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/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/module.h>
   8 #include <linux/spinlock.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/swap.h>
  11 #include <linux/writeback.h>
  12 #include <linux/pagevec.h>
  13 #include "extent_io.h"
  14 #include "extent_map.h"
  15 #include "compat.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18
  19 static struct kmem_cache *extent_state_cache;
  20 static struct kmem_cache *extent_buffer_cache;
  21
  22 static LIST_HEAD(buffers);
  23 static LIST_HEAD(states);
  24
  25 #define LEAK_DEBUG 0
  26 #if LEAK_DEBUG
  27 static DEFINE_SPINLOCK(leak_lock);
  28 #endif
  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         /* tells writepage not to lock the state bits for this range
  44          * it still does the unlocking
  45          */
  46         unsigned int extent_locked:1;
  47
  48         /* tells the submit_bio code to use a WRITE_SYNC */
  49         unsigned int sync_io:1;
  50 };
  51
  52 int __init extent_io_init(void)
  53 {
  54         extent_state_cache = kmem_cache_create("extent_state",
  55                         sizeof(struct extent_state), 0,
  56                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  57         if (!extent_state_cache)
  58                 return -ENOMEM;
  59
  60         extent_buffer_cache = kmem_cache_create("extent_buffers",
  61                         sizeof(struct extent_buffer), 0,
  62                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  63         if (!extent_buffer_cache)
  64                 goto free_state_cache;
  65         return 0;
  66
  67 free_state_cache:
  68         kmem_cache_destroy(extent_state_cache);
  69         return -ENOMEM;
  70 }
  71
  72 void extent_io_exit(void)
  73 {
  74         struct extent_state *state;
  75         struct extent_buffer *eb;
  76
  77         while (!list_empty(&states)) {
  78                 state = list_entry(states.next, struct extent_state, leak_list);
  79                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  80                        "state %lu in tree %p refs %d\n",
  81                        (unsigned long long)state->start,
  82                        (unsigned long long)state->end,
  83                        state->state, state->tree, atomic_read(&state->refs));
  84                 list_del(&state->leak_list);
  85                 kmem_cache_free(extent_state_cache, state);
  86
  87         }
  88
  89         while (!list_empty(&buffers)) {
  90                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  91                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  92                        "refs %d\n", (unsigned long long)eb->start,
  93                        eb->len, atomic_read(&eb->refs));
  94                 list_del(&eb->leak_list);
  95                 kmem_cache_free(extent_buffer_cache, eb);
  96         }
  97         if (extent_state_cache)
  98                 kmem_cache_destroy(extent_state_cache);
  99         if (extent_buffer_cache)
 100                 kmem_cache_destroy(extent_buffer_cache);
 101 }
 102
 103 void extent_io_tree_init(struct extent_io_tree *tree,
 104                           struct address_space *mapping, gfp_t mask)
 105 {
 106         tree->state = RB_ROOT;
 107         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 108         tree->ops = NULL;
 109         tree->dirty_bytes = 0;
 110         spin_lock_init(&tree->lock);
 111         spin_lock_init(&tree->buffer_lock);
 112         tree->mapping = mapping;
 113 }
 114
 115 static struct extent_state *alloc_extent_state(gfp_t mask)
 116 {
 117         struct extent_state *state;
 118 #if LEAK_DEBUG
 119         unsigned long flags;
 120 #endif
 121
 122         state = kmem_cache_alloc(extent_state_cache, mask);
 123         if (!state)
 124                 return state;
 125         state->state = 0;
 126         state->private = 0;
 127         state->tree = NULL;
 128 #if LEAK_DEBUG
 129         spin_lock_irqsave(&leak_lock, flags);
 130         list_add(&state->leak_list, &states);
 131         spin_unlock_irqrestore(&leak_lock, flags);
 132 #endif
 133         atomic_set(&state->refs, 1);
 134         init_waitqueue_head(&state->wq);
 135         return state;
 136 }
 137
 138 void free_extent_state(struct extent_state *state)
 139 {
 140         if (!state)
 141                 return;
 142         if (atomic_dec_and_test(&state->refs)) {
 143 #if LEAK_DEBUG
 144                 unsigned long flags;
 145 #endif
 146                 WARN_ON(state->tree);
 147 #if LEAK_DEBUG
 148                 spin_lock_irqsave(&leak_lock, flags);
 149                 list_del(&state->leak_list);
 150                 spin_unlock_irqrestore(&leak_lock, flags);
 151 #endif
 152                 kmem_cache_free(extent_state_cache, state);
 153         }
 154 }
 155
 156 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 157                                    struct rb_node *node)
 158 {
 159         struct rb_node **p = &root->rb_node;
 160         struct rb_node *parent = NULL;
 161         struct tree_entry *entry;
 162
 163         while (*p) {
 164                 parent = *p;
 165                 entry = rb_entry(parent, struct tree_entry, rb_node);
 166
 167                 if (offset < entry->start)
 168                         p = &(*p)->rb_left;
 169                 else if (offset > entry->end)
 170                         p = &(*p)->rb_right;
 171                 else
 172                         return parent;
 173         }
 174
 175         entry = rb_entry(node, struct tree_entry, rb_node);
 176         rb_link_node(node, parent, p);
 177         rb_insert_color(node, root);
 178         return NULL;
 179 }
 180
 181 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 182                                      struct rb_node **prev_ret,
 183                                      struct rb_node **next_ret)
 184 {
 185         struct rb_root *root = &tree->state;
 186         struct rb_node *n = root->rb_node;
 187         struct rb_node *prev = NULL;
 188         struct rb_node *orig_prev = NULL;
 189         struct tree_entry *entry;
 190         struct tree_entry *prev_entry = NULL;
 191
 192         while (n) {
 193                 entry = rb_entry(n, struct tree_entry, rb_node);
 194                 prev = n;
 195                 prev_entry = entry;
 196
 197                 if (offset < entry->start)
 198                         n = n->rb_left;
 199                 else if (offset > entry->end)
 200                         n = n->rb_right;
 201                 else
 202                         return n;
 203         }
 204
 205         if (prev_ret) {
 206                 orig_prev = prev;
 207                 while (prev && offset > prev_entry->end) {
 208                         prev = rb_next(prev);
 209                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 210                 }
 211                 *prev_ret = prev;
 212                 prev = orig_prev;
 213         }
 214
 215         if (next_ret) {
 216                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 217                 while (prev && offset < prev_entry->start) {
 218                         prev = rb_prev(prev);
 219                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 220                 }
 221                 *next_ret = prev;
 222         }
 223         return NULL;
 224 }
 225
 226 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 227                                           u64 offset)
 228 {
 229         struct rb_node *prev = NULL;
 230         struct rb_node *ret;
 231
 232         ret = __etree_search(tree, offset, &prev, NULL);
 233         if (!ret)
 234                 return prev;
 235         return ret;
 236 }
 237
 238 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 239                      struct extent_state *other)
 240 {
 241         if (tree->ops && tree->ops->merge_extent_hook)
 242                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 243                                              other);
 244 }
 245
 246 /*
 247  * utility function to look for merge candidates inside a given range.
 248  * Any extents with matching state are merged together into a single
 249  * extent in the tree.  Extents with EXTENT_IO in their state field
 250  * are not merged because the end_io handlers need to be able to do
 251  * operations on them without sleeping (or doing allocations/splits).
 252  *
 253  * This should be called with the tree lock held.
 254  */
 255 static int merge_state(struct extent_io_tree *tree,
 256                        struct extent_state *state)
 257 {
 258         struct extent_state *other;
 259         struct rb_node *other_node;
 260
 261         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 262                 return 0;
 263
 264         other_node = rb_prev(&state->rb_node);
 265         if (other_node) {
 266                 other = rb_entry(other_node, struct extent_state, rb_node);
 267                 if (other->end == state->start - 1 &&
 268                     other->state == state->state) {
 269                         merge_cb(tree, state, other);
 270                         state->start = other->start;
 271                         other->tree = NULL;
 272                         rb_erase(&other->rb_node, &tree->state);
 273                         free_extent_state(other);
 274                 }
 275         }
 276         other_node = rb_next(&state->rb_node);
 277         if (other_node) {
 278                 other = rb_entry(other_node, struct extent_state, rb_node);
 279                 if (other->start == state->end + 1 &&
 280                     other->state == state->state) {
 281                         merge_cb(tree, state, other);
 282                         other->start = state->start;
 283                         state->tree = NULL;
 284                         rb_erase(&state->rb_node, &tree->state);
 285                         free_extent_state(state);
 286                         state = NULL;
 287                 }
 288         }
 289
 290         return 0;
 291 }
 292
 293 static int set_state_cb(struct extent_io_tree *tree,
 294                          struct extent_state *state, int *bits)
 295 {
 296         if (tree->ops && tree->ops->set_bit_hook) {
 297                 return tree->ops->set_bit_hook(tree->mapping->host,
 298                                                state, bits);
 299         }
 300
 301         return 0;
 302 }
 303
 304 static void clear_state_cb(struct extent_io_tree *tree,
 305                            struct extent_state *state, int *bits)
 306 {
 307         if (tree->ops && tree->ops->clear_bit_hook)
 308                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 309 }
 310
 311 /*
 312  * insert an extent_state struct into the tree.  'bits' are set on the
 313  * struct before it is inserted.
 314  *
 315  * This may return -EEXIST if the extent is already there, in which case the
 316  * state struct is freed.
 317  *
 318  * The tree lock is not taken internally.  This is a utility function and
 319  * probably isn't what you want to call (see set/clear_extent_bit).
 320  */
 321 static int insert_state(struct extent_io_tree *tree,
 322                         struct extent_state *state, u64 start, u64 end,
 323                         int *bits)
 324 {
 325         struct rb_node *node;
 326         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 327         int ret;
 328
 329         if (end < start) {
 330                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
 331                        (unsigned long long)end,
 332                        (unsigned long long)start);
 333                 WARN_ON(1);
 334         }
 335         state->start = start;
 336         state->end = end;
 337         ret = set_state_cb(tree, state, bits);
 338         if (ret)
 339                 return ret;
 340
 341         if (bits_to_set & EXTENT_DIRTY)
 342                 tree->dirty_bytes += end - start + 1;
 343         state->state |= bits_to_set;
 344         node = tree_insert(&tree->state, end, &state->rb_node);
 345         if (node) {
 346                 struct extent_state *found;
 347                 found = rb_entry(node, struct extent_state, rb_node);
 348                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 349                        "%llu %llu\n", (unsigned long long)found->start,
 350                        (unsigned long long)found->end,
 351                        (unsigned long long)start, (unsigned long long)end);
 352                 free_extent_state(state);
 353                 return -EEXIST;
 354         }
 355         state->tree = tree;
 356         merge_state(tree, state);
 357         return 0;
 358 }
 359
 360 static int split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 361                      u64 split)
 362 {
 363         if (tree->ops && tree->ops->split_extent_hook)
 364                 return tree->ops->split_extent_hook(tree->mapping->host,
 365                                                     orig, split);
 366         return 0;
 367 }
 368
 369 /*
 370  * split a given extent state struct in two, inserting the preallocated
 371  * struct 'prealloc' as the newly created second half.  'split' indicates an
 372  * offset inside 'orig' where it should be split.
 373  *
 374  * Before calling,
 375  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 376  * are two extent state structs in the tree:
 377  * prealloc: [orig->start, split - 1]
 378  * orig: [ split, orig->end ]
 379  *
 380  * The tree locks are not taken by this function. They need to be held
 381  * by the caller.
 382  */
 383 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 384                        struct extent_state *prealloc, u64 split)
 385 {
 386         struct rb_node *node;
 387
 388         split_cb(tree, orig, split);
 389
 390         prealloc->start = orig->start;
 391         prealloc->end = split - 1;
 392         prealloc->state = orig->state;
 393         orig->start = split;
 394
 395         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 396         if (node) {
 397                 free_extent_state(prealloc);
 398                 return -EEXIST;
 399         }
 400         prealloc->tree = tree;
 401         return 0;
 402 }
 403
 404 /*
 405  * utility function to clear some bits in an extent state struct.
 406  * it will optionally wake up any one waiting on this state (wake == 1), or
 407  * forcibly remove the state from the tree (delete == 1).
 408  *
 409  * If no bits are set on the state struct after clearing things, the
 410  * struct is freed and removed from the tree
 411  */
 412 static int clear_state_bit(struct extent_io_tree *tree,
 413                             struct extent_state *state,
 414                             int *bits, int wake)
 415 {
 416         int bits_to_clear = *bits & ~EXTENT_CTLBITS;
 417         int ret = state->state & bits_to_clear;
 418
 419         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 420                 u64 range = state->end - state->start + 1;
 421                 WARN_ON(range > tree->dirty_bytes);
 422                 tree->dirty_bytes -= range;
 423         }
 424         clear_state_cb(tree, state, bits);
 425         state->state &= ~bits_to_clear;
 426         if (wake)
 427                 wake_up(&state->wq);
 428         if (state->state == 0) {
 429                 if (state->tree) {
 430                         rb_erase(&state->rb_node, &tree->state);
 431                         state->tree = NULL;
 432                         free_extent_state(state);
 433                 } else {
 434                         WARN_ON(1);
 435                 }
 436         } else {
 437                 merge_state(tree, state);
 438         }
 439         return ret;
 440 }
 441
 442 /*
 443  * clear some bits on a range in the tree.  This may require splitting
 444  * or inserting elements in the tree, so the gfp mask is used to
 445  * indicate which allocations or sleeping are allowed.
 446  *
 447  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 448  * the given range from the tree regardless of state (ie for truncate).
 449  *
 450  * the range [start, end] is inclusive.
 451  *
 452  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 453  * bits were already set, or zero if none of the bits were already set.
 454  */
 455 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 456                      int bits, int wake, int delete,
 457                      struct extent_state **cached_state,
 458                      gfp_t mask)
 459 {
 460         struct extent_state *state;
 461         struct extent_state *cached;
 462         struct extent_state *prealloc = NULL;
 463         struct rb_node *next_node;
 464         struct rb_node *node;
 465         u64 last_end;
 466         int err;
 467         int set = 0;
 468         int clear = 0;
 469
 470         if (delete)
 471                 bits |= ~EXTENT_CTLBITS;
 472         bits |= EXTENT_FIRST_DELALLOC;
 473
 474         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 475                 clear = 1;
 476 again:
 477         if (!prealloc && (mask & __GFP_WAIT)) {
 478                 prealloc = alloc_extent_state(mask);
 479                 if (!prealloc)
 480                         return -ENOMEM;
 481         }
 482
 483         spin_lock(&tree->lock);
 484         if (cached_state) {
 485                 cached = *cached_state;
 486
 487                 if (clear) {
 488                         *cached_state = NULL;
 489                         cached_state = NULL;
 490                 }
 491
 492                 if (cached && cached->tree && cached->start == start) {
 493                         if (clear)
 494                                 atomic_dec(&cached->refs);
 495                         state = cached;
 496                         goto hit_next;
 497                 }
 498                 if (clear)
 499                         free_extent_state(cached);
 500         }
 501         /*
 502          * this search will find the extents that end after
 503          * our range starts
 504          */
 505         node = tree_search(tree, start);
 506         if (!node)
 507                 goto out;
 508         state = rb_entry(node, struct extent_state, rb_node);
 509 hit_next:
 510         if (state->start > end)
 511                 goto out;
 512         WARN_ON(state->end < start);
 513         last_end = state->end;
 514
 515         /*
 516          *     | ---- desired range ---- |
 517          *  | state | or
 518          *  | ------------- state -------------- |
 519          *
 520          * We need to split the extent we found, and may flip
 521          * bits on second half.
 522          *
 523          * If the extent we found extends past our range, we
 524          * just split and search again.  It'll get split again
 525          * the next time though.
 526          *
 527          * If the extent we found is inside our range, we clear
 528          * the desired bit on it.
 529          */
 530
 531         if (state->start < start) {
 532                 if (!prealloc)
 533                         prealloc = alloc_extent_state(GFP_ATOMIC);
 534                 err = split_state(tree, state, prealloc, start);
 535                 BUG_ON(err == -EEXIST);
 536                 prealloc = NULL;
 537                 if (err)
 538                         goto out;
 539                 if (state->end <= end) {
 540                         set |= clear_state_bit(tree, state, &bits, wake);
 541                         if (last_end == (u64)-1)
 542                                 goto out;
 543                         start = last_end + 1;
 544                 }
 545                 goto search_again;
 546         }
 547         /*
 548          * | ---- desired range ---- |
 549          *                        | state |
 550          * We need to split the extent, and clear the bit
 551          * on the first half
 552          */
 553         if (state->start <= end && state->end > end) {
 554                 if (!prealloc)
 555                         prealloc = alloc_extent_state(GFP_ATOMIC);
 556                 err = split_state(tree, state, prealloc, end + 1);
 557                 BUG_ON(err == -EEXIST);
 558                 if (wake)
 559                         wake_up(&state->wq);
 560
 561                 set |= clear_state_bit(tree, prealloc, &bits, wake);
 562
 563                 prealloc = NULL;
 564                 goto out;
 565         }
 566
 567         if (state->end < end && prealloc && !need_resched())
 568                 next_node = rb_next(&state->rb_node);
 569         else
 570                 next_node = NULL;
 571
 572         set |= clear_state_bit(tree, state, &bits, wake);
 573         if (last_end == (u64)-1)
 574                 goto out;
 575         start = last_end + 1;
 576         if (start <= end && next_node) {
 577                 state = rb_entry(next_node, struct extent_state,
 578                                  rb_node);
 579                 if (state->start == start)
 580                         goto hit_next;
 581         }
 582         goto search_again;
 583
 584 out:
 585         spin_unlock(&tree->lock);
 586         if (prealloc)
 587                 free_extent_state(prealloc);
 588
 589         return set;
 590
 591 search_again:
 592         if (start > end)
 593                 goto out;
 594         spin_unlock(&tree->lock);
 595         if (mask & __GFP_WAIT)
 596                 cond_resched();
 597         goto again;
 598 }
 599
 600 static int wait_on_state(struct extent_io_tree *tree,
 601                          struct extent_state *state)
 602                 __releases(tree->lock)
 603                 __acquires(tree->lock)
 604 {
 605         DEFINE_WAIT(wait);
 606         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 607         spin_unlock(&tree->lock);
 608         schedule();
 609         spin_lock(&tree->lock);
 610         finish_wait(&state->wq, &wait);
 611         return 0;
 612 }
 613
 614 /*
 615  * waits for one or more bits to clear on a range in the state tree.
 616  * The range [start, end] is inclusive.
 617  * The tree lock is taken by this function
 618  */
 619 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 620 {
 621         struct extent_state *state;
 622         struct rb_node *node;
 623
 624         spin_lock(&tree->lock);
 625 again:
 626         while (1) {
 627                 /*
 628                  * this search will find all the extents that end after
 629                  * our range starts
 630                  */
 631                 node = tree_search(tree, start);
 632                 if (!node)
 633                         break;
 634
 635                 state = rb_entry(node, struct extent_state, rb_node);
 636
 637                 if (state->start > end)
 638                         goto out;
 639
 640                 if (state->state & bits) {
 641                         start = state->start;
 642                         atomic_inc(&state->refs);
 643                         wait_on_state(tree, state);
 644                         free_extent_state(state);
 645                         goto again;
 646                 }
 647                 start = state->end + 1;
 648
 649                 if (start > end)
 650                         break;
 651
 652                 if (need_resched()) {
 653                         spin_unlock(&tree->lock);
 654                         cond_resched();
 655                         spin_lock(&tree->lock);
 656                 }
 657         }
 658 out:
 659         spin_unlock(&tree->lock);
 660         return 0;
 661 }
 662
 663 static int set_state_bits(struct extent_io_tree *tree,
 664                            struct extent_state *state,
 665                            int *bits)
 666 {
 667         int ret;
 668         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 669
 670         ret = set_state_cb(tree, state, bits);
 671         if (ret)
 672                 return ret;
 673         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 674                 u64 range = state->end - state->start + 1;
 675                 tree->dirty_bytes += range;
 676         }
 677         state->state |= bits_to_set;
 678
 679         return 0;
 680 }
 681
 682 static void cache_state(struct extent_state *state,
 683                         struct extent_state **cached_ptr)
 684 {
 685         if (cached_ptr && !(*cached_ptr)) {
 686                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 687                         *cached_ptr = state;
 688                         atomic_inc(&state->refs);
 689                 }
 690         }
 691 }
 692
 693 /*
 694  * set some bits on a range in the tree.  This may require allocations or
 695  * sleeping, so the gfp mask is used to indicate what is allowed.
 696  *
 697  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 698  * part of the range already has the desired bits set.  The start of the
 699  * existing range is returned in failed_start in this case.
 700  *
 701  * [start, end] is inclusive This takes the tree lock.
 702  */
 703
 704 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 705                    int bits, int exclusive_bits, u64 *failed_start,
 706                    struct extent_state **cached_state, gfp_t mask)
 707 {
 708         struct extent_state *state;
 709         struct extent_state *prealloc = NULL;
 710         struct rb_node *node;
 711         int err = 0;
 712         u64 last_start;
 713         u64 last_end;
 714
 715         bits |= EXTENT_FIRST_DELALLOC;
 716 again:
 717         if (!prealloc && (mask & __GFP_WAIT)) {
 718                 prealloc = alloc_extent_state(mask);
 719                 if (!prealloc)
 720                         return -ENOMEM;
 721         }
 722
 723         spin_lock(&tree->lock);
 724         if (cached_state && *cached_state) {
 725                 state = *cached_state;
 726                 if (state->start == start && state->tree) {
 727                         node = &state->rb_node;
 728                         goto hit_next;
 729                 }
 730         }
 731         /*
 732          * this search will find all the extents that end after
 733          * our range starts.
 734          */
 735         node = tree_search(tree, start);
 736         if (!node) {
 737                 err = insert_state(tree, prealloc, start, end, &bits);
 738                 prealloc = NULL;
 739                 BUG_ON(err == -EEXIST);
 740                 goto out;
 741         }
 742         state = rb_entry(node, struct extent_state, rb_node);
 743 hit_next:
 744         last_start = state->start;
 745         last_end = state->end;
 746
 747         /*
 748          * | ---- desired range ---- |
 749          * | state |
 750          *
 751          * Just lock what we found and keep going
 752          */
 753         if (state->start == start && state->end <= end) {
 754                 struct rb_node *next_node;
 755                 if (state->state & exclusive_bits) {
 756                         *failed_start = state->start;
 757                         err = -EEXIST;
 758                         goto out;
 759                 }
 760
 761                 err = set_state_bits(tree, state, &bits);
 762                 if (err)
 763                         goto out;
 764
 765                 cache_state(state, cached_state);
 766                 merge_state(tree, state);
 767                 if (last_end == (u64)-1)
 768                         goto out;
 769
 770                 start = last_end + 1;
 771                 if (start < end && prealloc && !need_resched()) {
 772                         next_node = rb_next(node);
 773                         if (next_node) {
 774                                 state = rb_entry(next_node, struct extent_state,
 775                                                  rb_node);
 776                                 if (state->start == start)
 777                                         goto hit_next;
 778                         }
 779                 }
 780                 goto search_again;
 781         }
 782
 783         /*
 784          *     | ---- desired range ---- |
 785          * | state |
 786          *   or
 787          * | ------------- state -------------- |
 788          *
 789          * We need to split the extent we found, and may flip bits on
 790          * second half.
 791          *
 792          * If the extent we found extends past our
 793          * range, we just split and search again.  It'll get split
 794          * again the next time though.
 795          *
 796          * If the extent we found is inside our range, we set the
 797          * desired bit on it.
 798          */
 799         if (state->start < start) {
 800                 if (state->state & exclusive_bits) {
 801                         *failed_start = start;
 802                         err = -EEXIST;
 803                         goto out;
 804                 }
 805                 err = split_state(tree, state, prealloc, start);
 806                 BUG_ON(err == -EEXIST);
 807                 prealloc = NULL;
 808                 if (err)
 809                         goto out;
 810                 if (state->end <= end) {
 811                         err = set_state_bits(tree, state, &bits);
 812                         if (err)
 813                                 goto out;
 814                         cache_state(state, cached_state);
 815                         merge_state(tree, state);
 816                         if (last_end == (u64)-1)
 817                                 goto out;
 818                         start = last_end + 1;
 819                 }
 820                 goto search_again;
 821         }
 822         /*
 823          * | ---- desired range ---- |
 824          *     | state | or               | state |
 825          *
 826          * There's a hole, we need to insert something in it and
 827          * ignore the extent we found.
 828          */
 829         if (state->start > start) {
 830                 u64 this_end;
 831                 if (end < last_start)
 832                         this_end = end;
 833                 else
 834                         this_end = last_start - 1;
 835                 err = insert_state(tree, prealloc, start, this_end,
 836                                    &bits);
 837                 BUG_ON(err == -EEXIST);
 838                 if (err) {
 839                         prealloc = NULL;
 840                         goto out;
 841                 }
 842                 cache_state(prealloc, cached_state);
 843                 prealloc = NULL;
 844                 start = this_end + 1;
 845                 goto search_again;
 846         }
 847         /*
 848          * | ---- desired range ---- |
 849          *                        | state |
 850          * We need to split the extent, and set the bit
 851          * on the first half
 852          */
 853         if (state->start <= end && state->end > end) {
 854                 if (state->state & exclusive_bits) {
 855                         *failed_start = start;
 856                         err = -EEXIST;
 857                         goto out;
 858                 }
 859                 err = split_state(tree, state, prealloc, end + 1);
 860                 BUG_ON(err == -EEXIST);
 861
 862                 err = set_state_bits(tree, prealloc, &bits);
 863                 if (err) {
 864                         prealloc = NULL;
 865                         goto out;
 866                 }
 867                 cache_state(prealloc, cached_state);
 868                 merge_state(tree, prealloc);
 869                 prealloc = NULL;
 870                 goto out;
 871         }
 872
 873         goto search_again;
 874
 875 out:
 876         spin_unlock(&tree->lock);
 877         if (prealloc)
 878                 free_extent_state(prealloc);
 879
 880         return err;
 881
 882 search_again:
 883         if (start > end)
 884                 goto out;
 885         spin_unlock(&tree->lock);
 886         if (mask & __GFP_WAIT)
 887                 cond_resched();
 888         goto again;
 889 }
 890
 891 /* wrappers around set/clear extent bit */
 892 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 893                      gfp_t mask)
 894 {
 895         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 896                               NULL, mask);
 897 }
 898
 899 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 900                     int bits, gfp_t mask)
 901 {
 902         return set_extent_bit(tree, start, end, bits, 0, NULL,
 903                               NULL, mask);
 904 }
 905
 906 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 907                       int bits, gfp_t mask)
 908 {
 909         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
 910 }
 911
 912 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
 913                         struct extent_state **cached_state, gfp_t mask)
 914 {
 915         return set_extent_bit(tree, start, end,
 916                               EXTENT_DELALLOC | EXTENT_DIRTY | EXTENT_UPTODATE,
 917                               0, NULL, cached_state, mask);
 918 }
 919
 920 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 921                        gfp_t mask)
 922 {
 923         return clear_extent_bit(tree, start, end,
 924                                 EXTENT_DIRTY | EXTENT_DELALLOC |
 925                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
 926 }
 927
 928 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 929                      gfp_t mask)
 930 {
 931         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 932                               NULL, mask);
 933 }
 934
 935 static int clear_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 936                        gfp_t mask)
 937 {
 938         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0,
 939                                 NULL, mask);
 940 }
 941
 942 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 943                         gfp_t mask)
 944 {
 945         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, NULL,
 946                               NULL, mask);
 947 }
 948
 949 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
 950                                  u64 end, struct extent_state **cached_state,
 951                                  gfp_t mask)
 952 {
 953         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
 954                                 cached_state, mask);
 955 }
 956
 957 int wait_on_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 958 {
 959         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 960 }
 961
 962 /*
 963  * either insert or lock state struct between start and end use mask to tell
 964  * us if waiting is desired.
 965  */
 966 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 967                      int bits, struct extent_state **cached_state, gfp_t mask)
 968 {
 969         int err;
 970         u64 failed_start;
 971         while (1) {
 972                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
 973                                      EXTENT_LOCKED, &failed_start,
 974                                      cached_state, mask);
 975                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
 976                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
 977                         start = failed_start;
 978                 } else {
 979                         break;
 980                 }
 981                 WARN_ON(start > end);
 982         }
 983         return err;
 984 }
 985
 986 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
 987 {
 988         return lock_extent_bits(tree, start, end, 0, NULL, mask);
 989 }
 990
 991 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
 992                     gfp_t mask)
 993 {
 994         int err;
 995         u64 failed_start;
 996
 997         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
 998                              &failed_start, NULL, mask);
 999         if (err == -EEXIST) {
1000                 if (failed_start > start)
1001                         clear_extent_bit(tree, start, failed_start - 1,
1002                                          EXTENT_LOCKED, 1, 0, NULL, mask);
1003                 return 0;
1004         }
1005         return 1;
1006 }
1007
1008 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1009                          struct extent_state **cached, gfp_t mask)
1010 {
1011         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1012                                 mask);
1013 }
1014
1015 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1016                   gfp_t mask)
1017 {
1018         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1019                                 mask);
1020 }
1021
1022 /*
1023  * helper function to set pages and extents in the tree dirty
1024  */
1025 int set_range_dirty(struct extent_io_tree *tree, u64 start, u64 end)
1026 {
1027         unsigned long index = start >> PAGE_CACHE_SHIFT;
1028         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1029         struct page *page;
1030
1031         while (index <= end_index) {
1032                 page = find_get_page(tree->mapping, index);
1033                 BUG_ON(!page);
1034                 __set_page_dirty_nobuffers(page);
1035                 page_cache_release(page);
1036                 index++;
1037         }
1038         return 0;
1039 }
1040
1041 /*
1042  * helper function to set both pages and extents in the tree writeback
1043  */
1044 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1045 {
1046         unsigned long index = start >> PAGE_CACHE_SHIFT;
1047         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1048         struct page *page;
1049
1050         while (index <= end_index) {
1051                 page = find_get_page(tree->mapping, index);
1052                 BUG_ON(!page);
1053                 set_page_writeback(page);
1054                 page_cache_release(page);
1055                 index++;
1056         }
1057         return 0;
1058 }
1059
1060 /*
1061  * find the first offset in the io tree with 'bits' set. zero is
1062  * returned if we find something, and *start_ret and *end_ret are
1063  * set to reflect the state struct that was found.
1064  *
1065  * If nothing was found, 1 is returned, < 0 on error
1066  */
1067 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1068                           u64 *start_ret, u64 *end_ret, int bits)
1069 {
1070         struct rb_node *node;
1071         struct extent_state *state;
1072         int ret = 1;
1073
1074         spin_lock(&tree->lock);
1075         /*
1076          * this search will find all the extents that end after
1077          * our range starts.
1078          */
1079         node = tree_search(tree, start);
1080         if (!node)
1081                 goto out;
1082
1083         while (1) {
1084                 state = rb_entry(node, struct extent_state, rb_node);
1085                 if (state->end >= start && (state->state & bits)) {
1086                         *start_ret = state->start;
1087                         *end_ret = state->end;
1088                         ret = 0;
1089                         break;
1090                 }
1091                 node = rb_next(node);
1092                 if (!node)
1093                         break;
1094         }
1095 out:
1096         spin_unlock(&tree->lock);
1097         return ret;
1098 }
1099
1100 /* find the first state struct with 'bits' set after 'start', and
1101  * return it.  tree->lock must be held.  NULL will returned if
1102  * nothing was found after 'start'
1103  */
1104 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1105                                                  u64 start, int bits)
1106 {
1107         struct rb_node *node;
1108         struct extent_state *state;
1109
1110         /*
1111          * this search will find all the extents that end after
1112          * our range starts.
1113          */
1114         node = tree_search(tree, start);
1115         if (!node)
1116                 goto out;
1117
1118         while (1) {
1119                 state = rb_entry(node, struct extent_state, rb_node);
1120                 if (state->end >= start && (state->state & bits))
1121                         return state;
1122
1123                 node = rb_next(node);
1124                 if (!node)
1125                         break;
1126         }
1127 out:
1128         return NULL;
1129 }
1130
1131 /*
1132  * find a contiguous range of bytes in the file marked as delalloc, not
1133  * more than 'max_bytes'.  start and end are used to return the range,
1134  *
1135  * 1 is returned if we find something, 0 if nothing was in the tree
1136  */
1137 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1138                                         u64 *start, u64 *end, u64 max_bytes,
1139                                         struct extent_state **cached_state)
1140 {
1141         struct rb_node *node;
1142         struct extent_state *state;
1143         u64 cur_start = *start;
1144         u64 found = 0;
1145         u64 total_bytes = 0;
1146
1147         spin_lock(&tree->lock);
1148
1149         /*
1150          * this search will find all the extents that end after
1151          * our range starts.
1152          */
1153         node = tree_search(tree, cur_start);
1154         if (!node) {
1155                 if (!found)
1156                         *end = (u64)-1;
1157                 goto out;
1158         }
1159
1160         while (1) {
1161                 state = rb_entry(node, struct extent_state, rb_node);
1162                 if (found && (state->start != cur_start ||
1163                               (state->state & EXTENT_BOUNDARY))) {
1164                         goto out;
1165                 }
1166                 if (!(state->state & EXTENT_DELALLOC)) {
1167                         if (!found)
1168                                 *end = state->end;
1169                         goto out;
1170                 }
1171                 if (!found) {
1172                         *start = state->start;
1173                         *cached_state = state;
1174                         atomic_inc(&state->refs);
1175                 }
1176                 found++;
1177                 *end = state->end;
1178                 cur_start = state->end + 1;
1179                 node = rb_next(node);
1180                 if (!node)
1181                         break;
1182                 total_bytes += state->end - state->start + 1;
1183                 if (total_bytes >= max_bytes)
1184                         break;
1185         }
1186 out:
1187         spin_unlock(&tree->lock);
1188         return found;
1189 }
1190
1191 static noinline int __unlock_for_delalloc(struct inode *inode,
1192                                           struct page *locked_page,
1193                                           u64 start, u64 end)
1194 {
1195         int ret;
1196         struct page *pages[16];
1197         unsigned long index = start >> PAGE_CACHE_SHIFT;
1198         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1199         unsigned long nr_pages = end_index - index + 1;
1200         int i;
1201
1202         if (index == locked_page->index && end_index == index)
1203                 return 0;
1204
1205         while (nr_pages > 0) {
1206                 ret = find_get_pages_contig(inode->i_mapping, index,
1207                                      min_t(unsigned long, nr_pages,
1208                                      ARRAY_SIZE(pages)), pages);
1209                 for (i = 0; i < ret; i++) {
1210                         if (pages[i] != locked_page)
1211                                 unlock_page(pages[i]);
1212                         page_cache_release(pages[i]);
1213                 }
1214                 nr_pages -= ret;
1215                 index += ret;
1216                 cond_resched();
1217         }
1218         return 0;
1219 }
1220
1221 static noinline int lock_delalloc_pages(struct inode *inode,
1222                                         struct page *locked_page,
1223                                         u64 delalloc_start,
1224                                         u64 delalloc_end)
1225 {
1226         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1227         unsigned long start_index = index;
1228         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1229         unsigned long pages_locked = 0;
1230         struct page *pages[16];
1231         unsigned long nrpages;
1232         int ret;
1233         int i;
1234
1235         /* the caller is responsible for locking the start index */
1236         if (index == locked_page->index && index == end_index)
1237                 return 0;
1238
1239         /* skip the page at the start index */
1240         nrpages = end_index - index + 1;
1241         while (nrpages > 0) {
1242                 ret = find_get_pages_contig(inode->i_mapping, index,
1243                                      min_t(unsigned long,
1244                                      nrpages, ARRAY_SIZE(pages)), pages);
1245                 if (ret == 0) {
1246                         ret = -EAGAIN;
1247                         goto done;
1248                 }
1249                 /* now we have an array of pages, lock them all */
1250                 for (i = 0; i < ret; i++) {
1251                         /*
1252                          * the caller is taking responsibility for
1253                          * locked_page
1254                          */
1255                         if (pages[i] != locked_page) {
1256                                 lock_page(pages[i]);
1257                                 if (!PageDirty(pages[i]) ||
1258                                     pages[i]->mapping != inode->i_mapping) {
1259                                         ret = -EAGAIN;
1260                                         unlock_page(pages[i]);
1261                                         page_cache_release(pages[i]);
1262                                         goto done;
1263                                 }
1264                         }
1265                         page_cache_release(pages[i]);
1266                         pages_locked++;
1267                 }
1268                 nrpages -= ret;
1269                 index += ret;
1270                 cond_resched();
1271         }
1272         ret = 0;
1273 done:
1274         if (ret && pages_locked) {
1275                 __unlock_for_delalloc(inode, locked_page,
1276                               delalloc_start,
1277                               ((u64)(start_index + pages_locked - 1)) <<
1278                               PAGE_CACHE_SHIFT);
1279         }
1280         return ret;
1281 }
1282
1283 /*
1284  * find a contiguous range of bytes in the file marked as delalloc, not
1285  * more than 'max_bytes'.  start and end are used to return the range,
1286  *
1287  * 1 is returned if we find something, 0 if nothing was in the tree
1288  */
1289 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1290                                              struct extent_io_tree *tree,
1291                                              struct page *locked_page,
1292                                              u64 *start, u64 *end,
1293                                              u64 max_bytes)
1294 {
1295         u64 delalloc_start;
1296         u64 delalloc_end;
1297         u64 found;
1298         struct extent_state *cached_state = NULL;
1299         int ret;
1300         int loops = 0;
1301
1302 again:
1303         /* step one, find a bunch of delalloc bytes starting at start */
1304         delalloc_start = *start;
1305         delalloc_end = 0;
1306         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1307                                     max_bytes, &cached_state);
1308         if (!found || delalloc_end <= *start) {
1309                 *start = delalloc_start;
1310                 *end = delalloc_end;
1311                 free_extent_state(cached_state);
1312                 return found;
1313         }
1314
1315         /*
1316          * start comes from the offset of locked_page.  We have to lock
1317          * pages in order, so we can't process delalloc bytes before
1318          * locked_page
1319          */
1320         if (delalloc_start < *start)
1321                 delalloc_start = *start;
1322
1323         /*
1324          * make sure to limit the number of pages we try to lock down
1325          * if we're looping.
1326          */
1327         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1328                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1329
1330         /* step two, lock all the pages after the page that has start */
1331         ret = lock_delalloc_pages(inode, locked_page,
1332                                   delalloc_start, delalloc_end);
1333         if (ret == -EAGAIN) {
1334                 /* some of the pages are gone, lets avoid looping by
1335                  * shortening the size of the delalloc range we're searching
1336                  */
1337                 free_extent_state(cached_state);
1338                 if (!loops) {
1339                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1340                         max_bytes = PAGE_CACHE_SIZE - offset;
1341                         loops = 1;
1342                         goto again;
1343                 } else {
1344                         found = 0;
1345                         goto out_failed;
1346                 }
1347         }
1348         BUG_ON(ret);
1349
1350         /* step three, lock the state bits for the whole range */
1351         lock_extent_bits(tree, delalloc_start, delalloc_end,
1352                          0, &cached_state, GFP_NOFS);
1353
1354         /* then test to make sure it is all still delalloc */
1355         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1356                              EXTENT_DELALLOC, 1, cached_state);
1357         if (!ret) {
1358                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1359                                      &cached_state, GFP_NOFS);
1360                 __unlock_for_delalloc(inode, locked_page,
1361                               delalloc_start, delalloc_end);
1362                 cond_resched();
1363                 goto again;
1364         }
1365         free_extent_state(cached_state);
1366         *start = delalloc_start;
1367         *end = delalloc_end;
1368 out_failed:
1369         return found;
1370 }
1371
1372 int extent_clear_unlock_delalloc(struct inode *inode,
1373                                 struct extent_io_tree *tree,
1374                                 u64 start, u64 end, struct page *locked_page,
1375                                 unsigned long op)
1376 {
1377         int ret;
1378         struct page *pages[16];
1379         unsigned long index = start >> PAGE_CACHE_SHIFT;
1380         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1381         unsigned long nr_pages = end_index - index + 1;
1382         int i;
1383         int clear_bits = 0;
1384
1385         if (op & EXTENT_CLEAR_UNLOCK)
1386                 clear_bits |= EXTENT_LOCKED;
1387         if (op & EXTENT_CLEAR_DIRTY)
1388                 clear_bits |= EXTENT_DIRTY;
1389
1390         if (op & EXTENT_CLEAR_DELALLOC)
1391                 clear_bits |= EXTENT_DELALLOC;
1392
1393         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1394         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1395                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1396                     EXTENT_SET_PRIVATE2)))
1397                 return 0;
1398
1399         while (nr_pages > 0) {
1400                 ret = find_get_pages_contig(inode->i_mapping, index,
1401                                      min_t(unsigned long,
1402                                      nr_pages, ARRAY_SIZE(pages)), pages);
1403                 for (i = 0; i < ret; i++) {
1404
1405                         if (op & EXTENT_SET_PRIVATE2)
1406                                 SetPagePrivate2(pages[i]);
1407
1408                         if (pages[i] == locked_page) {
1409                                 page_cache_release(pages[i]);
1410                                 continue;
1411                         }
1412                         if (op & EXTENT_CLEAR_DIRTY)
1413                                 clear_page_dirty_for_io(pages[i]);
1414                         if (op & EXTENT_SET_WRITEBACK)
1415                                 set_page_writeback(pages[i]);
1416                         if (op & EXTENT_END_WRITEBACK)
1417                                 end_page_writeback(pages[i]);
1418                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1419                                 unlock_page(pages[i]);
1420                         page_cache_release(pages[i]);
1421                 }
1422                 nr_pages -= ret;
1423                 index += ret;
1424                 cond_resched();
1425         }
1426         return 0;
1427 }
1428
1429 /*
1430  * count the number of bytes in the tree that have a given bit(s)
1431  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1432  * cached.  The total number found is returned.
1433  */
1434 u64 count_range_bits(struct extent_io_tree *tree,
1435                      u64 *start, u64 search_end, u64 max_bytes,
1436                      unsigned long bits)
1437 {
1438         struct rb_node *node;
1439         struct extent_state *state;
1440         u64 cur_start = *start;
1441         u64 total_bytes = 0;
1442         int found = 0;
1443
1444         if (search_end <= cur_start) {
1445                 WARN_ON(1);
1446                 return 0;
1447         }
1448
1449         spin_lock(&tree->lock);
1450         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1451                 total_bytes = tree->dirty_bytes;
1452                 goto out;
1453         }
1454         /*
1455          * this search will find all the extents that end after
1456          * our range starts.
1457          */
1458         node = tree_search(tree, cur_start);
1459         if (!node)
1460                 goto out;
1461
1462         while (1) {
1463                 state = rb_entry(node, struct extent_state, rb_node);
1464                 if (state->start > search_end)
1465                         break;
1466                 if (state->end >= cur_start && (state->state & bits)) {
1467                         total_bytes += min(search_end, state->end) + 1 -
1468                                        max(cur_start, state->start);
1469                         if (total_bytes >= max_bytes)
1470                                 break;
1471                         if (!found) {
1472                                 *start = state->start;
1473                                 found = 1;
1474                         }
1475                 }
1476                 node = rb_next(node);
1477                 if (!node)
1478                         break;
1479         }
1480 out:
1481         spin_unlock(&tree->lock);
1482         return total_bytes;
1483 }
1484
1485 /*
1486  * set the private field for a given byte offset in the tree.  If there isn't
1487  * an extent_state there already, this does nothing.
1488  */
1489 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1490 {
1491         struct rb_node *node;
1492         struct extent_state *state;
1493         int ret = 0;
1494
1495         spin_lock(&tree->lock);
1496         /*
1497          * this search will find all the extents that end after
1498          * our range starts.
1499          */
1500         node = tree_search(tree, start);
1501         if (!node) {
1502                 ret = -ENOENT;
1503                 goto out;
1504         }
1505         state = rb_entry(node, struct extent_state, rb_node);
1506         if (state->start != start) {
1507                 ret = -ENOENT;
1508                 goto out;
1509         }
1510         state->private = private;
1511 out:
1512         spin_unlock(&tree->lock);
1513         return ret;
1514 }
1515
1516 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1517 {
1518         struct rb_node *node;
1519         struct extent_state *state;
1520         int ret = 0;
1521
1522         spin_lock(&tree->lock);
1523         /*
1524          * this search will find all the extents that end after
1525          * our range starts.
1526          */
1527         node = tree_search(tree, start);
1528         if (!node) {
1529                 ret = -ENOENT;
1530                 goto out;
1531         }
1532         state = rb_entry(node, struct extent_state, rb_node);
1533         if (state->start != start) {
1534                 ret = -ENOENT;
1535                 goto out;
1536         }
1537         *private = state->private;
1538 out:
1539         spin_unlock(&tree->lock);
1540         return ret;
1541 }
1542
1543 /*
1544  * searches a range in the state tree for a given mask.
1545  * If 'filled' == 1, this returns 1 only if every extent in the tree
1546  * has the bits set.  Otherwise, 1 is returned if any bit in the
1547  * range is found set.
1548  */
1549 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1550                    int bits, int filled, struct extent_state *cached)
1551 {
1552         struct extent_state *state = NULL;
1553         struct rb_node *node;
1554         int bitset = 0;
1555
1556         spin_lock(&tree->lock);
1557         if (cached && cached->tree && cached->start == start)
1558                 node = &cached->rb_node;
1559         else
1560                 node = tree_search(tree, start);
1561         while (node && start <= end) {
1562                 state = rb_entry(node, struct extent_state, rb_node);
1563
1564                 if (filled && state->start > start) {
1565                         bitset = 0;
1566                         break;
1567                 }
1568
1569                 if (state->start > end)
1570                         break;
1571
1572                 if (state->state & bits) {
1573                         bitset = 1;
1574                         if (!filled)
1575                                 break;
1576                 } else if (filled) {
1577                         bitset = 0;
1578                         break;
1579                 }
1580
1581                 if (state->end == (u64)-1)
1582                         break;
1583
1584                 start = state->end + 1;
1585                 if (start > end)
1586                         break;
1587                 node = rb_next(node);
1588                 if (!node) {
1589                         if (filled)
1590                                 bitset = 0;
1591                         break;
1592                 }
1593         }
1594         spin_unlock(&tree->lock);
1595         return bitset;
1596 }
1597
1598 /*
1599  * helper function to set a given page up to date if all the
1600  * extents in the tree for that page are up to date
1601  */
1602 static int check_page_uptodate(struct extent_io_tree *tree,
1603                                struct page *page)
1604 {
1605         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1606         u64 end = start + PAGE_CACHE_SIZE - 1;
1607         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1608                 SetPageUptodate(page);
1609         return 0;
1610 }
1611
1612 /*
1613  * helper function to unlock a page if all the extents in the tree
1614  * for that page are unlocked
1615  */
1616 static int check_page_locked(struct extent_io_tree *tree,
1617                              struct page *page)
1618 {
1619         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1620         u64 end = start + PAGE_CACHE_SIZE - 1;
1621         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1622                 unlock_page(page);
1623         return 0;
1624 }
1625
1626 /*
1627  * helper function to end page writeback if all the extents
1628  * in the tree for that page are done with writeback
1629  */
1630 static int check_page_writeback(struct extent_io_tree *tree,
1631                              struct page *page)
1632 {
1633         end_page_writeback(page);
1634         return 0;
1635 }
1636
1637 /* lots and lots of room for performance fixes in the end_bio funcs */
1638
1639 /*
1640  * after a writepage IO is done, we need to:
1641  * clear the uptodate bits on error
1642  * clear the writeback bits in the extent tree for this IO
1643  * end_page_writeback if the page has no more pending IO
1644  *
1645  * Scheduling is not allowed, so the extent state tree is expected
1646  * to have one and only one object corresponding to this IO.
1647  */
1648 static void end_bio_extent_writepage(struct bio *bio, int err)
1649 {
1650         int uptodate = err == 0;
1651         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1652         struct extent_io_tree *tree;
1653         u64 start;
1654         u64 end;
1655         int whole_page;
1656         int ret;
1657
1658         do {
1659                 struct page *page = bvec->bv_page;
1660                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1661
1662                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1663                          bvec->bv_offset;
1664                 end = start + bvec->bv_len - 1;
1665
1666                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1667                         whole_page = 1;
1668                 else
1669                         whole_page = 0;
1670
1671                 if (--bvec >= bio->bi_io_vec)
1672                         prefetchw(&bvec->bv_page->flags);
1673                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1674                         ret = tree->ops->writepage_end_io_hook(page, start,
1675                                                        end, NULL, uptodate);
1676                         if (ret)
1677                                 uptodate = 0;
1678                 }
1679
1680                 if (!uptodate && tree->ops &&
1681                     tree->ops->writepage_io_failed_hook) {
1682                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1683                                                          start, end, NULL);
1684                         if (ret == 0) {
1685                                 uptodate = (err == 0);
1686                                 continue;
1687                         }
1688                 }
1689
1690                 if (!uptodate) {
1691                         clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
1692                         ClearPageUptodate(page);
1693                         SetPageError(page);
1694                 }
1695
1696                 if (whole_page)
1697                         end_page_writeback(page);
1698                 else
1699                         check_page_writeback(tree, page);
1700         } while (bvec >= bio->bi_io_vec);
1701
1702         bio_put(bio);
1703 }
1704
1705 /*
1706  * after a readpage IO is done, we need to:
1707  * clear the uptodate bits on error
1708  * set the uptodate bits if things worked
1709  * set the page up to date if all extents in the tree are uptodate
1710  * clear the lock bit in the extent tree
1711  * unlock the page if there are no other extents locked for it
1712  *
1713  * Scheduling is not allowed, so the extent state tree is expected
1714  * to have one and only one object corresponding to this IO.
1715  */
1716 static void end_bio_extent_readpage(struct bio *bio, int err)
1717 {
1718         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1719         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
1720         struct bio_vec *bvec = bio->bi_io_vec;
1721         struct extent_io_tree *tree;
1722         u64 start;
1723         u64 end;
1724         int whole_page;
1725         int ret;
1726
1727         if (err)
1728                 uptodate = 0;
1729
1730         do {
1731                 struct page *page = bvec->bv_page;
1732                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1733
1734                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1735                         bvec->bv_offset;
1736                 end = start + bvec->bv_len - 1;
1737
1738                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1739                         whole_page = 1;
1740                 else
1741                         whole_page = 0;
1742
1743                 if (++bvec <= bvec_end)
1744                         prefetchw(&bvec->bv_page->flags);
1745
1746                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1747                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1748                                                               NULL);
1749                         if (ret)
1750                                 uptodate = 0;
1751                 }
1752                 if (!uptodate && tree->ops &&
1753                     tree->ops->readpage_io_failed_hook) {
1754                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1755                                                          start, end, NULL);
1756                         if (ret == 0) {
1757                                 uptodate =
1758                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1759                                 if (err)
1760                                         uptodate = 0;
1761                                 continue;
1762                         }
1763                 }
1764
1765                 if (uptodate) {
1766                         set_extent_uptodate(tree, start, end,
1767                                             GFP_ATOMIC);
1768                 }
1769                 unlock_extent(tree, start, end, GFP_ATOMIC);
1770
1771                 if (whole_page) {
1772                         if (uptodate) {
1773                                 SetPageUptodate(page);
1774                         } else {
1775                                 ClearPageUptodate(page);
1776                                 SetPageError(page);
1777                         }
1778                         unlock_page(page);
1779                 } else {
1780                         if (uptodate) {
1781                                 check_page_uptodate(tree, page);
1782                         } else {
1783                                 ClearPageUptodate(page);
1784                                 SetPageError(page);
1785                         }
1786                         check_page_locked(tree, page);
1787                 }
1788         } while (bvec <= bvec_end);
1789
1790         bio_put(bio);
1791 }
1792
1793 /*
1794  * IO done from prepare_write is pretty simple, we just unlock
1795  * the structs in the extent tree when done, and set the uptodate bits
1796  * as appropriate.
1797  */
1798 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1799 {
1800         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1801         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1802         struct extent_io_tree *tree;
1803         u64 start;
1804         u64 end;
1805
1806         do {
1807                 struct page *page = bvec->bv_page;
1808                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1809
1810                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1811                         bvec->bv_offset;
1812                 end = start + bvec->bv_len - 1;
1813
1814                 if (--bvec >= bio->bi_io_vec)
1815                         prefetchw(&bvec->bv_page->flags);
1816
1817                 if (uptodate) {
1818                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1819                 } else {
1820                         ClearPageUptodate(page);
1821                         SetPageError(page);
1822                 }
1823
1824                 unlock_extent(tree, start, end, GFP_ATOMIC);
1825
1826         } while (bvec >= bio->bi_io_vec);
1827
1828         bio_put(bio);
1829 }
1830
1831 struct bio *
1832 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1833                 gfp_t gfp_flags)
1834 {
1835         struct bio *bio;
1836
1837         bio = bio_alloc(gfp_flags, nr_vecs);
1838
1839         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1840                 while (!bio && (nr_vecs /= 2))
1841                         bio = bio_alloc(gfp_flags, nr_vecs);
1842         }
1843
1844         if (bio) {
1845                 bio->bi_size = 0;
1846                 bio->bi_bdev = bdev;
1847                 bio->bi_sector = first_sector;
1848         }
1849         return bio;
1850 }
1851
1852 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1853                           unsigned long bio_flags)
1854 {
1855         int ret = 0;
1856         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1857         struct page *page = bvec->bv_page;
1858         struct extent_io_tree *tree = bio->bi_private;
1859         u64 start;
1860
1861         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1862
1863         bio->bi_private = NULL;
1864
1865         bio_get(bio);
1866
1867         if (tree->ops && tree->ops->submit_bio_hook)
1868                 tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1869                                            mirror_num, bio_flags, start);
1870         else
1871                 submit_bio(rw, bio);
1872         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1873                 ret = -EOPNOTSUPP;
1874         bio_put(bio);
1875         return ret;
1876 }
1877
1878 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1879                               struct page *page, sector_t sector,
1880                               size_t size, unsigned long offset,
1881                               struct block_device *bdev,
1882                               struct bio **bio_ret,
1883                               unsigned long max_pages,
1884                               bio_end_io_t end_io_func,
1885                               int mirror_num,
1886                               unsigned long prev_bio_flags,
1887                               unsigned long bio_flags)
1888 {
1889         int ret = 0;
1890         struct bio *bio;
1891         int nr;
1892         int contig = 0;
1893         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1894         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1895         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1896
1897         if (bio_ret && *bio_ret) {
1898                 bio = *bio_ret;
1899                 if (old_compressed)
1900                         contig = bio->bi_sector == sector;
1901                 else
1902                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
1903                                 sector;
1904
1905                 if (prev_bio_flags != bio_flags || !contig ||
1906                     (tree->ops && tree->ops->merge_bio_hook &&
1907                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
1908                                                bio_flags)) ||
1909                     bio_add_page(bio, page, page_size, offset) < page_size) {
1910                         ret = submit_one_bio(rw, bio, mirror_num,
1911                                              prev_bio_flags);
1912                         bio = NULL;
1913                 } else {
1914                         return 0;
1915                 }
1916         }
1917         if (this_compressed)
1918                 nr = BIO_MAX_PAGES;
1919         else
1920                 nr = bio_get_nr_vecs(bdev);
1921
1922         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1923
1924         bio_add_page(bio, page, page_size, offset);
1925         bio->bi_end_io = end_io_func;
1926         bio->bi_private = tree;
1927
1928         if (bio_ret)
1929                 *bio_ret = bio;
1930         else
1931                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1932
1933         return ret;
1934 }
1935
1936 void set_page_extent_mapped(struct page *page)
1937 {
1938         if (!PagePrivate(page)) {
1939                 SetPagePrivate(page);
1940                 page_cache_get(page);
1941                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1942         }
1943 }
1944
1945 static void set_page_extent_head(struct page *page, unsigned long len)
1946 {
1947         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1948 }
1949
1950 /*
1951  * basic readpage implementation.  Locked extent state structs are inserted
1952  * into the tree that are removed when the IO is done (by the end_io
1953  * handlers)
1954  */
1955 static int __extent_read_full_page(struct extent_io_tree *tree,
1956                                    struct page *page,
1957                                    get_extent_t *get_extent,
1958                                    struct bio **bio, int mirror_num,
1959                                    unsigned long *bio_flags)
1960 {
1961         struct inode *inode = page->mapping->host;
1962         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1963         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1964         u64 end;
1965         u64 cur = start;
1966         u64 extent_offset;
1967         u64 last_byte = i_size_read(inode);
1968         u64 block_start;
1969         u64 cur_end;
1970         sector_t sector;
1971         struct extent_map *em;
1972         struct block_device *bdev;
1973         struct btrfs_ordered_extent *ordered;
1974         int ret;
1975         int nr = 0;
1976         size_t page_offset = 0;
1977         size_t iosize;
1978         size_t disk_io_size;
1979         size_t blocksize = inode->i_sb->s_blocksize;
1980         unsigned long this_bio_flag = 0;
1981
1982         set_page_extent_mapped(page);
1983
1984         end = page_end;
1985         while (1) {
1986                 lock_extent(tree, start, end, GFP_NOFS);
1987                 ordered = btrfs_lookup_ordered_extent(inode, start);
1988                 if (!ordered)
1989                         break;
1990                 unlock_extent(tree, start, end, GFP_NOFS);
1991                 btrfs_start_ordered_extent(inode, ordered, 1);
1992                 btrfs_put_ordered_extent(ordered);
1993         }
1994
1995         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
1996                 char *userpage;
1997                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
1998
1999                 if (zero_offset) {
2000                         iosize = PAGE_CACHE_SIZE - zero_offset;
2001                         userpage = kmap_atomic(page, KM_USER0);
2002                         memset(userpage + zero_offset, 0, iosize);
2003                         flush_dcache_page(page);
2004                         kunmap_atomic(userpage, KM_USER0);
2005                 }
2006         }
2007         while (cur <= end) {
2008                 if (cur >= last_byte) {
2009                         char *userpage;
2010                         iosize = PAGE_CACHE_SIZE - page_offset;
2011                         userpage = kmap_atomic(page, KM_USER0);
2012                         memset(userpage + page_offset, 0, iosize);
2013                         flush_dcache_page(page);
2014                         kunmap_atomic(userpage, KM_USER0);
2015                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2016                                             GFP_NOFS);
2017                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2018                         break;
2019                 }
2020                 em = get_extent(inode, page, page_offset, cur,
2021                                 end - cur + 1, 0);
2022                 if (IS_ERR(em) || !em) {
2023                         SetPageError(page);
2024                         unlock_extent(tree, cur, end, GFP_NOFS);
2025                         break;
2026                 }
2027                 extent_offset = cur - em->start;
2028                 BUG_ON(extent_map_end(em) <= cur);
2029                 BUG_ON(end < cur);
2030
2031                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2032                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2033
2034                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2035                 cur_end = min(extent_map_end(em) - 1, end);
2036                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2037                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2038                         disk_io_size = em->block_len;
2039                         sector = em->block_start >> 9;
2040                 } else {
2041                         sector = (em->block_start + extent_offset) >> 9;
2042                         disk_io_size = iosize;
2043                 }
2044                 bdev = em->bdev;
2045                 block_start = em->block_start;
2046                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2047                         block_start = EXTENT_MAP_HOLE;
2048                 free_extent_map(em);
2049                 em = NULL;
2050
2051                 /* we've found a hole, just zero and go on */
2052                 if (block_start == EXTENT_MAP_HOLE) {
2053                         char *userpage;
2054                         userpage = kmap_atomic(page, KM_USER0);
2055                         memset(userpage + page_offset, 0, iosize);
2056                         flush_dcache_page(page);
2057                         kunmap_atomic(userpage, KM_USER0);
2058
2059                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2060                                             GFP_NOFS);
2061                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2062                         cur = cur + iosize;
2063                         page_offset += iosize;
2064                         continue;
2065                 }
2066                 /* the get_extent function already copied into the page */
2067                 if (test_range_bit(tree, cur, cur_end,
2068                                    EXTENT_UPTODATE, 1, NULL)) {
2069                         check_page_uptodate(tree, page);
2070                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2071                         cur = cur + iosize;
2072                         page_offset += iosize;
2073                         continue;
2074                 }
2075                 /* we have an inline extent but it didn't get marked up
2076                  * to date.  Error out
2077                  */
2078                 if (block_start == EXTENT_MAP_INLINE) {
2079                         SetPageError(page);
2080                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2081                         cur = cur + iosize;
2082                         page_offset += iosize;
2083                         continue;
2084                 }
2085
2086                 ret = 0;
2087                 if (tree->ops && tree->ops->readpage_io_hook) {
2088                         ret = tree->ops->readpage_io_hook(page, cur,
2089                                                           cur + iosize - 1);
2090                 }
2091                 if (!ret) {
2092                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2093                         pnr -= page->index;
2094                         ret = submit_extent_page(READ, tree, page,
2095                                          sector, disk_io_size, page_offset,
2096                                          bdev, bio, pnr,
2097                                          end_bio_extent_readpage, mirror_num,
2098                                          *bio_flags,
2099                                          this_bio_flag);
2100                         nr++;
2101                         *bio_flags = this_bio_flag;
2102                 }
2103                 if (ret)
2104                         SetPageError(page);
2105                 cur = cur + iosize;
2106                 page_offset += iosize;
2107         }
2108         if (!nr) {
2109                 if (!PageError(page))
2110                         SetPageUptodate(page);
2111                 unlock_page(page);
2112         }
2113         return 0;
2114 }
2115
2116 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2117                             get_extent_t *get_extent)
2118 {
2119         struct bio *bio = NULL;
2120         unsigned long bio_flags = 0;
2121         int ret;
2122
2123         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2124                                       &bio_flags);
2125         if (bio)
2126                 submit_one_bio(READ, bio, 0, bio_flags);
2127         return ret;
2128 }
2129
2130 static noinline void update_nr_written(struct page *page,
2131                                       struct writeback_control *wbc,
2132                                       unsigned long nr_written)
2133 {
2134         wbc->nr_to_write -= nr_written;
2135         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2136             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2137                 page->mapping->writeback_index = page->index + nr_written;
2138 }
2139
2140 /*
2141  * the writepage semantics are similar to regular writepage.  extent
2142  * records are inserted to lock ranges in the tree, and as dirty areas
2143  * are found, they are marked writeback.  Then the lock bits are removed
2144  * and the end_io handler clears the writeback ranges
2145  */
2146 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2147                               void *data)
2148 {
2149         struct inode *inode = page->mapping->host;
2150         struct extent_page_data *epd = data;
2151         struct extent_io_tree *tree = epd->tree;
2152         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2153         u64 delalloc_start;
2154         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2155         u64 end;
2156         u64 cur = start;
2157         u64 extent_offset;
2158         u64 last_byte = i_size_read(inode);
2159         u64 block_start;
2160         u64 iosize;
2161         sector_t sector;
2162         struct extent_state *cached_state = NULL;
2163         struct extent_map *em;
2164         struct block_device *bdev;
2165         int ret;
2166         int nr = 0;
2167         size_t pg_offset = 0;
2168         size_t blocksize;
2169         loff_t i_size = i_size_read(inode);
2170         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2171         u64 nr_delalloc;
2172         u64 delalloc_end;
2173         int page_started;
2174         int compressed;
2175         int write_flags;
2176         unsigned long nr_written = 0;
2177
2178         if (wbc->sync_mode == WB_SYNC_ALL)
2179                 write_flags = WRITE_SYNC_PLUG;
2180         else
2181                 write_flags = WRITE;
2182
2183         WARN_ON(!PageLocked(page));
2184         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2185         if (page->index > end_index ||
2186            (page->index == end_index && !pg_offset)) {
2187                 page->mapping->a_ops->invalidatepage(page, 0);
2188                 unlock_page(page);
2189                 return 0;
2190         }
2191
2192         if (page->index == end_index) {
2193                 char *userpage;
2194
2195                 userpage = kmap_atomic(page, KM_USER0);
2196                 memset(userpage + pg_offset, 0,
2197                        PAGE_CACHE_SIZE - pg_offset);
2198                 kunmap_atomic(userpage, KM_USER0);
2199                 flush_dcache_page(page);
2200         }
2201         pg_offset = 0;
2202
2203         set_page_extent_mapped(page);
2204
2205         delalloc_start = start;
2206         delalloc_end = 0;
2207         page_started = 0;
2208         if (!epd->extent_locked) {
2209                 u64 delalloc_to_write = 0;
2210                 /*
2211                  * make sure the wbc mapping index is at least updated
2212                  * to this page.
2213                  */
2214                 update_nr_written(page, wbc, 0);
2215
2216                 while (delalloc_end < page_end) {
2217                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2218                                                        page,
2219                                                        &delalloc_start,
2220                                                        &delalloc_end,
2221                                                        128 * 1024 * 1024);
2222                         if (nr_delalloc == 0) {
2223                                 delalloc_start = delalloc_end + 1;
2224                                 continue;
2225                         }
2226                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2227                                                  delalloc_end, &page_started,
2228                                                  &nr_written);
2229                         /*
2230                          * delalloc_end is already one less than the total
2231                          * length, so we don't subtract one from
2232                          * PAGE_CACHE_SIZE
2233                          */
2234                         delalloc_to_write += (delalloc_end - delalloc_start +
2235                                               PAGE_CACHE_SIZE) >>
2236                                               PAGE_CACHE_SHIFT;
2237                         delalloc_start = delalloc_end + 1;
2238                 }
2239                 if (wbc->nr_to_write < delalloc_to_write) {
2240                         int thresh = 8192;
2241
2242                         if (delalloc_to_write < thresh * 2)
2243                                 thresh = delalloc_to_write;
2244                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2245                                                  thresh);
2246                 }
2247
2248                 /* did the fill delalloc function already unlock and start
2249                  * the IO?
2250                  */
2251                 if (page_started) {
2252                         ret = 0;
2253                         /*
2254                          * we've unlocked the page, so we can't update
2255                          * the mapping's writeback index, just update
2256                          * nr_to_write.
2257                          */
2258                         wbc->nr_to_write -= nr_written;
2259                         goto done_unlocked;
2260                 }
2261         }
2262         if (tree->ops && tree->ops->writepage_start_hook) {
2263                 ret = tree->ops->writepage_start_hook(page, start,
2264                                                       page_end);
2265                 if (ret == -EAGAIN) {
2266                         redirty_page_for_writepage(wbc, page);
2267                         update_nr_written(page, wbc, nr_written);
2268                         unlock_page(page);
2269                         ret = 0;
2270                         goto done_unlocked;
2271                 }
2272         }
2273
2274         /*
2275          * we don't want to touch the inode after unlocking the page,
2276          * so we update the mapping writeback index now
2277          */
2278         update_nr_written(page, wbc, nr_written + 1);
2279
2280         end = page_end;
2281         if (last_byte <= start) {
2282                 if (tree->ops && tree->ops->writepage_end_io_hook)
2283                         tree->ops->writepage_end_io_hook(page, start,
2284                                                          page_end, NULL, 1);
2285                 goto done;
2286         }
2287
2288         blocksize = inode->i_sb->s_blocksize;
2289
2290         while (cur <= end) {
2291                 if (cur >= last_byte) {
2292                         if (tree->ops && tree->ops->writepage_end_io_hook)
2293                                 tree->ops->writepage_end_io_hook(page, cur,
2294                                                          page_end, NULL, 1);
2295                         break;
2296                 }
2297                 em = epd->get_extent(inode, page, pg_offset, cur,
2298                                      end - cur + 1, 1);
2299                 if (IS_ERR(em) || !em) {
2300                         SetPageError(page);
2301                         break;
2302                 }
2303
2304                 extent_offset = cur - em->start;
2305                 BUG_ON(extent_map_end(em) <= cur);
2306                 BUG_ON(end < cur);
2307                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2308                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2309                 sector = (em->block_start + extent_offset) >> 9;
2310                 bdev = em->bdev;
2311                 block_start = em->block_start;
2312                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2313                 free_extent_map(em);
2314                 em = NULL;
2315
2316                 /*
2317                  * compressed and inline extents are written through other
2318                  * paths in the FS
2319                  */
2320                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2321                     block_start == EXTENT_MAP_INLINE) {
2322                         /*
2323                          * end_io notification does not happen here for
2324                          * compressed extents
2325                          */
2326                         if (!compressed && tree->ops &&
2327                             tree->ops->writepage_end_io_hook)
2328                                 tree->ops->writepage_end_io_hook(page, cur,
2329                                                          cur + iosize - 1,
2330                                                          NULL, 1);
2331                         else if (compressed) {
2332                                 /* we don't want to end_page_writeback on
2333                                  * a compressed extent.  this happens
2334                                  * elsewhere
2335                                  */
2336                                 nr++;
2337                         }
2338
2339                         cur += iosize;
2340                         pg_offset += iosize;
2341                         continue;
2342                 }
2343                 /* leave this out until we have a page_mkwrite call */
2344                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2345                                    EXTENT_DIRTY, 0, NULL)) {
2346                         cur = cur + iosize;
2347                         pg_offset += iosize;
2348                         continue;
2349                 }
2350
2351                 if (tree->ops && tree->ops->writepage_io_hook) {
2352                         ret = tree->ops->writepage_io_hook(page, cur,
2353                                                 cur + iosize - 1);
2354                 } else {
2355                         ret = 0;
2356                 }
2357                 if (ret) {
2358                         SetPageError(page);
2359                 } else {
2360                         unsigned long max_nr = end_index + 1;
2361
2362                         set_range_writeback(tree, cur, cur + iosize - 1);
2363                         if (!PageWriteback(page)) {
2364                                 printk(KERN_ERR "btrfs warning page %lu not "
2365                                        "writeback, cur %llu end %llu\n",
2366                                        page->index, (unsigned long long)cur,
2367                                        (unsigned long long)end);
2368                         }
2369
2370                         ret = submit_extent_page(write_flags, tree, page,
2371                                                  sector, iosize, pg_offset,
2372                                                  bdev, &epd->bio, max_nr,
2373                                                  end_bio_extent_writepage,
2374                                                  0, 0, 0);
2375                         if (ret)
2376                                 SetPageError(page);
2377                 }
2378                 cur = cur + iosize;
2379                 pg_offset += iosize;
2380                 nr++;
2381         }
2382 done:
2383         if (nr == 0) {
2384                 /* make sure the mapping tag for page dirty gets cleared */
2385                 set_page_writeback(page);
2386                 end_page_writeback(page);
2387         }
2388         unlock_page(page);
2389
2390 done_unlocked:
2391
2392         /* drop our reference on any cached states */
2393         free_extent_state(cached_state);
2394         return 0;
2395 }
2396
2397 /**
2398  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2399  * @mapping: address space structure to write
2400  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2401  * @writepage: function called for each page
2402  * @data: data passed to writepage function
2403  *
2404  * If a page is already under I/O, write_cache_pages() skips it, even
2405  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2406  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2407  * and msync() need to guarantee that all the data which was dirty at the time
2408  * the call was made get new I/O started against them.  If wbc->sync_mode is
2409  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2410  * existing IO to complete.
2411  */
2412 static int extent_write_cache_pages(struct extent_io_tree *tree,
2413                              struct address_space *mapping,
2414                              struct writeback_control *wbc,
2415                              writepage_t writepage, void *data,
2416                              void (*flush_fn)(void *))
2417 {
2418         int ret = 0;
2419         int done = 0;
2420         int nr_to_write_done = 0;
2421         struct pagevec pvec;
2422         int nr_pages;
2423         pgoff_t index;
2424         pgoff_t end;            /* Inclusive */
2425         int scanned = 0;
2426
2427         pagevec_init(&pvec, 0);
2428         if (wbc->range_cyclic) {
2429                 index = mapping->writeback_index; /* Start from prev offset */
2430                 end = -1;
2431         } else {
2432                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2433                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2434                 scanned = 1;
2435         }
2436 retry:
2437         while (!done && !nr_to_write_done && (index <= end) &&
2438                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2439                               PAGECACHE_TAG_DIRTY, min(end - index,
2440                                   (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2441                 unsigned i;
2442
2443                 scanned = 1;
2444                 for (i = 0; i < nr_pages; i++) {
2445                         struct page *page = pvec.pages[i];
2446
2447                         /*
2448                          * At this point we hold neither mapping->tree_lock nor
2449                          * lock on the page itself: the page may be truncated or
2450                          * invalidated (changing page->mapping to NULL), or even
2451                          * swizzled back from swapper_space to tmpfs file
2452                          * mapping
2453                          */
2454                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2455                                 tree->ops->write_cache_pages_lock_hook(page);
2456                         else
2457                                 lock_page(page);
2458
2459                         if (unlikely(page->mapping != mapping)) {
2460                                 unlock_page(page);
2461                                 continue;
2462                         }
2463
2464                         if (!wbc->range_cyclic && page->index > end) {
2465                                 done = 1;
2466                                 unlock_page(page);
2467                                 continue;
2468                         }
2469
2470                         if (wbc->sync_mode != WB_SYNC_NONE) {
2471                                 if (PageWriteback(page))
2472                                         flush_fn(data);
2473                                 wait_on_page_writeback(page);
2474                         }
2475
2476                         if (PageWriteback(page) ||
2477                             !clear_page_dirty_for_io(page)) {
2478                                 unlock_page(page);
2479                                 continue;
2480                         }
2481
2482                         ret = (*writepage)(page, wbc, data);
2483
2484                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2485                                 unlock_page(page);
2486                                 ret = 0;
2487                         }
2488                         if (ret)
2489                                 done = 1;
2490
2491                         /*
2492                          * the filesystem may choose to bump up nr_to_write.
2493                          * We have to make sure to honor the new nr_to_write
2494                          * at any time
2495                          */
2496                         nr_to_write_done = wbc->nr_to_write <= 0;
2497                 }
2498                 pagevec_release(&pvec);
2499                 cond_resched();
2500         }
2501         if (!scanned && !done) {
2502                 /*
2503                  * We hit the last page and there is more work to be done: wrap
2504                  * back to the start of the file
2505                  */
2506                 scanned = 1;
2507                 index = 0;
2508                 goto retry;
2509         }
2510         return ret;
2511 }
2512
2513 static void flush_epd_write_bio(struct extent_page_data *epd)
2514 {
2515         if (epd->bio) {
2516                 if (epd->sync_io)
2517                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2518                 else
2519                         submit_one_bio(WRITE, epd->bio, 0, 0);
2520                 epd->bio = NULL;
2521         }
2522 }
2523
2524 static noinline void flush_write_bio(void *data)
2525 {
2526         struct extent_page_data *epd = data;
2527         flush_epd_write_bio(epd);
2528 }
2529
2530 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2531                           get_extent_t *get_extent,
2532                           struct writeback_control *wbc)
2533 {
2534         int ret;
2535         struct address_space *mapping = page->mapping;
2536         struct extent_page_data epd = {
2537                 .bio = NULL,
2538                 .tree = tree,
2539                 .get_extent = get_extent,
2540                 .extent_locked = 0,
2541                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2542         };
2543         struct writeback_control wbc_writepages = {
2544                 .sync_mode      = wbc->sync_mode,
2545                 .older_than_this = NULL,
2546                 .nr_to_write    = 64,
2547                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2548                 .range_end      = (loff_t)-1,
2549         };
2550
2551         ret = __extent_writepage(page, wbc, &epd);
2552
2553         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2554                                  __extent_writepage, &epd, flush_write_bio);
2555         flush_epd_write_bio(&epd);
2556         return ret;
2557 }
2558
2559 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2560                               u64 start, u64 end, get_extent_t *get_extent,
2561                               int mode)
2562 {
2563         int ret = 0;
2564         struct address_space *mapping = inode->i_mapping;
2565         struct page *page;
2566         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2567                 PAGE_CACHE_SHIFT;
2568
2569         struct extent_page_data epd = {
2570                 .bio = NULL,
2571                 .tree = tree,
2572                 .get_extent = get_extent,
2573                 .extent_locked = 1,
2574                 .sync_io = mode == WB_SYNC_ALL,
2575         };
2576         struct writeback_control wbc_writepages = {
2577                 .sync_mode      = mode,
2578                 .older_than_this = NULL,
2579                 .nr_to_write    = nr_pages * 2,
2580                 .range_start    = start,
2581                 .range_end      = end + 1,
2582         };
2583
2584         while (start <= end) {
2585                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2586                 if (clear_page_dirty_for_io(page))
2587                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2588                 else {
2589                         if (tree->ops && tree->ops->writepage_end_io_hook)
2590                                 tree->ops->writepage_end_io_hook(page, start,
2591                                                  start + PAGE_CACHE_SIZE - 1,
2592                                                  NULL, 1);
2593                         unlock_page(page);
2594                 }
2595                 page_cache_release(page);
2596                 start += PAGE_CACHE_SIZE;
2597         }
2598
2599         flush_epd_write_bio(&epd);
2600         return ret;
2601 }
2602
2603 int extent_writepages(struct extent_io_tree *tree,
2604                       struct address_space *mapping,
2605                       get_extent_t *get_extent,
2606                       struct writeback_control *wbc)
2607 {
2608         int ret = 0;
2609         struct extent_page_data epd = {
2610                 .bio = NULL,
2611                 .tree = tree,
2612                 .get_extent = get_extent,
2613                 .extent_locked = 0,
2614                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2615         };
2616
2617         ret = extent_write_cache_pages(tree, mapping, wbc,
2618                                        __extent_writepage, &epd,
2619                                        flush_write_bio);
2620         flush_epd_write_bio(&epd);
2621         return ret;
2622 }
2623
2624 int extent_readpages(struct extent_io_tree *tree,
2625                      struct address_space *mapping,
2626                      struct list_head *pages, unsigned nr_pages,
2627                      get_extent_t get_extent)
2628 {
2629         struct bio *bio = NULL;
2630         unsigned page_idx;
2631         unsigned long bio_flags = 0;
2632
2633         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2634                 struct page *page = list_entry(pages->prev, struct page, lru);
2635
2636                 prefetchw(&page->flags);
2637                 list_del(&page->lru);
2638                 if (!add_to_page_cache_lru(page, mapping,
2639                                         page->index, GFP_KERNEL)) {
2640                         __extent_read_full_page(tree, page, get_extent,
2641                                                 &bio, 0, &bio_flags);
2642                 }
2643                 page_cache_release(page);
2644         }
2645         BUG_ON(!list_empty(pages));
2646         if (bio)
2647                 submit_one_bio(READ, bio, 0, bio_flags);
2648         return 0;
2649 }
2650
2651 /*
2652  * basic invalidatepage code, this waits on any locked or writeback
2653  * ranges corresponding to the page, and then deletes any extent state
2654  * records from the tree
2655  */
2656 int extent_invalidatepage(struct extent_io_tree *tree,
2657                           struct page *page, unsigned long offset)
2658 {
2659         struct extent_state *cached_state = NULL;
2660         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2661         u64 end = start + PAGE_CACHE_SIZE - 1;
2662         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2663
2664         start += (offset + blocksize - 1) & ~(blocksize - 1);
2665         if (start > end)
2666                 return 0;
2667
2668         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
2669         wait_on_page_writeback(page);
2670         clear_extent_bit(tree, start, end,
2671                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2672                          EXTENT_DO_ACCOUNTING,
2673                          1, 1, &cached_state, GFP_NOFS);
2674         return 0;
2675 }
2676
2677 /*
2678  * simple commit_write call, set_range_dirty is used to mark both
2679  * the pages and the extent records as dirty
2680  */
2681 int extent_commit_write(struct extent_io_tree *tree,
2682                         struct inode *inode, struct page *page,
2683                         unsigned from, unsigned to)
2684 {
2685         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2686
2687         set_page_extent_mapped(page);
2688         set_page_dirty(page);
2689
2690         if (pos > inode->i_size) {
2691                 i_size_write(inode, pos);
2692                 mark_inode_dirty(inode);
2693         }
2694         return 0;
2695 }
2696
2697 int extent_prepare_write(struct extent_io_tree *tree,
2698                          struct inode *inode, struct page *page,
2699                          unsigned from, unsigned to, get_extent_t *get_extent)
2700 {
2701         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2702         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2703         u64 block_start;
2704         u64 orig_block_start;
2705         u64 block_end;
2706         u64 cur_end;
2707         struct extent_map *em;
2708         unsigned blocksize = 1 << inode->i_blkbits;
2709         size_t page_offset = 0;
2710         size_t block_off_start;
2711         size_t block_off_end;
2712         int err = 0;
2713         int iocount = 0;
2714         int ret = 0;
2715         int isnew;
2716
2717         set_page_extent_mapped(page);
2718
2719         block_start = (page_start + from) & ~((u64)blocksize - 1);
2720         block_end = (page_start + to - 1) | (blocksize - 1);
2721         orig_block_start = block_start;
2722
2723         lock_extent(tree, page_start, page_end, GFP_NOFS);
2724         while (block_start <= block_end) {
2725                 em = get_extent(inode, page, page_offset, block_start,
2726                                 block_end - block_start + 1, 1);
2727                 if (IS_ERR(em) || !em)
2728                         goto err;
2729
2730                 cur_end = min(block_end, extent_map_end(em) - 1);
2731                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2732                 block_off_end = block_off_start + blocksize;
2733                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2734
2735                 if (!PageUptodate(page) && isnew &&
2736                     (block_off_end > to || block_off_start < from)) {
2737                         void *kaddr;
2738
2739                         kaddr = kmap_atomic(page, KM_USER0);
2740                         if (block_off_end > to)
2741                                 memset(kaddr + to, 0, block_off_end - to);
2742                         if (block_off_start < from)
2743                                 memset(kaddr + block_off_start, 0,
2744                                        from - block_off_start);
2745                         flush_dcache_page(page);
2746                         kunmap_atomic(kaddr, KM_USER0);
2747                 }
2748                 if ((em->block_start != EXTENT_MAP_HOLE &&
2749                      em->block_start != EXTENT_MAP_INLINE) &&
2750                     !isnew && !PageUptodate(page) &&
2751                     (block_off_end > to || block_off_start < from) &&
2752                     !test_range_bit(tree, block_start, cur_end,
2753                                     EXTENT_UPTODATE, 1, NULL)) {
2754                         u64 sector;
2755                         u64 extent_offset = block_start - em->start;
2756                         size_t iosize;
2757                         sector = (em->block_start + extent_offset) >> 9;
2758                         iosize = (cur_end - block_start + blocksize) &
2759                                 ~((u64)blocksize - 1);
2760                         /*
2761                          * we've already got the extent locked, but we
2762                          * need to split the state such that our end_bio
2763                          * handler can clear the lock.
2764                          */
2765                         set_extent_bit(tree, block_start,
2766                                        block_start + iosize - 1,
2767                                        EXTENT_LOCKED, 0, NULL, NULL, GFP_NOFS);
2768                         ret = submit_extent_page(READ, tree, page,
2769                                          sector, iosize, page_offset, em->bdev,
2770                                          NULL, 1,
2771                                          end_bio_extent_preparewrite, 0,
2772                                          0, 0);
2773                         if (ret && !err)
2774                                 err = ret;
2775                         iocount++;
2776                         block_start = block_start + iosize;
2777                 } else {
2778                         set_extent_uptodate(tree, block_start, cur_end,
2779                                             GFP_NOFS);
2780                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2781                         block_start = cur_end + 1;
2782                 }
2783                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2784                 free_extent_map(em);
2785         }
2786         if (iocount) {
2787                 wait_extent_bit(tree, orig_block_start,
2788                                 block_end, EXTENT_LOCKED);
2789         }
2790         check_page_uptodate(tree, page);
2791 err:
2792         /* FIXME, zero out newly allocated blocks on error */
2793         return err;
2794 }
2795
2796 /*
2797  * a helper for releasepage, this tests for areas of the page that
2798  * are locked or under IO and drops the related state bits if it is safe
2799  * to drop the page.
2800  */
2801 int try_release_extent_state(struct extent_map_tree *map,
2802                              struct extent_io_tree *tree, struct page *page,
2803                              gfp_t mask)
2804 {
2805         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2806         u64 end = start + PAGE_CACHE_SIZE - 1;
2807         int ret = 1;
2808
2809         if (test_range_bit(tree, start, end,
2810                            EXTENT_IOBITS, 0, NULL))
2811                 ret = 0;
2812         else {
2813                 if ((mask & GFP_NOFS) == GFP_NOFS)
2814                         mask = GFP_NOFS;
2815                 /*
2816                  * at this point we can safely clear everything except the
2817                  * locked bit and the nodatasum bit
2818                  */
2819                 clear_extent_bit(tree, start, end,
2820                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2821                                  0, 0, NULL, mask);
2822         }
2823         return ret;
2824 }
2825
2826 /*
2827  * a helper for releasepage.  As long as there are no locked extents
2828  * in the range corresponding to the page, both state records and extent
2829  * map records are removed
2830  */
2831 int try_release_extent_mapping(struct extent_map_tree *map,
2832                                struct extent_io_tree *tree, struct page *page,
2833                                gfp_t mask)
2834 {
2835         struct extent_map *em;
2836         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2837         u64 end = start + PAGE_CACHE_SIZE - 1;
2838
2839         if ((mask & __GFP_WAIT) &&
2840             page->mapping->host->i_size > 16 * 1024 * 1024) {
2841                 u64 len;
2842                 while (start <= end) {
2843                         len = end - start + 1;
2844                         write_lock(&map->lock);
2845                         em = lookup_extent_mapping(map, start, len);
2846                         if (!em || IS_ERR(em)) {
2847                                 write_unlock(&map->lock);
2848                                 break;
2849                         }
2850                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2851                             em->start != start) {
2852                                 write_unlock(&map->lock);
2853                                 free_extent_map(em);
2854                                 break;
2855                         }
2856                         if (!test_range_bit(tree, em->start,
2857                                             extent_map_end(em) - 1,
2858                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
2859                                             0, NULL)) {
2860                                 remove_extent_mapping(map, em);
2861                                 /* once for the rb tree */
2862                                 free_extent_map(em);
2863                         }
2864                         start = extent_map_end(em);
2865                         write_unlock(&map->lock);
2866
2867                         /* once for us */
2868                         free_extent_map(em);
2869                 }
2870         }
2871         return try_release_extent_state(map, tree, page, mask);
2872 }
2873
2874 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2875                 get_extent_t *get_extent)
2876 {
2877         struct inode *inode = mapping->host;
2878         struct extent_state *cached_state = NULL;
2879         u64 start = iblock << inode->i_blkbits;
2880         sector_t sector = 0;
2881         size_t blksize = (1 << inode->i_blkbits);
2882         struct extent_map *em;
2883
2884         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2885                          0, &cached_state, GFP_NOFS);
2886         em = get_extent(inode, NULL, 0, start, blksize, 0);
2887         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start,
2888                              start + blksize - 1, &cached_state, GFP_NOFS);
2889         if (!em || IS_ERR(em))
2890                 return 0;
2891
2892         if (em->block_start > EXTENT_MAP_LAST_BYTE)
2893                 goto out;
2894
2895         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2896 out:
2897         free_extent_map(em);
2898         return sector;
2899 }
2900
2901 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2902                 __u64 start, __u64 len, get_extent_t *get_extent)
2903 {
2904         int ret = 0;
2905         u64 off = start;
2906         u64 max = start + len;
2907         u32 flags = 0;
2908         u32 found_type;
2909         u64 last;
2910         u64 disko = 0;
2911         struct btrfs_key found_key;
2912         struct extent_map *em = NULL;
2913         struct extent_state *cached_state = NULL;
2914         struct btrfs_path *path;
2915         struct btrfs_file_extent_item *item;
2916         int end = 0;
2917         u64 em_start = 0, em_len = 0;
2918         unsigned long emflags;
2919         int hole = 0;
2920
2921         if (len == 0)
2922                 return -EINVAL;
2923
2924         path = btrfs_alloc_path();
2925         if (!path)
2926                 return -ENOMEM;
2927         path->leave_spinning = 1;
2928
2929         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
2930                                        path, inode->i_ino, -1, 0);
2931         if (ret < 0) {
2932                 btrfs_free_path(path);
2933                 return ret;
2934         }
2935         WARN_ON(!ret);
2936         path->slots[0]--;
2937         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2938                               struct btrfs_file_extent_item);
2939         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
2940         found_type = btrfs_key_type(&found_key);
2941
2942         /* No extents, just return */
2943         if (found_key.objectid != inode->i_ino ||
2944             found_type != BTRFS_EXTENT_DATA_KEY) {
2945                 btrfs_free_path(path);
2946                 return 0;
2947         }
2948         last = found_key.offset;
2949         btrfs_free_path(path);
2950
2951         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
2952                          &cached_state, GFP_NOFS);
2953         em = get_extent(inode, NULL, 0, off, max - off, 0);
2954         if (!em)
2955                 goto out;
2956         if (IS_ERR(em)) {
2957                 ret = PTR_ERR(em);
2958                 goto out;
2959         }
2960
2961         while (!end) {
2962                 hole = 0;
2963                 off = em->start + em->len;
2964                 if (off >= max)
2965                         end = 1;
2966
2967                 if (em->block_start == EXTENT_MAP_HOLE) {
2968                         hole = 1;
2969                         goto next;
2970                 }
2971
2972                 em_start = em->start;
2973                 em_len = em->len;
2974
2975                 disko = 0;
2976                 flags = 0;
2977
2978                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
2979                         end = 1;
2980                         flags |= FIEMAP_EXTENT_LAST;
2981                 } else if (em->block_start == EXTENT_MAP_INLINE) {
2982                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
2983                                   FIEMAP_EXTENT_NOT_ALIGNED);
2984                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
2985                         flags |= (FIEMAP_EXTENT_DELALLOC |
2986                                   FIEMAP_EXTENT_UNKNOWN);
2987                 } else {
2988                         disko = em->block_start;
2989                 }
2990                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2991                         flags |= FIEMAP_EXTENT_ENCODED;
2992
2993 next:
2994                 emflags = em->flags;
2995                 free_extent_map(em);
2996                 em = NULL;
2997                 if (!end) {
2998                         em = get_extent(inode, NULL, 0, off, max - off, 0);
2999                         if (!em)
3000                                 goto out;
3001                         if (IS_ERR(em)) {
3002                                 ret = PTR_ERR(em);
3003                                 goto out;
3004                         }
3005                         emflags = em->flags;
3006                 }
3007
3008                 if (test_bit(EXTENT_FLAG_VACANCY, &emflags)) {
3009                         flags |= FIEMAP_EXTENT_LAST;
3010                         end = 1;
3011                 }
3012
3013                 if (em_start == last) {
3014                         flags |= FIEMAP_EXTENT_LAST;
3015                         end = 1;
3016                 }
3017
3018                 if (!hole) {
3019                         ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3020                                                 em_len, flags);
3021                         if (ret)
3022                                 goto out_free;
3023                 }
3024         }
3025 out_free:
3026         free_extent_map(em);
3027 out:
3028         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3029                              &cached_state, GFP_NOFS);
3030         return ret;
3031 }
3032
3033 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
3034                                               unsigned long i)
3035 {
3036         struct page *p;
3037         struct address_space *mapping;
3038
3039         if (i == 0)
3040                 return eb->first_page;
3041         i += eb->start >> PAGE_CACHE_SHIFT;
3042         mapping = eb->first_page->mapping;
3043         if (!mapping)
3044                 return NULL;
3045
3046         /*
3047          * extent_buffer_page is only called after pinning the page
3048          * by increasing the reference count.  So we know the page must
3049          * be in the radix tree.
3050          */
3051         rcu_read_lock();
3052         p = radix_tree_lookup(&mapping->page_tree, i);
3053         rcu_read_unlock();
3054
3055         return p;
3056 }
3057
3058 static inline unsigned long num_extent_pages(u64 start, u64 len)
3059 {
3060         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3061                 (start >> PAGE_CACHE_SHIFT);
3062 }
3063
3064 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3065                                                    u64 start,
3066                                                    unsigned long len,
3067                                                    gfp_t mask)
3068 {
3069         struct extent_buffer *eb = NULL;
3070 #if LEAK_DEBUG
3071         unsigned long flags;
3072 #endif
3073
3074         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3075         eb->start = start;
3076         eb->len = len;
3077         spin_lock_init(&eb->lock);
3078         init_waitqueue_head(&eb->lock_wq);
3079         INIT_RCU_HEAD(&eb->rcu_head);
3080
3081 #if LEAK_DEBUG
3082         spin_lock_irqsave(&leak_lock, flags);
3083         list_add(&eb->leak_list, &buffers);
3084         spin_unlock_irqrestore(&leak_lock, flags);
3085 #endif
3086         atomic_set(&eb->refs, 1);
3087
3088         return eb;
3089 }
3090
3091 static void __free_extent_buffer(struct extent_buffer *eb)
3092 {
3093 #if LEAK_DEBUG
3094         unsigned long flags;
3095         spin_lock_irqsave(&leak_lock, flags);
3096         list_del(&eb->leak_list);
3097         spin_unlock_irqrestore(&leak_lock, flags);
3098 #endif
3099         kmem_cache_free(extent_buffer_cache, eb);
3100 }
3101
3102 /*
3103  * Helper for releasing extent buffer page.
3104  */
3105 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3106                                                 unsigned long start_idx)
3107 {
3108         unsigned long index;
3109         struct page *page;
3110
3111         if (!eb->first_page)
3112                 return;
3113
3114         index = num_extent_pages(eb->start, eb->len);
3115         if (start_idx >= index)
3116                 return;
3117
3118         do {
3119                 index--;
3120                 page = extent_buffer_page(eb, index);
3121                 if (page)
3122                         page_cache_release(page);
3123         } while (index != start_idx);
3124 }
3125
3126 /*
3127  * Helper for releasing the extent buffer.
3128  */
3129 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3130 {
3131         btrfs_release_extent_buffer_page(eb, 0);
3132         __free_extent_buffer(eb);
3133 }
3134
3135 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3136                                           u64 start, unsigned long len,
3137                                           struct page *page0,
3138                                           gfp_t mask)
3139 {
3140         unsigned long num_pages = num_extent_pages(start, len);
3141         unsigned long i;
3142         unsigned long index = start >> PAGE_CACHE_SHIFT;
3143         struct extent_buffer *eb;
3144         struct extent_buffer *exists = NULL;
3145         struct page *p;
3146         struct address_space *mapping = tree->mapping;
3147         int uptodate = 1;
3148         int ret;
3149
3150         rcu_read_lock();
3151         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3152         if (eb && atomic_inc_not_zero(&eb->refs)) {
3153                 rcu_read_unlock();
3154                 mark_page_accessed(eb->first_page);
3155                 return eb;
3156         }
3157         rcu_read_unlock();
3158
3159         eb = __alloc_extent_buffer(tree, start, len, mask);
3160         if (!eb)
3161                 return NULL;
3162
3163         if (page0) {
3164                 eb->first_page = page0;
3165                 i = 1;
3166                 index++;
3167                 page_cache_get(page0);
3168                 mark_page_accessed(page0);
3169                 set_page_extent_mapped(page0);
3170                 set_page_extent_head(page0, len);
3171                 uptodate = PageUptodate(page0);
3172         } else {
3173                 i = 0;
3174         }
3175         for (; i < num_pages; i++, index++) {
3176                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
3177                 if (!p) {
3178                         WARN_ON(1);
3179                         goto free_eb;
3180                 }
3181                 set_page_extent_mapped(p);
3182                 mark_page_accessed(p);
3183                 if (i == 0) {
3184                         eb->first_page = p;
3185                         set_page_extent_head(p, len);
3186                 } else {
3187                         set_page_private(p, EXTENT_PAGE_PRIVATE);
3188                 }
3189                 if (!PageUptodate(p))
3190                         uptodate = 0;
3191                 unlock_page(p);
3192         }
3193         if (uptodate)
3194                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3195
3196         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
3197         if (ret)
3198                 goto free_eb;
3199
3200         spin_lock(&tree->buffer_lock);
3201         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
3202         if (ret == -EEXIST) {
3203                 exists = radix_tree_lookup(&tree->buffer,
3204                                                 start >> PAGE_CACHE_SHIFT);
3205                 /* add one reference for the caller */
3206                 atomic_inc(&exists->refs);
3207                 spin_unlock(&tree->buffer_lock);
3208                 radix_tree_preload_end();
3209                 goto free_eb;
3210         }
3211         /* add one reference for the tree */
3212         atomic_inc(&eb->refs);
3213         spin_unlock(&tree->buffer_lock);
3214         radix_tree_preload_end();
3215         return eb;
3216
3217 free_eb:
3218         if (!atomic_dec_and_test(&eb->refs))
3219                 return exists;
3220         btrfs_release_extent_buffer(eb);
3221         return exists;
3222 }
3223
3224 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3225                                          u64 start, unsigned long len,
3226                                           gfp_t mask)
3227 {
3228         struct extent_buffer *eb;
3229
3230         rcu_read_lock();
3231         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3232         if (eb && atomic_inc_not_zero(&eb->refs)) {
3233                 rcu_read_unlock();
3234                 mark_page_accessed(eb->first_page);
3235                 return eb;
3236         }
3237         rcu_read_unlock();
3238
3239         return NULL;
3240 }
3241
3242 void free_extent_buffer(struct extent_buffer *eb)
3243 {
3244         if (!eb)
3245                 return;
3246
3247         if (!atomic_dec_and_test(&eb->refs))
3248                 return;
3249
3250         WARN_ON(1);
3251 }
3252
3253 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3254                               struct extent_buffer *eb)
3255 {
3256         unsigned long i;
3257         unsigned long num_pages;
3258         struct page *page;
3259
3260         num_pages = num_extent_pages(eb->start, eb->len);
3261
3262         for (i = 0; i < num_pages; i++) {
3263                 page = extent_buffer_page(eb, i);
3264                 if (!PageDirty(page))
3265                         continue;
3266
3267                 lock_page(page);
3268                 if (i == 0)
3269                         set_page_extent_head(page, eb->len);
3270                 else
3271                         set_page_private(page, EXTENT_PAGE_PRIVATE);
3272
3273                 clear_page_dirty_for_io(page);
3274                 spin_lock_irq(&page->mapping->tree_lock);
3275                 if (!PageDirty(page)) {
3276                         radix_tree_tag_clear(&page->mapping->page_tree,
3277                                                 page_index(page),
3278                                                 PAGECACHE_TAG_DIRTY);
3279                 }
3280                 spin_unlock_irq(&page->mapping->tree_lock);
3281                 unlock_page(page);
3282         }
3283         return 0;
3284 }
3285
3286 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
3287                                     struct extent_buffer *eb)
3288 {
3289         return wait_on_extent_writeback(tree, eb->start,
3290                                         eb->start + eb->len - 1);
3291 }
3292
3293 int set_extent_buffer_dirty(struct extent_io_tree *tree,
3294                              struct extent_buffer *eb)
3295 {
3296         unsigned long i;
3297         unsigned long num_pages;
3298         int was_dirty = 0;
3299
3300         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3301         num_pages = num_extent_pages(eb->start, eb->len);
3302         for (i = 0; i < num_pages; i++)
3303                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3304         return was_dirty;
3305 }
3306
3307 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3308                                 struct extent_buffer *eb,
3309                                 struct extent_state **cached_state)
3310 {
3311         unsigned long i;
3312         struct page *page;
3313         unsigned long num_pages;
3314
3315         num_pages = num_extent_pages(eb->start, eb->len);
3316         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3317
3318         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3319                               cached_state, GFP_NOFS);
3320         for (i = 0; i < num_pages; i++) {
3321                 page = extent_buffer_page(eb, i);
3322                 if (page)
3323                         ClearPageUptodate(page);
3324         }
3325         return 0;
3326 }
3327
3328 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3329                                 struct extent_buffer *eb)
3330 {
3331         unsigned long i;
3332         struct page *page;
3333         unsigned long num_pages;
3334
3335         num_pages = num_extent_pages(eb->start, eb->len);
3336
3337         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3338                             GFP_NOFS);
3339         for (i = 0; i < num_pages; i++) {
3340                 page = extent_buffer_page(eb, i);
3341                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3342                     ((i == num_pages - 1) &&
3343                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3344                         check_page_uptodate(tree, page);
3345                         continue;
3346                 }
3347                 SetPageUptodate(page);
3348         }
3349         return 0;
3350 }
3351
3352 int extent_range_uptodate(struct extent_io_tree *tree,
3353                           u64 start, u64 end)
3354 {
3355         struct page *page;
3356         int ret;
3357         int pg_uptodate = 1;
3358         int uptodate;
3359         unsigned long index;
3360
3361         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL);
3362         if (ret)
3363                 return 1;
3364         while (start <= end) {
3365                 index = start >> PAGE_CACHE_SHIFT;
3366                 page = find_get_page(tree->mapping, index);
3367                 uptodate = PageUptodate(page);
3368                 page_cache_release(page);
3369                 if (!uptodate) {
3370                         pg_uptodate = 0;
3371                         break;
3372                 }
3373                 start += PAGE_CACHE_SIZE;
3374         }
3375         return pg_uptodate;
3376 }
3377
3378 int extent_buffer_uptodate(struct extent_io_tree *tree,
3379                            struct extent_buffer *eb,
3380                            struct extent_state *cached_state)
3381 {
3382         int ret = 0;
3383         unsigned long num_pages;
3384         unsigned long i;
3385         struct page *page;
3386         int pg_uptodate = 1;
3387
3388         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3389                 return 1;
3390
3391         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3392                            EXTENT_UPTODATE, 1, cached_state);
3393         if (ret)
3394                 return ret;
3395
3396         num_pages = num_extent_pages(eb->start, eb->len);
3397         for (i = 0; i < num_pages; i++) {
3398                 page = extent_buffer_page(eb, i);
3399                 if (!PageUptodate(page)) {
3400                         pg_uptodate = 0;
3401                         break;
3402                 }
3403         }
3404         return pg_uptodate;
3405 }
3406
3407 int read_extent_buffer_pages(struct extent_io_tree *tree,
3408                              struct extent_buffer *eb,
3409                              u64 start, int wait,
3410                              get_extent_t *get_extent, int mirror_num)
3411 {
3412         unsigned long i;
3413         unsigned long start_i;
3414         struct page *page;
3415         int err;
3416         int ret = 0;
3417         int locked_pages = 0;
3418         int all_uptodate = 1;
3419         int inc_all_pages = 0;
3420         unsigned long num_pages;
3421         struct bio *bio = NULL;
3422         unsigned long bio_flags = 0;
3423
3424         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3425                 return 0;
3426
3427         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3428                            EXTENT_UPTODATE, 1, NULL)) {
3429                 return 0;
3430         }
3431
3432         if (start) {
3433                 WARN_ON(start < eb->start);
3434                 start_i = (start >> PAGE_CACHE_SHIFT) -
3435                         (eb->start >> PAGE_CACHE_SHIFT);
3436         } else {
3437                 start_i = 0;
3438         }
3439
3440         num_pages = num_extent_pages(eb->start, eb->len);
3441         for (i = start_i; i < num_pages; i++) {
3442                 page = extent_buffer_page(eb, i);
3443                 if (!wait) {
3444                         if (!trylock_page(page))
3445                                 goto unlock_exit;
3446                 } else {
3447                         lock_page(page);
3448                 }
3449                 locked_pages++;
3450                 if (!PageUptodate(page))
3451                         all_uptodate = 0;
3452         }
3453         if (all_uptodate) {
3454                 if (start_i == 0)
3455                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3456                 goto unlock_exit;
3457         }
3458
3459         for (i = start_i; i < num_pages; i++) {
3460                 page = extent_buffer_page(eb, i);
3461                 if (inc_all_pages)
3462                         page_cache_get(page);
3463                 if (!PageUptodate(page)) {
3464                         if (start_i == 0)
3465                                 inc_all_pages = 1;
3466                         ClearPageError(page);
3467                         err = __extent_read_full_page(tree, page,
3468                                                       get_extent, &bio,
3469                                                       mirror_num, &bio_flags);
3470                         if (err)
3471                                 ret = err;
3472                 } else {
3473                         unlock_page(page);
3474                 }
3475         }
3476
3477         if (bio)
3478                 submit_one_bio(READ, bio, mirror_num, bio_flags);
3479
3480         if (ret || !wait)
3481                 return ret;
3482
3483         for (i = start_i; i < num_pages; i++) {
3484                 page = extent_buffer_page(eb, i);
3485                 wait_on_page_locked(page);
3486                 if (!PageUptodate(page))
3487                         ret = -EIO;
3488         }
3489
3490         if (!ret)
3491                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3492         return ret;
3493
3494 unlock_exit:
3495         i = start_i;
3496         while (locked_pages > 0) {
3497                 page = extent_buffer_page(eb, i);
3498                 i++;
3499                 unlock_page(page);
3500                 locked_pages--;
3501         }
3502         return ret;
3503 }
3504
3505 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3506                         unsigned long start,
3507                         unsigned long len)
3508 {
3509         size_t cur;
3510         size_t offset;
3511         struct page *page;
3512         char *kaddr;
3513         char *dst = (char *)dstv;
3514         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3515         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3516
3517         WARN_ON(start > eb->len);
3518         WARN_ON(start + len > eb->start + eb->len);
3519
3520         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3521
3522         while (len > 0) {
3523                 page = extent_buffer_page(eb, i);
3524
3525                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3526                 kaddr = kmap_atomic(page, KM_USER1);
3527                 memcpy(dst, kaddr + offset, cur);
3528                 kunmap_atomic(kaddr, KM_USER1);
3529
3530                 dst += cur;
3531                 len -= cur;
3532                 offset = 0;
3533                 i++;
3534         }
3535 }
3536
3537 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3538                                unsigned long min_len, char **token, char **map,
3539                                unsigned long *map_start,
3540                                unsigned long *map_len, int km)
3541 {
3542         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3543         char *kaddr;
3544         struct page *p;
3545         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3546         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3547         unsigned long end_i = (start_offset + start + min_len - 1) >>
3548                 PAGE_CACHE_SHIFT;
3549
3550         if (i != end_i)
3551                 return -EINVAL;
3552
3553         if (i == 0) {
3554                 offset = start_offset;
3555                 *map_start = 0;
3556         } else {
3557                 offset = 0;
3558                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3559         }
3560
3561         if (start + min_len > eb->len) {
3562                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3563                        "wanted %lu %lu\n", (unsigned long long)eb->start,
3564                        eb->len, start, min_len);
3565                 WARN_ON(1);
3566         }
3567
3568         p = extent_buffer_page(eb, i);
3569         kaddr = kmap_atomic(p, km);
3570         *token = kaddr;
3571         *map = kaddr + offset;
3572         *map_len = PAGE_CACHE_SIZE - offset;
3573         return 0;
3574 }
3575
3576 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3577                       unsigned long min_len,
3578                       char **token, char **map,
3579                       unsigned long *map_start,
3580                       unsigned long *map_len, int km)
3581 {
3582         int err;
3583         int save = 0;
3584         if (eb->map_token) {
3585                 unmap_extent_buffer(eb, eb->map_token, km);
3586                 eb->map_token = NULL;
3587                 save = 1;
3588         }
3589         err = map_private_extent_buffer(eb, start, min_len, token, map,
3590                                        map_start, map_len, km);
3591         if (!err && save) {
3592                 eb->map_token = *token;
3593                 eb->kaddr = *map;
3594                 eb->map_start = *map_start;
3595                 eb->map_len = *map_len;
3596         }
3597         return err;
3598 }
3599
3600 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3601 {
3602         kunmap_atomic(token, km);
3603 }
3604
3605 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3606                           unsigned long start,
3607                           unsigned long len)
3608 {
3609         size_t cur;
3610         size_t offset;
3611         struct page *page;
3612         char *kaddr;
3613         char *ptr = (char *)ptrv;
3614         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3615         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3616         int ret = 0;
3617
3618         WARN_ON(start > eb->len);
3619         WARN_ON(start + len > eb->start + eb->len);
3620
3621         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3622
3623         while (len > 0) {
3624                 page = extent_buffer_page(eb, i);
3625
3626                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3627
3628                 kaddr = kmap_atomic(page, KM_USER0);
3629                 ret = memcmp(ptr, kaddr + offset, cur);
3630                 kunmap_atomic(kaddr, KM_USER0);
3631                 if (ret)
3632                         break;
3633
3634                 ptr += cur;
3635                 len -= cur;
3636                 offset = 0;
3637                 i++;
3638         }
3639         return ret;
3640 }
3641
3642 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3643                          unsigned long start, unsigned long len)
3644 {
3645         size_t cur;
3646         size_t offset;
3647         struct page *page;
3648         char *kaddr;
3649         char *src = (char *)srcv;
3650         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3651         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3652
3653         WARN_ON(start > eb->len);
3654         WARN_ON(start + len > eb->start + eb->len);
3655
3656         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3657
3658         while (len > 0) {
3659                 page = extent_buffer_page(eb, i);
3660                 WARN_ON(!PageUptodate(page));
3661
3662                 cur = min(len, PAGE_CACHE_SIZE - offset);
3663                 kaddr = kmap_atomic(page, KM_USER1);
3664                 memcpy(kaddr + offset, src, cur);
3665                 kunmap_atomic(kaddr, KM_USER1);
3666
3667                 src += cur;
3668                 len -= cur;
3669                 offset = 0;
3670                 i++;
3671         }
3672 }
3673
3674 void memset_extent_buffer(struct extent_buffer *eb, char c,
3675                           unsigned long start, unsigned long len)
3676 {
3677         size_t cur;
3678         size_t offset;
3679         struct page *page;
3680         char *kaddr;
3681         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3682         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3683
3684         WARN_ON(start > eb->len);
3685         WARN_ON(start + len > eb->start + eb->len);
3686
3687         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3688
3689         while (len > 0) {
3690                 page = extent_buffer_page(eb, i);
3691                 WARN_ON(!PageUptodate(page));
3692
3693                 cur = min(len, PAGE_CACHE_SIZE - offset);
3694                 kaddr = kmap_atomic(page, KM_USER0);
3695                 memset(kaddr + offset, c, cur);
3696                 kunmap_atomic(kaddr, KM_USER0);
3697
3698                 len -= cur;
3699                 offset = 0;
3700                 i++;
3701         }
3702 }
3703
3704 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3705                         unsigned long dst_offset, unsigned long src_offset,
3706                         unsigned long len)
3707 {
3708         u64 dst_len = dst->len;
3709         size_t cur;
3710         size_t offset;
3711         struct page *page;
3712         char *kaddr;
3713         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3714         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3715
3716         WARN_ON(src->len != dst_len);
3717
3718         offset = (start_offset + dst_offset) &
3719                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3720
3721         while (len > 0) {
3722                 page = extent_buffer_page(dst, i);
3723                 WARN_ON(!PageUptodate(page));
3724
3725                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3726
3727                 kaddr = kmap_atomic(page, KM_USER0);
3728                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3729                 kunmap_atomic(kaddr, KM_USER0);
3730
3731                 src_offset += cur;
3732                 len -= cur;
3733                 offset = 0;
3734                 i++;
3735         }
3736 }
3737
3738 static void move_pages(struct page *dst_page, struct page *src_page,
3739                        unsigned long dst_off, unsigned long src_off,
3740                        unsigned long len)
3741 {
3742         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3743         if (dst_page == src_page) {
3744                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3745         } else {
3746                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3747                 char *p = dst_kaddr + dst_off + len;
3748                 char *s = src_kaddr + src_off + len;
3749
3750                 while (len--)
3751                         *--p = *--s;
3752
3753                 kunmap_atomic(src_kaddr, KM_USER1);
3754         }
3755         kunmap_atomic(dst_kaddr, KM_USER0);
3756 }
3757
3758 static void copy_pages(struct page *dst_page, struct page *src_page,
3759                        unsigned long dst_off, unsigned long src_off,
3760                        unsigned long len)
3761 {
3762         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3763         char *src_kaddr;
3764
3765         if (dst_page != src_page)
3766                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3767         else
3768                 src_kaddr = dst_kaddr;
3769
3770         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3771         kunmap_atomic(dst_kaddr, KM_USER0);
3772         if (dst_page != src_page)
3773                 kunmap_atomic(src_kaddr, KM_USER1);
3774 }
3775
3776 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3777                            unsigned long src_offset, unsigned long len)
3778 {
3779         size_t cur;
3780         size_t dst_off_in_page;
3781         size_t src_off_in_page;
3782         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3783         unsigned long dst_i;
3784         unsigned long src_i;
3785
3786         if (src_offset + len > dst->len) {
3787                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3788                        "len %lu dst len %lu\n", src_offset, len, dst->len);
3789                 BUG_ON(1);
3790         }
3791         if (dst_offset + len > dst->len) {
3792                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3793                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
3794                 BUG_ON(1);
3795         }
3796
3797         while (len > 0) {
3798                 dst_off_in_page = (start_offset + dst_offset) &
3799                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3800                 src_off_in_page = (start_offset + src_offset) &
3801                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3802
3803                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3804                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3805
3806                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3807                                                src_off_in_page));
3808                 cur = min_t(unsigned long, cur,
3809                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3810
3811                 copy_pages(extent_buffer_page(dst, dst_i),
3812                            extent_buffer_page(dst, src_i),
3813                            dst_off_in_page, src_off_in_page, cur);
3814
3815                 src_offset += cur;
3816                 dst_offset += cur;
3817                 len -= cur;
3818         }
3819 }
3820
3821 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3822                            unsigned long src_offset, unsigned long len)
3823 {
3824         size_t cur;
3825         size_t dst_off_in_page;
3826         size_t src_off_in_page;
3827         unsigned long dst_end = dst_offset + len - 1;
3828         unsigned long src_end = src_offset + len - 1;
3829         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3830         unsigned long dst_i;
3831         unsigned long src_i;
3832
3833         if (src_offset + len > dst->len) {
3834                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3835                        "len %lu len %lu\n", src_offset, len, dst->len);
3836                 BUG_ON(1);
3837         }
3838         if (dst_offset + len > dst->len) {
3839                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3840                        "len %lu len %lu\n", dst_offset, len, dst->len);
3841                 BUG_ON(1);
3842         }
3843         if (dst_offset < src_offset) {
3844                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3845                 return;
3846         }
3847         while (len > 0) {
3848                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3849                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3850
3851                 dst_off_in_page = (start_offset + dst_end) &
3852                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3853                 src_off_in_page = (start_offset + src_end) &
3854                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3855
3856                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3857                 cur = min(cur, dst_off_in_page + 1);
3858                 move_pages(extent_buffer_page(dst, dst_i),
3859                            extent_buffer_page(dst, src_i),
3860                            dst_off_in_page - cur + 1,
3861                            src_off_in_page - cur + 1, cur);
3862
3863                 dst_end -= cur;
3864                 src_end -= cur;
3865                 len -= cur;
3866         }
3867 }
3868
3869 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3870 {
3871         struct extent_buffer *eb =
3872                         container_of(head, struct extent_buffer, rcu_head);
3873
3874         btrfs_release_extent_buffer(eb);
3875 }
3876
3877 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3878 {
3879         u64 start = page_offset(page);
3880         struct extent_buffer *eb;
3881         int ret = 1;
3882
3883         spin_lock(&tree->buffer_lock);
3884         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3885         if (!eb) {
3886                 spin_unlock(&tree->buffer_lock);
3887                 return ret;
3888         }
3889
3890         if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3891                 ret = 0;
3892                 goto out;
3893         }
3894
3895         /*
3896          * set @eb->refs to 0 if it is already 1, and then release the @eb.
3897          * Or go back.
3898          */
3899         if (atomic_cmpxchg(&eb->refs, 1, 0) != 1) {
3900                 ret = 0;
3901                 goto out;
3902         }
3903
3904         radix_tree_delete(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3905 out:
3906         spin_unlock(&tree->buffer_lock);
3907
3908         /* at this point we can safely release the extent buffer */
3909         if (atomic_read(&eb->refs) == 0)
3910                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3911         return ret;
3912 }