ON(start + len > eb->start + eb->len);
3128
3129         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3130
3131         while(len > 0) {
3132                 page = extent_buffer_page(eb, i);
3133                 WARN_ON(!PageUptodate(page));
3134
3135                 cur = min(len, PAGE_CACHE_SIZE - offset);
3136                 kaddr = kmap_atomic(page, KM_USER1);
3137                 memcpy(kaddr + offset, src, cur);
3138                 kunmap_atomic(kaddr, KM_USER1);
3139
3140                 src += cur;
3141                 len -= cur;
3142                 offset = 0;
3143                 i++;
3144         }
3145 }
3146 EXPORT_SYMBOL(write_extent_buffer);
3147
3148 void memset_extent_buffer(struct extent_buffer *eb, char c,
3149                           unsigned long start, unsigned long len)
3150 {
3151         size_t cur;
3152         size_t offset;
3153         struct page *page;
3154         char *kaddr;
3155         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3156         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3157
3158         WARN_ON(start > eb->len);
3159         WARN_ON(start + len > eb->start + eb->len);
3160
3161         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3162
3163         while(len > 0) {
3164                 page = extent_buffer_page(eb, i);
3165                 WARN_ON(!PageUptodate(page));
3166
3167                 cur = min(len, PAGE_CACHE_SIZE - offset);
3168                 kaddr = kmap_atomic(page, KM_USER0);
3169                 memset(kaddr + offset, c, cur);
3170                 kunmap_atomic(kaddr, KM_USER0);
3171
3172                 len -= cur;
3173                 offset = 0;
3174                 i++;
3175         }
3176 }
3177 EXPORT_SYMBOL(memset_extent_buffer);
3178
3179 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3180                         unsigned long dst_offset, unsigned long src_offset,
3181                         unsigned long len)
3182 {
3183         u64 dst_len = dst->len;
3184         size_t cur;
3185         size_t offset;
3186         struct page *page;
3187         char *kaddr;
3188         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3189         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3190
3191         WARN_ON(src->len != dst_len);
3192
3193         offset = (start_offset + dst_offset) &
3194                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3195
3196         while(len > 0) {
3197                 page = extent_buffer_page(dst, i);
3198                 WARN_ON(!PageUptodate(page));
3199
3200                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3201
3202                 kaddr = kmap_atomic(page, KM_USER0);
3203                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3204                 kunmap_atomic(kaddr, KM_USER0);
3205
3206                 src_offset += cur;
3207                 len -= cur;
3208                 offset = 0;
3209                 i++;
3210         }
3211 }
3212 EXPORT_SYMBOL(copy_extent_buffer);
3213
3214 static void move_pages(struct page *dst_page, struct page *src_page,
3215                        unsigned long dst_off, unsigned long src_off,
3216                        unsigned long len)
3217 {
3218         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3219         if (dst_page == src_page) {
3220                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3221         } else {
3222                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3223                 char *p = dst_kaddr + dst_off + len;
3224                 char *s = src_kaddr + src_off + len;
3225
3226                 while (len--)
3227                         *--p = *--s;
3228
3229                 kunmap_atomic(src_kaddr, KM_USER1);
3230         }
3231         kunmap_atomic(dst_kaddr, KM_USER0);
3232 }
3233
3234 static void copy_pages(struct page *dst_page, struct page *src_page,
3235                        unsigned long dst_off, unsigned long src_off,
3236                        unsigned long len)
3237 {
3238         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3239         char *src_kaddr;
3240
3241         if (dst_page != src_page)
3242                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3243         else
3244                 src_kaddr = dst_kaddr;
3245
3246         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3247         kunmap_atomic(dst_kaddr, KM_USER0);
3248         if (dst_page != src_page)
3249                 kunmap_atomic(src_kaddr, KM_USER1);
3250 }
3251
3252 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3253                            unsigned long src_offset, unsigned long len)
3254 {
3255         size_t cur;
3256         size_t dst_off_in_page;
3257         size_t src_off_in_page;
3258         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3259         unsigned long dst_i;
3260         unsigned long src_i;
3261
3262         if (src_offset + len > dst->len) {
3263                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
3264                        src_offset, len, dst->len);
3265                 BUG_ON(1);
3266         }
3267         if (dst_offset + len > dst->len) {
3268                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
3269                        dst_offset, len, dst->len);
3270                 BUG_ON(1);
3271         }
3272
3273         while(len > 0) {
3274                 dst_off_in_page = (start_offset + dst_offset) &
3275                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3276                 src_off_in_page = (start_offset + src_offset) &
3277                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3278
3279                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3280                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3281
3282                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3283                                                src_off_in_page));
3284                 cur = min_t(unsigned long, cur,
3285                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3286
3287                 copy_pages(extent_buffer_page(dst, dst_i),
3288                            extent_buffer_page(dst, src_i),
3289                            dst_off_in_page, src_off_in_page, cur);
3290
3291                 src_offset += cur;
3292                 dst_offset += cur;
3293                 len -= cur;
3294         }
3295 }
3296 EXPORT_SYMBOL(memcpy_extent_buffer);
3297
3298 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3299                            unsigned long src_offset, unsigned long len)
3300 {
3301         size_t cur;
3302         size_t dst_off_in_page;
3303         size_t src_off_in_page;
3304         unsigned long dst_end = dst_offset + len - 1;
3305         unsigned long src_end = src_offset + len - 1;
3306         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3307         unsigned long dst_i;
3308         unsigned long src_i;
3309
3310         if (src_offset + len > dst->len) {
3311                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
3312                        src_offset, len, dst->len);
3313                 BUG_ON(1);
3314         }
3315         if (dst_offset + len > dst->len) {
3316                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
3317                        dst_offset, len, dst->len);
3318                 BUG_ON(1);
3319         }
3320         if (dst_offset < src_offset) {
3321                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3322                 return;
3323         }
3324         while(len > 0) {
3325                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3326                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3327
3328                 dst_off_in_page = (start_offset + dst_end) &
3329                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3330                 src_off_in_page = (start_offset + src_end) &
3331                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3332
3333                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3334                 cur = min(cur, dst_off_in_page + 1);
3335                 move_pages(extent_buffer_page(dst, dst_i),
3336                            extent_buffer_page(dst, src_i),
3337                            dst_off_in_page - cur + 1,
3338                            src_off_in_page - cur + 1, cur);
3339
3340                 dst_end -= cur;
3341                 src_end -= cur;
3342                 len -= cur;
3343         }
3344 }
3345 EXPORT_SYMBOL(memmove_extent_buffer);