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                         extent_set_compress_type(&this_bio_flag,
2034                                                  em->compress_type);
2035                 }
2036
2037                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2038                 cur_end = min(extent_map_end(em) - 1, end);
2039                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2040                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2041                         disk_io_size = em->block_len;
2042                         sector = em->block_start >> 9;
2043                 } else {
2044                         sector = (em->block_start + extent_offset) >> 9;
2045                         disk_io_size = iosize;
2046                 }
2047                 bdev = em->bdev;
2048                 block_start = em->block_start;
2049                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2050                         block_start = EXTENT_MAP_HOLE;
2051                 free_extent_map(em);
2052                 em = NULL;
2053
2054                 /* we've found a hole, just zero and go on */
2055                 if (block_start == EXTENT_MAP_HOLE) {
2056                         char *userpage;
2057                         userpage = kmap_atomic(page, KM_USER0);
2058                         memset(userpage + page_offset, 0, iosize);
2059                         flush_dcache_page(page);
2060                         kunmap_atomic(userpage, KM_USER0);
2061
2062                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2063                                             GFP_NOFS);
2064                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2065                         cur = cur + iosize;
2066                         page_offset += iosize;
2067                         continue;
2068                 }
2069                 /* the get_extent function already copied into the page */
2070                 if (test_range_bit(tree, cur, cur_end,
2071                                    EXTENT_UPTODATE, 1, NULL)) {
2072                         check_page_uptodate(tree, page);
2073                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2074                         cur = cur + iosize;
2075                         page_offset += iosize;
2076                         continue;
2077                 }
2078                 /* we have an inline extent but it didn't get marked up
2079                  * to date.  Error out
2080                  */
2081                 if (block_start == EXTENT_MAP_INLINE) {
2082                         SetPageError(page);
2083                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2084                         cur = cur + iosize;
2085                         page_offset += iosize;
2086                         continue;
2087                 }
2088
2089                 ret = 0;
2090                 if (tree->ops && tree->ops->readpage_io_hook) {
2091                         ret = tree->ops->readpage_io_hook(page, cur,
2092                                                           cur + iosize - 1);
2093                 }
2094                 if (!ret) {
2095                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2096                         pnr -= page->index;
2097                         ret = submit_extent_page(READ, tree, page,
2098                                          sector, disk_io_size, page_offset,
2099                                          bdev, bio, pnr,
2100                                          end_bio_extent_readpage, mirror_num,
2101                                          *bio_flags,
2102                                          this_bio_flag);
2103                         nr++;
2104                         *bio_flags = this_bio_flag;
2105                 }
2106                 if (ret)
2107                         SetPageError(page);
2108                 cur = cur + iosize;
2109                 page_offset += iosize;
2110         }
2111         if (!nr) {
2112                 if (!PageError(page))
2113                         SetPageUptodate(page);
2114                 unlock_page(page);
2115         }
2116         return 0;
2117 }
2118
2119 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2120                             get_extent_t *get_extent)
2121 {
2122         struct bio *bio = NULL;
2123         unsigned long bio_flags = 0;
2124         int ret;
2125
2126         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2127                                       &bio_flags);
2128         if (bio)
2129                 submit_one_bio(READ, bio, 0, bio_flags);
2130         return ret;
2131 }
2132
2133 static noinline void update_nr_written(struct page *page,
2134                                       struct writeback_control *wbc,
2135                                       unsigned long nr_written)
2136 {
2137         wbc->nr_to_write -= nr_written;
2138         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2139             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2140                 page->mapping->writeback_index = page->index + nr_written;
2141 }
2142
2143 /*
2144  * the writepage semantics are similar to regular writepage.  extent
2145  * records are inserted to lock ranges in the tree, and as dirty areas
2146  * are found, they are marked writeback.  Then the lock bits are removed
2147  * and the end_io handler clears the writeback ranges
2148  */
2149 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2150                               void *data)
2151 {
2152         struct inode *inode = page->mapping->host;
2153         struct extent_page_data *epd = data;
2154         struct extent_io_tree *tree = epd->tree;
2155         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2156         u64 delalloc_start;
2157         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2158         u64 end;
2159         u64 cur = start;
2160         u64 extent_offset;
2161         u64 last_byte = i_size_read(inode);
2162         u64 block_start;
2163         u64 iosize;
2164         sector_t sector;
2165         struct extent_state *cached_state = NULL;
2166         struct extent_map *em;
2167         struct block_device *bdev;
2168         int ret;
2169         int nr = 0;
2170         size_t pg_offset = 0;
2171         size_t blocksize;
2172         loff_t i_size = i_size_read(inode);
2173         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2174         u64 nr_delalloc;
2175         u64 delalloc_end;
2176         int page_started;
2177         int compressed;
2178         int write_flags;
2179         unsigned long nr_written = 0;
2180
2181         if (wbc->sync_mode == WB_SYNC_ALL)
2182                 write_flags = WRITE_SYNC_PLUG;
2183         else
2184                 write_flags = WRITE;
2185
2186         WARN_ON(!PageLocked(page));
2187         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2188         if (page->index > end_index ||
2189            (page->index == end_index && !pg_offset)) {
2190                 page->mapping->a_ops->invalidatepage(page, 0);
2191                 unlock_page(page);
2192                 return 0;
2193         }
2194
2195         if (page->index == end_index) {
2196                 char *userpage;
2197
2198                 userpage = kmap_atomic(page, KM_USER0);
2199                 memset(userpage + pg_offset, 0,
2200                        PAGE_CACHE_SIZE - pg_offset);
2201                 kunmap_atomic(userpage, KM_USER0);
2202                 flush_dcache_page(page);
2203         }
2204         pg_offset = 0;
2205
2206         set_page_extent_mapped(page);
2207
2208         delalloc_start = start;
2209         delalloc_end = 0;
2210         page_started = 0;
2211         if (!epd->extent_locked) {
2212                 u64 delalloc_to_write = 0;
2213                 /*
2214                  * make sure the wbc mapping index is at least updated
2215                  * to this page.
2216                  */
2217                 update_nr_written(page, wbc, 0);
2218
2219                 while (delalloc_end < page_end) {
2220                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2221                                                        page,
2222                                                        &delalloc_start,
2223                                                        &delalloc_end,
2224                                                        128 * 1024 * 1024);
2225                         if (nr_delalloc == 0) {
2226                                 delalloc_start = delalloc_end + 1;
2227                                 continue;
2228                         }
2229                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2230                                                  delalloc_end, &page_started,
2231                                                  &nr_written);
2232                         /*
2233                          * delalloc_end is already one less than the total
2234                          * length, so we don't subtract one from
2235                          * PAGE_CACHE_SIZE
2236                          */
2237                         delalloc_to_write += (delalloc_end - delalloc_start +
2238                                               PAGE_CACHE_SIZE) >>
2239                                               PAGE_CACHE_SHIFT;
2240                         delalloc_start = delalloc_end + 1;
2241                 }
2242                 if (wbc->nr_to_write < delalloc_to_write) {
2243                         int thresh = 8192;
2244
2245                         if (delalloc_to_write < thresh * 2)
2246                                 thresh = delalloc_to_write;
2247                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2248                                                  thresh);
2249                 }
2250
2251                 /* did the fill delalloc function already unlock and start
2252                  * the IO?
2253                  */
2254                 if (page_started) {
2255                         ret = 0;
2256                         /*
2257                          * we've unlocked the page, so we can't update
2258                          * the mapping's writeback index, just update
2259                          * nr_to_write.
2260                          */
2261                         wbc->nr_to_write -= nr_written;
2262                         goto done_unlocked;
2263                 }
2264         }
2265         if (tree->ops && tree->ops->writepage_start_hook) {
2266                 ret = tree->ops->writepage_start_hook(page, start,
2267                                                       page_end);
2268                 if (ret == -EAGAIN) {
2269                         redirty_page_for_writepage(wbc, page);
2270                         update_nr_written(page, wbc, nr_written);
2271                         unlock_page(page);
2272                         ret = 0;
2273                         goto done_unlocked;
2274                 }
2275         }
2276
2277         /*
2278          * we don't want to touch the inode after unlocking the page,
2279          * so we update the mapping writeback index now
2280          */
2281         update_nr_written(page, wbc, nr_written + 1);
2282
2283         end = page_end;
2284         if (last_byte <= start) {
2285                 if (tree->ops && tree->ops->writepage_end_io_hook)
2286                         tree->ops->writepage_end_io_hook(page, start,
2287                                                          page_end, NULL, 1);
2288                 goto done;
2289         }
2290
2291         blocksize = inode->i_sb->s_blocksize;
2292
2293         while (cur <= end) {
2294                 if (cur >= last_byte) {
2295                         if (tree->ops && tree->ops->writepage_end_io_hook)
2296                                 tree->ops->writepage_end_io_hook(page, cur,
2297                                                          page_end, NULL, 1);
2298                         break;
2299                 }
2300                 em = epd->get_extent(inode, page, pg_offset, cur,
2301                                      end - cur + 1, 1);
2302                 if (IS_ERR(em) || !em) {
2303                         SetPageError(page);
2304                         break;
2305                 }
2306
2307                 extent_offset = cur - em->start;
2308                 BUG_ON(extent_map_end(em) <= cur);
2309                 BUG_ON(end < cur);
2310                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2311                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2312                 sector = (em->block_start + extent_offset) >> 9;
2313                 bdev = em->bdev;
2314                 block_start = em->block_start;
2315                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2316                 free_extent_map(em);
2317                 em = NULL;
2318
2319                 /*
2320                  * compressed and inline extents are written through other
2321                  * paths in the FS
2322                  */
2323                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2324                     block_start == EXTENT_MAP_INLINE) {
2325                         /*
2326                          * end_io notification does not happen here for
2327                          * compressed extents
2328                          */
2329                         if (!compressed && tree->ops &&
2330                             tree->ops->writepage_end_io_hook)
2331                                 tree->ops->writepage_end_io_hook(page, cur,
2332                                                          cur + iosize - 1,
2333                                                          NULL, 1);
2334                         else if (compressed) {
2335                                 /* we don't want to end_page_writeback on
2336                                  * a compressed extent.  this happens
2337                                  * elsewhere
2338                                  */
2339                                 nr++;
2340                         }
2341
2342                         cur += iosize;
2343                         pg_offset += iosize;
2344                         continue;
2345                 }
2346                 /* leave this out until we have a page_mkwrite call */
2347                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2348                                    EXTENT_DIRTY, 0, NULL)) {
2349                         cur = cur + iosize;
2350                         pg_offset += iosize;
2351                         continue;
2352                 }
2353
2354                 if (tree->ops && tree->ops->writepage_io_hook) {
2355                         ret = tree->ops->writepage_io_hook(page, cur,
2356                                                 cur + iosize - 1);
2357                 } else {
2358                         ret = 0;
2359                 }
2360                 if (ret) {
2361                         SetPageError(page);
2362                 } else {
2363                         unsigned long max_nr = end_index + 1;
2364
2365                         set_range_writeback(tree, cur, cur + iosize - 1);
2366                         if (!PageWriteback(page)) {
2367                                 printk(KERN_ERR "btrfs warning page %lu not "
2368                                        "writeback, cur %llu end %llu\n",
2369                                        page->index, (unsigned long long)cur,
2370                                        (unsigned long long)end);
2371                         }
2372
2373                         ret = submit_extent_page(write_flags, tree, page,
2374                                                  sector, iosize, pg_offset,
2375                                                  bdev, &epd->bio, max_nr,
2376                                                  end_bio_extent_writepage,
2377                                                  0, 0, 0);
2378                         if (ret)
2379                                 SetPageError(page);
2380                 }
2381                 cur = cur + iosize;
2382                 pg_offset += iosize;
2383                 nr++;
2384         }
2385 done:
2386         if (nr == 0) {
2387                 /* make sure the mapping tag for page dirty gets cleared */
2388                 set_page_writeback(page);
2389                 end_page_writeback(page);
2390         }
2391         unlock_page(page);
2392
2393 done_unlocked:
2394
2395         /* drop our reference on any cached states */
2396         free_extent_state(cached_state);
2397         return 0;
2398 }
2399
2400 /**
2401  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2402  * @mapping: address space structure to write
2403  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2404  * @writepage: function called for each page
2405  * @data: data passed to writepage function
2406  *
2407  * If a page is already under I/O, write_cache_pages() skips it, even
2408  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2409  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2410  * and msync() need to guarantee that all the data which was dirty at the time
2411  * the call was made get new I/O started against them.  If wbc->sync_mode is
2412  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2413  * existing IO to complete.
2414  */
2415 static int extent_write_cache_pages(struct extent_io_tree *tree,
2416                              struct address_space *mapping,
2417                              struct writeback_control *wbc,
2418                              writepage_t writepage, void *data,
2419                              void (*flush_fn)(void *))
2420 {
2421         int ret = 0;
2422         int done = 0;
2423         int nr_to_write_done = 0;
2424         struct pagevec pvec;
2425         int nr_pages;
2426         pgoff_t index;
2427         pgoff_t end;            /* Inclusive */
2428         int scanned = 0;
2429
2430         pagevec_init(&pvec, 0);
2431         if (wbc->range_cyclic) {
2432                 index = mapping->writeback_index; /* Start from prev offset */
2433                 end = -1;
2434         } else {
2435                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2436                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2437                 scanned = 1;
2438         }
2439 retry:
2440         while (!done && !nr_to_write_done && (index <= end) &&
2441                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2442                               PAGECACHE_TAG_DIRTY, min(end - index,
2443                                   (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2444                 unsigned i;
2445
2446                 scanned = 1;
2447                 for (i = 0; i < nr_pages; i++) {
2448                         struct page *page = pvec.pages[i];
2449
2450                         /*
2451                          * At this point we hold neither mapping->tree_lock nor
2452                          * lock on the page itself: the page may be truncated or
2453                          * invalidated (changing page->mapping to NULL), or even
2454                          * swizzled back from swapper_space to tmpfs file
2455                          * mapping
2456                          */
2457                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2458                                 tree->ops->write_cache_pages_lock_hook(page);
2459                         else
2460                                 lock_page(page);
2461
2462                         if (unlikely(page->mapping != mapping)) {
2463                                 unlock_page(page);
2464                                 continue;
2465                         }
2466
2467                         if (!wbc->range_cyclic && page->index > end) {
2468                                 done = 1;
2469                                 unlock_page(page);
2470                                 continue;
2471                         }
2472
2473                         if (wbc->sync_mode != WB_SYNC_NONE) {
2474                                 if (PageWriteback(page))
2475                                         flush_fn(data);
2476                                 wait_on_page_writeback(page);
2477                         }
2478
2479                         if (PageWriteback(page) ||
2480                             !clear_page_dirty_for_io(page)) {
2481                                 unlock_page(page);
2482                                 continue;
2483                         }
2484
2485                         ret = (*writepage)(page, wbc, data);
2486
2487                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2488                                 unlock_page(page);
2489                                 ret = 0;
2490                         }
2491                         if (ret)
2492                                 done = 1;
2493
2494                         /*
2495                          * the filesystem may choose to bump up nr_to_write.
2496                          * We have to make sure to honor the new nr_to_write
2497                          * at any time
2498                          */
2499                         nr_to_write_done = wbc->nr_to_write <= 0;
2500                 }
2501                 pagevec_release(&pvec);
2502                 cond_resched();
2503         }
2504         if (!scanned && !done) {
2505                 /*
2506                  * We hit the last page and there is more work to be done: wrap
2507                  * back to the start of the file
2508                  */
2509                 scanned = 1;
2510                 index = 0;
2511                 goto retry;
2512         }
2513         return ret;
2514 }
2515
2516 static void flush_epd_write_bio(struct extent_page_data *epd)
2517 {
2518         if (epd->bio) {
2519                 if (epd->sync_io)
2520                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2521                 else
2522                         submit_one_bio(WRITE, epd->bio, 0, 0);
2523                 epd->bio = NULL;
2524         }
2525 }
2526
2527 static noinline void flush_write_bio(void *data)
2528 {
2529         struct extent_page_data *epd = data;
2530         flush_epd_write_bio(epd);
2531 }
2532
2533 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2534                           get_extent_t *get_extent,
2535                           struct writeback_control *wbc)
2536 {
2537         int ret;
2538         struct address_space *mapping = page->mapping;
2539         struct extent_page_data epd = {
2540                 .bio = NULL,
2541                 .tree = tree,
2542                 .get_extent = get_extent,
2543                 .extent_locked = 0,
2544                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2545         };
2546         struct writeback_control wbc_writepages = {
2547                 .sync_mode      = wbc->sync_mode,
2548                 .older_than_this = NULL,
2549                 .nr_to_write    = 64,
2550                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2551                 .range_end      = (loff_t)-1,
2552         };
2553
2554         ret = __extent_writepage(page, wbc, &epd);
2555
2556         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2557                                  __extent_writepage, &epd, flush_write_bio);
2558         flush_epd_write_bio(&epd);
2559         return ret;
2560 }
2561
2562 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2563                               u64 start, u64 end, get_extent_t *get_extent,
2564                               int mode)
2565 {
2566         int ret = 0;
2567         struct address_space *mapping = inode->i_mapping;
2568         struct page *page;
2569         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2570                 PAGE_CACHE_SHIFT;
2571
2572         struct extent_page_data epd = {
2573                 .bio = NULL,
2574                 .tree = tree,
2575                 .get_extent = get_extent,
2576                 .extent_locked = 1,
2577                 .sync_io = mode == WB_SYNC_ALL,
2578         };
2579         struct writeback_control wbc_writepages = {
2580                 .sync_mode      = mode,
2581                 .older_than_this = NULL,
2582                 .nr_to_write    = nr_pages * 2,
2583                 .range_start    = start,
2584                 .range_end      = end + 1,
2585         };
2586
2587         while (start <= end) {
2588                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2589                 if (clear_page_dirty_for_io(page))
2590                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2591                 else {
2592                         if (tree->ops && tree->ops->writepage_end_io_hook)
2593                                 tree->ops->writepage_end_io_hook(page, start,
2594                                                  start + PAGE_CACHE_SIZE - 1,
2595                                                  NULL, 1);
2596                         unlock_page(page);
2597                 }
2598                 page_cache_release(page);
2599                 start += PAGE_CACHE_SIZE;
2600         }
2601
2602         flush_epd_write_bio(&epd);
2603         return ret;
2604 }
2605
2606 int extent_writepages(struct extent_io_tree *tree,
2607                       struct address_space *mapping,
2608                       get_extent_t *get_extent,
2609                       struct writeback_control *wbc)
2610 {
2611         int ret = 0;
2612         struct extent_page_data epd = {
2613                 .bio = NULL,
2614                 .tree = tree,
2615                 .get_extent = get_extent,
2616                 .extent_locked = 0,
2617                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2618         };
2619
2620         ret = extent_write_cache_pages(tree, mapping, wbc,
2621                                        __extent_writepage, &epd,
2622                                        flush_write_bio);
2623         flush_epd_write_bio(&epd);
2624         return ret;
2625 }
2626
2627 int extent_readpages(struct extent_io_tree *tree,
2628                      struct address_space *mapping,
2629                      struct list_head *pages, unsigned nr_pages,
2630                      get_extent_t get_extent)
2631 {
2632         struct bio *bio = NULL;
2633         unsigned page_idx;
2634         unsigned long bio_flags = 0;
2635
2636         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2637                 struct page *page = list_entry(pages->prev, struct page, lru);
2638
2639                 prefetchw(&page->flags);
2640                 list_del(&page->lru);
2641                 if (!add_to_page_cache_lru(page, mapping,
2642                                         page->index, GFP_KERNEL)) {
2643                         __extent_read_full_page(tree, page, get_extent,
2644                                                 &bio, 0, &bio_flags);
2645                 }
2646                 page_cache_release(page);
2647         }
2648         BUG_ON(!list_empty(pages));
2649         if (bio)
2650                 submit_one_bio(READ, bio, 0, bio_flags);
2651         return 0;
2652 }
2653
2654 /*
2655  * basic invalidatepage code, this waits on any locked or writeback
2656  * ranges corresponding to the page, and then deletes any extent state
2657  * records from the tree
2658  */
2659 int extent_invalidatepage(struct extent_io_tree *tree,
2660                           struct page *page, unsigned long offset)
2661 {
2662         struct extent_state *cached_state = NULL;
2663         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2664         u64 end = start + PAGE_CACHE_SIZE - 1;
2665         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2666
2667         start += (offset + blocksize - 1) & ~(blocksize - 1);
2668         if (start > end)
2669                 return 0;
2670
2671         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
2672         wait_on_page_writeback(page);
2673         clear_extent_bit(tree, start, end,
2674                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2675                          EXTENT_DO_ACCOUNTING,
2676                          1, 1, &cached_state, GFP_NOFS);
2677         return 0;
2678 }
2679
2680 /*
2681  * simple commit_write call, set_range_dirty is used to mark both
2682  * the pages and the extent records as dirty
2683  */
2684 int extent_commit_write(struct extent_io_tree *tree,
2685                         struct inode *inode, struct page *page,
2686                         unsigned from, unsigned to)
2687 {
2688         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2689
2690         set_page_extent_mapped(page);
2691         set_page_dirty(page);
2692
2693         if (pos > inode->i_size) {
2694                 i_size_write(inode, pos);
2695                 mark_inode_dirty(inode);
2696         }
2697         return 0;
2698 }
2699
2700 int extent_prepare_write(struct extent_io_tree *tree,
2701                          struct inode *inode, struct page *page,
2702                          unsigned from, unsigned to, get_extent_t *get_extent)
2703 {
2704         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2705         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2706         u64 block_start;
2707         u64 orig_block_start;
2708         u64 block_end;
2709         u64 cur_end;
2710         struct extent_map *em;
2711         unsigned blocksize = 1 << inode->i_blkbits;
2712         size_t page_offset = 0;
2713         size_t block_off_start;
2714         size_t block_off_end;
2715         int err = 0;
2716         int iocount = 0;
2717         int ret = 0;
2718         int isnew;
2719
2720         set_page_extent_mapped(page);
2721
2722         block_start = (page_start + from) & ~((u64)blocksize - 1);
2723         block_end = (page_start + to - 1) | (blocksize - 1);
2724         orig_block_start = block_start;
2725
2726         lock_extent(tree, page_start, page_end, GFP_NOFS);
2727         while (block_start <= block_end) {
2728                 em = get_extent(inode, page, page_offset, block_start,
2729                                 block_end - block_start + 1, 1);
2730                 if (IS_ERR(em) || !em)
2731                         goto err;
2732
2733                 cur_end = min(block_end, extent_map_end(em) - 1);
2734                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2735                 block_off_end = block_off_start + blocksize;
2736                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2737
2738                 if (!PageUptodate(page) && isnew &&
2739                     (block_off_end > to || block_off_start < from)) {
2740                         void *kaddr;
2741
2742                         kaddr = kmap_atomic(page, KM_USER0);
2743                         if (block_off_end > to)
2744                                 memset(kaddr + to, 0, block_off_end - to);
2745                         if (block_off_start < from)
2746                                 memset(kaddr + block_off_start, 0,
2747                                        from - block_off_start);
2748                         flush_dcache_page(page);
2749                         kunmap_atomic(kaddr, KM_USER0);
2750                 }
2751                 if ((em->block_start != EXTENT_MAP_HOLE &&
2752                      em->block_start != EXTENT_MAP_INLINE) &&
2753                     !isnew && !PageUptodate(page) &&
2754                     (block_off_end > to || block_off_start < from) &&
2755                     !test_range_bit(tree, block_start, cur_end,
2756                                     EXTENT_UPTODATE, 1, NULL)) {
2757                         u64 sector;
2758                         u64 extent_offset = block_start - em->start;
2759                         size_t iosize;
2760                         sector = (em->block_start + extent_offset) >> 9;
2761                         iosize = (cur_end - block_start + blocksize) &
2762                                 ~((u64)blocksize - 1);
2763                         /*
2764                          * we've already got the extent locked, but we
2765                          * need to split the state such that our end_bio
2766                          * handler can clear the lock.
2767                          */
2768                         set_extent_bit(tree, block_start,
2769                                        block_start + iosize - 1,
2770                                        EXTENT_LOCKED, 0, NULL, NULL, GFP_NOFS);
2771                         ret = submit_extent_page(READ, tree, page,
2772                                          sector, iosize, page_offset, em->bdev,
2773                                          NULL, 1,
2774                                          end_bio_extent_preparewrite, 0,
2775                                          0, 0);
2776                         if (ret && !err)
2777                                 err = ret;
2778                         iocount++;
2779                         block_start = block_start + iosize;
2780                 } else {
2781                         set_extent_uptodate(tree, block_start, cur_end,
2782                                             GFP_NOFS);
2783                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2784                         block_start = cur_end + 1;
2785                 }
2786                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2787                 free_extent_map(em);
2788         }
2789         if (iocount) {
2790                 wait_extent_bit(tree, orig_block_start,
2791                                 block_end, EXTENT_LOCKED);
2792         }
2793         check_page_uptodate(tree, page);
2794 err:
2795         /* FIXME, zero out newly allocated blocks on error */
2796         return err;
2797 }
2798
2799 /*
2800  * a helper for releasepage, this tests for areas of the page that
2801  * are locked or under IO and drops the related state bits if it is safe
2802  * to drop the page.
2803  */
2804 int try_release_extent_state(struct extent_map_tree *map,
2805                              struct extent_io_tree *tree, struct page *page,
2806                              gfp_t mask)
2807 {
2808         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2809         u64 end = start + PAGE_CACHE_SIZE - 1;
2810         int ret = 1;
2811
2812         if (test_range_bit(tree, start, end,
2813                            EXTENT_IOBITS, 0, NULL))
2814                 ret = 0;
2815         else {
2816                 if ((mask & GFP_NOFS) == GFP_NOFS)
2817                         mask = GFP_NOFS;
2818                 /*
2819                  * at this point we can safely clear everything except the
2820                  * locked bit and the nodatasum bit
2821                  */
2822                 clear_extent_bit(tree, start, end,
2823                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2824                                  0, 0, NULL, mask);
2825         }
2826         return ret;
2827 }
2828
2829 /*
2830  * a helper for releasepage.  As long as there are no locked extents
2831  * in the range corresponding to the page, both state records and extent
2832  * map records are removed
2833  */
2834 int try_release_extent_mapping(struct extent_map_tree *map,
2835                                struct extent_io_tree *tree, struct page *page,
2836                                gfp_t mask)
2837 {
2838         struct extent_map *em;
2839         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2840         u64 end = start + PAGE_CACHE_SIZE - 1;
2841
2842         if ((mask & __GFP_WAIT) &&
2843             page->mapping->host->i_size > 16 * 1024 * 1024) {
2844                 u64 len;
2845                 while (start <= end) {
2846                         len = end - start + 1;
2847                         write_lock(&map->lock);
2848                         em = lookup_extent_mapping(map, start, len);
2849                         if (!em || IS_ERR(em)) {
2850                                 write_unlock(&map->lock);
2851                                 break;
2852                         }
2853                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2854                             em->start != start) {
2855                                 write_unlock(&map->lock);
2856                                 free_extent_map(em);
2857                                 break;
2858                         }
2859                         if (!test_range_bit(tree, em->start,
2860                                             extent_map_end(em) - 1,
2861                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
2862                                             0, NULL)) {
2863                                 remove_extent_mapping(map, em);
2864                                 /* once for the rb tree */
2865                                 free_extent_map(em);
2866                         }
2867                         start = extent_map_end(em);
2868                         write_unlock(&map->lock);
2869
2870                         /* once for us */
2871                         free_extent_map(em);
2872                 }
2873         }
2874         return try_release_extent_state(map, tree, page, mask);
2875 }
2876
2877 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2878                 get_extent_t *get_extent)
2879 {
2880         struct inode *inode = mapping->host;
2881         struct extent_state *cached_state = NULL;
2882         u64 start = iblock << inode->i_blkbits;
2883         sector_t sector = 0;
2884         size_t blksize = (1 << inode->i_blkbits);
2885         struct extent_map *em;
2886
2887         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2888                          0, &cached_state, GFP_NOFS);
2889         em = get_extent(inode, NULL, 0, start, blksize, 0);
2890         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start,
2891                              start + blksize - 1, &cached_state, GFP_NOFS);
2892         if (!em || IS_ERR(em))
2893                 return 0;
2894
2895         if (em->block_start > EXTENT_MAP_LAST_BYTE)
2896                 goto out;
2897
2898         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2899 out:
2900         free_extent_map(em);
2901         return sector;
2902 }
2903
2904 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2905                 __u64 start, __u64 len, get_extent_t *get_extent)
2906 {
2907         int ret = 0;
2908         u64 off = start;
2909         u64 max = start + len;
2910         u32 flags = 0;
2911         u32 found_type;
2912         u64 last;
2913         u64 disko = 0;
2914         struct btrfs_key found_key;
2915         struct extent_map *em = NULL;
2916         struct extent_state *cached_state = NULL;
2917         struct btrfs_path *path;
2918         struct btrfs_file_extent_item *item;
2919         int end = 0;
2920         u64 em_start = 0, em_len = 0;
2921         unsigned long emflags;
2922         int hole = 0;
2923
2924         if (len == 0)
2925                 return -EINVAL;
2926
2927         path = btrfs_alloc_path();
2928         if (!path)
2929                 return -ENOMEM;
2930         path->leave_spinning = 1;
2931
2932         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
2933                                        path, inode->i_ino, -1, 0);
2934         if (ret < 0) {
2935                 btrfs_free_path(path);
2936                 return ret;
2937         }
2938         WARN_ON(!ret);
2939         path->slots[0]--;
2940         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2941                               struct btrfs_file_extent_item);
2942         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
2943         found_type = btrfs_key_type(&found_key);
2944
2945         /* No extents, just return */
2946         if (found_key.objectid != inode->i_ino ||
2947             found_type != BTRFS_EXTENT_DATA_KEY) {
2948                 btrfs_free_path(path);
2949                 return 0;
2950         }
2951         last = found_key.offset;
2952         btrfs_free_path(path);
2953
2954         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
2955                          &cached_state, GFP_NOFS);
2956         em = get_extent(inode, NULL, 0, off, max - off, 0);
2957         if (!em)
2958                 goto out;
2959         if (IS_ERR(em)) {
2960                 ret = PTR_ERR(em);
2961                 goto out;
2962         }
2963
2964         while (!end) {
2965                 hole = 0;
2966                 off = em->start + em->len;
2967                 if (off >= max)
2968                         end = 1;
2969
2970                 if (em->block_start == EXTENT_MAP_HOLE) {
2971                         hole = 1;
2972                         goto next;
2973                 }
2974
2975                 em_start = em->start;
2976                 em_len = em->len;
2977
2978                 disko = 0;
2979                 flags = 0;
2980
2981                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
2982                         end = 1;
2983                         flags |= FIEMAP_EXTENT_LAST;
2984                 } else if (em->block_start == EXTENT_MAP_INLINE) {
2985                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
2986                                   FIEMAP_EXTENT_NOT_ALIGNED);
2987                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
2988                         flags |= (FIEMAP_EXTENT_DELALLOC |
2989                                   FIEMAP_EXTENT_UNKNOWN);
2990                 } else {
2991                         disko = em->block_start;
2992                 }
2993                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2994                         flags |= FIEMAP_EXTENT_ENCODED;
2995
2996 next:
2997                 emflags = em->flags;
2998                 free_extent_map(em);
2999                 em = NULL;
3000                 if (!end) {
3001                         em = get_extent(inode, NULL, 0, off, max - off, 0);
3002                         if (!em)
3003                                 goto out;
3004                         if (IS_ERR(em)) {
3005                                 ret = PTR_ERR(em);
3006                                 goto out;
3007                         }
3008                         emflags = em->flags;
3009                 }
3010
3011                 if (test_bit(EXTENT_FLAG_VACANCY, &emflags)) {
3012                         flags |= FIEMAP_EXTENT_LAST;
3013                         end = 1;
3014                 }
3015
3016                 if (em_start == last) {
3017                         flags |= FIEMAP_EXTENT_LAST;
3018                         end = 1;
3019                 }
3020
3021                 if (!hole) {
3022                         ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3023                                                 em_len, flags);
3024                         if (ret)
3025                                 goto out_free;
3026                 }
3027         }
3028 out_free:
3029         free_extent_map(em);
3030 out:
3031         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3032                              &cached_state, GFP_NOFS);
3033         return ret;
3034 }
3035
3036 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
3037                                               unsigned long i)
3038 {
3039         struct page *p;
3040         struct address_space *mapping;
3041
3042         if (i == 0)
3043                 return eb->first_page;
3044         i += eb->start >> PAGE_CACHE_SHIFT;
3045         mapping = eb->first_page->mapping;
3046         if (!mapping)
3047                 return NULL;
3048
3049         /*
3050          * extent_buffer_page is only called after pinning the page
3051          * by increasing the reference count.  So we know the page must
3052          * be in the radix tree.
3053          */
3054         rcu_read_lock();
3055         p = radix_tree_lookup(&mapping->page_tree, i);
3056         rcu_read_unlock();
3057
3058         return p;
3059 }
3060
3061 static inline unsigned long num_extent_pages(u64 start, u64 len)
3062 {
3063         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3064                 (start >> PAGE_CACHE_SHIFT);
3065 }
3066
3067 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3068                                                    u64 start,
3069                                                    unsigned long len,
3070                                                    gfp_t mask)
3071 {
3072         struct extent_buffer *eb = NULL;
3073 #if LEAK_DEBUG
3074         unsigned long flags;
3075 #endif
3076
3077         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3078         if (eb == NULL)
3079                 return NULL;
3080         eb->start = start;
3081         eb->len = len;
3082         spin_lock_init(&eb->lock);
3083         init_waitqueue_head(&eb->lock_wq);
3084
3085 #if LEAK_DEBUG
3086         spin_lock_irqsave(&leak_lock, flags);
3087         list_add(&eb->leak_list, &buffers);
3088         spin_unlock_irqrestore(&leak_lock, flags);
3089 #endif
3090         atomic_set(&eb->refs, 1);
3091
3092         return eb;
3093 }
3094
3095 static void __free_extent_buffer(struct extent_buffer *eb)
3096 {
3097 #if LEAK_DEBUG
3098         unsigned long flags;
3099         spin_lock_irqsave(&leak_lock, flags);
3100         list_del(&eb->leak_list);
3101         spin_unlock_irqrestore(&leak_lock, flags);
3102 #endif
3103         kmem_cache_free(extent_buffer_cache, eb);
3104 }
3105
3106 /*
3107  * Helper for releasing extent buffer page.
3108  */
3109 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3110                                                 unsigned long start_idx)
3111 {
3112         unsigned long index;
3113         struct page *page;
3114
3115         if (!eb->first_page)
3116                 return;
3117
3118         index = num_extent_pages(eb->start, eb->len);
3119         if (start_idx >= index)
3120                 return;
3121
3122         do {
3123                 index--;
3124                 page = extent_buffer_page(eb, index);
3125                 if (page)
3126                         page_cache_release(page);
3127         } while (index != start_idx);
3128 }
3129
3130 /*
3131  * Helper for releasing the extent buffer.
3132  */
3133 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3134 {
3135         btrfs_release_extent_buffer_page(eb, 0);
3136         __free_extent_buffer(eb);
3137 }
3138
3139 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3140                                           u64 start, unsigned long len,
3141                                           struct page *page0,
3142                                           gfp_t mask)
3143 {
3144         unsigned long num_pages = num_extent_pages(start, len);
3145         unsigned long i;
3146         unsigned long index = start >> PAGE_CACHE_SHIFT;
3147         struct extent_buffer *eb;
3148         struct extent_buffer *exists = NULL;
3149         struct page *p;
3150         struct address_space *mapping = tree->mapping;
3151         int uptodate = 1;
3152         int ret;
3153
3154         rcu_read_lock();
3155         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3156         if (eb && atomic_inc_not_zero(&eb->refs)) {
3157                 rcu_read_unlock();
3158                 mark_page_accessed(eb->first_page);
3159                 return eb;
3160         }
3161         rcu_read_unlock();
3162
3163         eb = __alloc_extent_buffer(tree, start, len, mask);
3164         if (!eb)
3165                 return NULL;
3166
3167         if (page0) {
3168                 eb->first_page = page0;
3169                 i = 1;
3170                 index++;
3171                 page_cache_get(page0);
3172                 mark_page_accessed(page0);
3173                 set_page_extent_mapped(page0);
3174                 set_page_extent_head(page0, len);
3175                 uptodate = PageUptodate(page0);
3176         } else {
3177                 i = 0;
3178         }
3179         for (; i < num_pages; i++, index++) {
3180                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
3181                 if (!p) {
3182                         WARN_ON(1);
3183                         goto free_eb;
3184                 }
3185                 set_page_extent_mapped(p);
3186                 mark_page_accessed(p);
3187                 if (i == 0) {
3188                         eb->first_page = p;
3189                         set_page_extent_head(p, len);
3190                 } else {
3191                         set_page_private(p, EXTENT_PAGE_PRIVATE);
3192                 }
3193                 if (!PageUptodate(p))
3194                         uptodate = 0;
3195                 unlock_page(p);
3196         }
3197         if (uptodate)
3198                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3199
3200         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
3201         if (ret)
3202                 goto free_eb;
3203
3204         spin_lock(&tree->buffer_lock);
3205         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
3206         if (ret == -EEXIST) {
3207                 exists = radix_tree_lookup(&tree->buffer,
3208                                                 start >> PAGE_CACHE_SHIFT);
3209                 /* add one reference for the caller */
3210                 atomic_inc(&exists->refs);
3211                 spin_unlock(&tree->buffer_lock);
3212                 radix_tree_preload_end();
3213                 goto free_eb;
3214         }
3215         /* add one reference for the tree */
3216         atomic_inc(&eb->refs);
3217         spin_unlock(&tree->buffer_lock);
3218         radix_tree_preload_end();
3219         return eb;
3220
3221 free_eb:
3222         if (!atomic_dec_and_test(&eb->refs))
3223                 return exists;
3224         btrfs_release_extent_buffer(eb);
3225         return exists;
3226 }
3227
3228 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3229                                          u64 start, unsigned long len,
3230                                           gfp_t mask)
3231 {
3232         struct extent_buffer *eb;
3233
3234         rcu_read_lock();
3235         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3236         if (eb && atomic_inc_not_zero(&eb->refs)) {
3237                 rcu_read_unlock();
3238                 mark_page_accessed(eb->first_page);
3239                 return eb;
3240         }
3241         rcu_read_unlock();
3242
3243         return NULL;
3244 }
3245
3246 void free_extent_buffer(struct extent_buffer *eb)
3247 {
3248         if (!eb)
3249                 return;
3250
3251         if (!atomic_dec_and_test(&eb->refs))
3252                 return;
3253
3254         WARN_ON(1);
3255 }
3256
3257 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3258                               struct extent_buffer *eb)
3259 {
3260         unsigned long i;
3261         unsigned long num_pages;
3262         struct page *page;
3263
3264         num_pages = num_extent_pages(eb->start, eb->len);
3265
3266         for (i = 0; i < num_pages; i++) {
3267                 page = extent_buffer_page(eb, i);
3268                 if (!PageDirty(page))
3269                         continue;
3270
3271                 lock_page(page);
3272                 if (i == 0)
3273                         set_page_extent_head(page, eb->len);
3274                 else
3275                         set_page_private(page, EXTENT_PAGE_PRIVATE);
3276
3277                 clear_page_dirty_for_io(page);
3278                 spin_lock_irq(&page->mapping->tree_lock);
3279                 if (!PageDirty(page)) {
3280                         radix_tree_tag_clear(&page->mapping->page_tree,
3281                                                 page_index(page),
3282                                                 PAGECACHE_TAG_DIRTY);
3283                 }
3284                 spin_unlock_irq(&page->mapping->tree_lock);
3285                 unlock_page(page);
3286         }
3287         return 0;
3288 }
3289
3290 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
3291                                     struct extent_buffer *eb)
3292 {
3293         return wait_on_extent_writeback(tree, eb->start,
3294                                         eb->start + eb->len - 1);
3295 }
3296
3297 int set_extent_buffer_dirty(struct extent_io_tree *tree,
3298                              struct extent_buffer *eb)
3299 {
3300         unsigned long i;
3301         unsigned long num_pages;
3302         int was_dirty = 0;
3303
3304         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3305         num_pages = num_extent_pages(eb->start, eb->len);
3306         for (i = 0; i < num_pages; i++)
3307                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3308         return was_dirty;
3309 }
3310
3311 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3312                                 struct extent_buffer *eb,
3313                                 struct extent_state **cached_state)
3314 {
3315         unsigned long i;
3316         struct page *page;
3317         unsigned long num_pages;
3318
3319         num_pages = num_extent_pages(eb->start, eb->len);
3320         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3321
3322         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3323                               cached_state, GFP_NOFS);
3324         for (i = 0; i < num_pages; i++) {
3325                 page = extent_buffer_page(eb, i);
3326                 if (page)
3327                         ClearPageUptodate(page);
3328         }
3329         return 0;
3330 }
3331
3332 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3333                                 struct extent_buffer *eb)
3334 {
3335         unsigned long i;
3336         struct page *page;
3337         unsigned long num_pages;
3338
3339         num_pages = num_extent_pages(eb->start, eb->len);
3340
3341         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3342                             GFP_NOFS);
3343         for (i = 0; i < num_pages; i++) {
3344                 page = extent_buffer_page(eb, i);
3345                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3346                     ((i == num_pages - 1) &&
3347                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3348                         check_page_uptodate(tree, page);
3349                         continue;
3350                 }
3351                 SetPageUptodate(page);
3352         }
3353         return 0;
3354 }
3355
3356 int extent_range_uptodate(struct extent_io_tree *tree,
3357                           u64 start, u64 end)
3358 {
3359         struct page *page;
3360         int ret;
3361         int pg_uptodate = 1;
3362         int uptodate;
3363         unsigned long index;
3364
3365         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL);
3366         if (ret)
3367                 return 1;
3368         while (start <= end) {
3369                 index = start >> PAGE_CACHE_SHIFT;
3370                 page = find_get_page(tree->mapping, index);
3371                 uptodate = PageUptodate(page);
3372                 page_cache_release(page);
3373                 if (!uptodate) {
3374                         pg_uptodate = 0;
3375                         break;
3376                 }
3377                 start += PAGE_CACHE_SIZE;
3378         }
3379         return pg_uptodate;
3380 }
3381
3382 int extent_buffer_uptodate(struct extent_io_tree *tree,
3383                            struct extent_buffer *eb,
3384                            struct extent_state *cached_state)
3385 {
3386         int ret = 0;
3387         unsigned long num_pages;
3388         unsigned long i;
3389         struct page *page;
3390         int pg_uptodate = 1;
3391
3392         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3393                 return 1;
3394
3395         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3396                            EXTENT_UPTODATE, 1, cached_state);
3397         if (ret)
3398                 return ret;
3399
3400         num_pages = num_extent_pages(eb->start, eb->len);
3401         for (i = 0; i < num_pages; i++) {
3402                 page = extent_buffer_page(eb, i);
3403                 if (!PageUptodate(page)) {
3404                         pg_uptodate = 0;
3405                         break;
3406                 }
3407         }
3408         return pg_uptodate;
3409 }
3410
3411 int read_extent_buffer_pages(struct extent_io_tree *tree,
3412                              struct extent_buffer *eb,
3413                              u64 start, int wait,
3414                              get_extent_t *get_extent, int mirror_num)
3415 {
3416         unsigned long i;
3417         unsigned long start_i;
3418         struct page *page;
3419         int err;
3420         int ret = 0;
3421         int locked_pages = 0;
3422         int all_uptodate = 1;
3423         int inc_all_pages = 0;
3424         unsigned long num_pages;
3425         struct bio *bio = NULL;
3426         unsigned long bio_flags = 0;
3427
3428         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3429                 return 0;
3430
3431         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3432                            EXTENT_UPTODATE, 1, NULL)) {
3433                 return 0;
3434         }
3435
3436         if (start) {
3437                 WARN_ON(start < eb->start);
3438                 start_i = (start >> PAGE_CACHE_SHIFT) -
3439                         (eb->start >> PAGE_CACHE_SHIFT);
3440         } else {
3441                 start_i = 0;
3442         }
3443
3444         num_pages = num_extent_pages(eb->start, eb->len);
3445         for (i = start_i; i < num_pages; i++) {
3446                 page = extent_buffer_page(eb, i);
3447                 if (!wait) {
3448                         if (!trylock_page(page))
3449                                 goto unlock_exit;
3450                 } else {
3451                         lock_page(page);
3452                 }
3453                 locked_pages++;
3454                 if (!PageUptodate(page))
3455                         all_uptodate = 0;
3456         }
3457         if (all_uptodate) {
3458                 if (start_i == 0)
3459                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3460                 goto unlock_exit;
3461         }
3462
3463         for (i = start_i; i < num_pages; i++) {
3464                 page = extent_buffer_page(eb, i);
3465                 if (inc_all_pages)
3466                         page_cache_get(page);
3467                 if (!PageUptodate(page)) {
3468                         if (start_i == 0)
3469                                 inc_all_pages = 1;
3470                         ClearPageError(page);
3471                         err = __extent_read_full_page(tree, page,
3472                                                       get_extent, &bio,
3473                                                       mirror_num, &bio_flags);
3474                         if (err)
3475                                 ret = err;
3476                 } else {
3477                         unlock_page(page);
3478                 }
3479         }
3480
3481         if (bio)
3482                 submit_one_bio(READ, bio, mirror_num, bio_flags);
3483
3484         if (ret || !wait)
3485                 return ret;
3486
3487         for (i = start_i; i < num_pages; i++) {
3488                 page = extent_buffer_page(eb, i);
3489                 wait_on_page_locked(page);
3490                 if (!PageUptodate(page))
3491                         ret = -EIO;
3492         }
3493
3494         if (!ret)
3495                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3496         return ret;
3497
3498 unlock_exit:
3499         i = start_i;
3500         while (locked_pages > 0) {
3501                 page = extent_buffer_page(eb, i);
3502                 i++;
3503                 unlock_page(page);
3504                 locked_pages--;
3505         }
3506         return ret;
3507 }
3508
3509 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3510                         unsigned long start,
3511                         unsigned long len)
3512 {
3513         size_t cur;
3514         size_t offset;
3515         struct page *page;
3516         char *kaddr;
3517         char *dst = (char *)dstv;
3518         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3519         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3520
3521         WARN_ON(start > eb->len);
3522         WARN_ON(start + len > eb->start + eb->len);
3523
3524         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3525
3526         while (len > 0) {
3527                 page = extent_buffer_page(eb, i);
3528
3529                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3530                 kaddr = kmap_atomic(page, KM_USER1);
3531                 memcpy(dst, kaddr + offset, cur);
3532                 kunmap_atomic(kaddr, KM_USER1);
3533
3534                 dst += cur;
3535                 len -= cur;
3536                 offset = 0;
3537                 i++;
3538         }
3539 }
3540
3541 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3542                                unsigned long min_len, char **token, char **map,
3543                                unsigned long *map_start,
3544                                unsigned long *map_len, int km)
3545 {
3546         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3547         char *kaddr;
3548         struct page *p;
3549         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3550         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3551         unsigned long end_i = (start_offset + start + min_len - 1) >>
3552                 PAGE_CACHE_SHIFT;
3553
3554         if (i != end_i)
3555                 return -EINVAL;
3556
3557         if (i == 0) {
3558                 offset = start_offset;
3559                 *map_start = 0;
3560         } else {
3561                 offset = 0;
3562                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3563         }
3564
3565         if (start + min_len > eb->len) {
3566                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3567                        "wanted %lu %lu\n", (unsigned long long)eb->start,
3568                        eb->len, start, min_len);
3569                 WARN_ON(1);
3570         }
3571
3572         p = extent_buffer_page(eb, i);
3573         kaddr = kmap_atomic(p, km);
3574         *token = kaddr;
3575         *map = kaddr + offset;
3576         *map_len = PAGE_CACHE_SIZE - offset;
3577         return 0;
3578 }
3579
3580 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3581                       unsigned long min_len,
3582                       char **token, char **map,
3583                       unsigned long *map_start,
3584                       unsigned long *map_len, int km)
3585 {
3586         int err;
3587         int save = 0;
3588         if (eb->map_token) {
3589                 unmap_extent_buffer(eb, eb->map_token, km);
3590                 eb->map_token = NULL;
3591                 save = 1;
3592         }
3593         err = map_private_extent_buffer(eb, start, min_len, token, map,
3594                                        map_start, map_len, km);
3595         if (!err && save) {
3596                 eb->map_token = *token;
3597                 eb->kaddr = *map;
3598                 eb->map_start = *map_start;
3599                 eb->map_len = *map_len;
3600         }
3601         return err;
3602 }
3603
3604 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3605 {
3606         kunmap_atomic(token, km);
3607 }
3608
3609 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3610                           unsigned long start,
3611                           unsigned long len)
3612 {
3613         size_t cur;
3614         size_t offset;
3615         struct page *page;
3616         char *kaddr;
3617         char *ptr = (char *)ptrv;
3618         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3619         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3620         int ret = 0;
3621
3622         WARN_ON(start > eb->len);
3623         WARN_ON(start + len > eb->start + eb->len);
3624
3625         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3626
3627         while (len > 0) {
3628                 page = extent_buffer_page(eb, i);
3629
3630                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3631
3632                 kaddr = kmap_atomic(page, KM_USER0);
3633                 ret = memcmp(ptr, kaddr + offset, cur);
3634                 kunmap_atomic(kaddr, KM_USER0);
3635                 if (ret)
3636                         break;
3637
3638                 ptr += cur;
3639                 len -= cur;
3640                 offset = 0;
3641                 i++;
3642         }
3643         return ret;
3644 }
3645
3646 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3647                          unsigned long start, unsigned long len)
3648 {
3649         size_t cur;
3650         size_t offset;
3651         struct page *page;
3652         char *kaddr;
3653         char *src = (char *)srcv;
3654         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3655         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3656
3657         WARN_ON(start > eb->len);
3658         WARN_ON(start + len > eb->start + eb->len);
3659
3660         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3661
3662         while (len > 0) {
3663                 page = extent_buffer_page(eb, i);
3664                 WARN_ON(!PageUptodate(page));
3665
3666                 cur = min(len, PAGE_CACHE_SIZE - offset);
3667                 kaddr = kmap_atomic(page, KM_USER1);
3668                 memcpy(kaddr + offset, src, cur);
3669                 kunmap_atomic(kaddr, KM_USER1);
3670
3671                 src += cur;
3672                 len -= cur;
3673                 offset = 0;
3674                 i++;
3675         }
3676 }
3677
3678 void memset_extent_buffer(struct extent_buffer *eb, char c,
3679                           unsigned long start, unsigned long len)
3680 {
3681         size_t cur;
3682         size_t offset;
3683         struct page *page;
3684         char *kaddr;
3685         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3686         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3687
3688         WARN_ON(start > eb->len);
3689         WARN_ON(start + len > eb->start + eb->len);
3690
3691         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3692
3693         while (len > 0) {
3694                 page = extent_buffer_page(eb, i);
3695                 WARN_ON(!PageUptodate(page));
3696
3697                 cur = min(len, PAGE_CACHE_SIZE - offset);
3698                 kaddr = kmap_atomic(page, KM_USER0);
3699                 memset(kaddr + offset, c, cur);
3700                 kunmap_atomic(kaddr, KM_USER0);
3701
3702                 len -= cur;
3703                 offset = 0;
3704                 i++;
3705         }
3706 }
3707
3708 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3709                         unsigned long dst_offset, unsigned long src_offset,
3710                         unsigned long len)
3711 {
3712         u64 dst_len = dst->len;
3713         size_t cur;
3714         size_t offset;
3715         struct page *page;
3716         char *kaddr;
3717         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3718         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3719
3720         WARN_ON(src->len != dst_len);
3721
3722         offset = (start_offset + dst_offset) &
3723                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3724
3725         while (len > 0) {
3726                 page = extent_buffer_page(dst, i);
3727                 WARN_ON(!PageUptodate(page));
3728
3729                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3730
3731                 kaddr = kmap_atomic(page, KM_USER0);
3732                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3733                 kunmap_atomic(kaddr, KM_USER0);
3734
3735                 src_offset += cur;
3736                 len -= cur;
3737                 offset = 0;
3738                 i++;
3739         }
3740 }
3741
3742 static void move_pages(struct page *dst_page, struct page *src_page,
3743                        unsigned long dst_off, unsigned long src_off,
3744                        unsigned long len)
3745 {
3746         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3747         if (dst_page == src_page) {
3748                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3749         } else {
3750                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3751                 char *p = dst_kaddr + dst_off + len;
3752                 char *s = src_kaddr + src_off + len;
3753
3754                 while (len--)
3755                         *--p = *--s;
3756
3757                 kunmap_atomic(src_kaddr, KM_USER1);
3758         }
3759         kunmap_atomic(dst_kaddr, KM_USER0);
3760 }
3761
3762 static void copy_pages(struct page *dst_page, struct page *src_page,
3763                        unsigned long dst_off, unsigned long src_off,
3764                        unsigned long len)
3765 {
3766         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3767         char *src_kaddr;
3768
3769         if (dst_page != src_page)
3770                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3771         else
3772                 src_kaddr = dst_kaddr;
3773
3774         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3775         kunmap_atomic(dst_kaddr, KM_USER0);
3776         if (dst_page != src_page)
3777                 kunmap_atomic(src_kaddr, KM_USER1);
3778 }
3779
3780 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3781                            unsigned long src_offset, unsigned long len)
3782 {
3783         size_t cur;
3784         size_t dst_off_in_page;
3785         size_t src_off_in_page;
3786         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3787         unsigned long dst_i;
3788         unsigned long src_i;
3789
3790         if (src_offset + len > dst->len) {
3791                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3792                        "len %lu dst len %lu\n", src_offset, len, dst->len);
3793                 BUG_ON(1);
3794         }
3795         if (dst_offset + len > dst->len) {
3796                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3797                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
3798                 BUG_ON(1);
3799         }
3800
3801         while (len > 0) {
3802                 dst_off_in_page = (start_offset + dst_offset) &
3803                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3804                 src_off_in_page = (start_offset + src_offset) &
3805                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3806
3807                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3808                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3809
3810                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3811                                                src_off_in_page));
3812                 cur = min_t(unsigned long, cur,
3813                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3814
3815                 copy_pages(extent_buffer_page(dst, dst_i),
3816                            extent_buffer_page(dst, src_i),
3817                            dst_off_in_page, src_off_in_page, cur);
3818
3819                 src_offset += cur;
3820                 dst_offset += cur;
3821                 len -= cur;
3822         }
3823 }
3824
3825 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3826                            unsigned long src_offset, unsigned long len)
3827 {
3828         size_t cur;
3829         size_t dst_off_in_page;
3830         size_t src_off_in_page;
3831         unsigned long dst_end = dst_offset + len - 1;
3832         unsigned long src_end = src_offset + len - 1;
3833         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3834         unsigned long dst_i;
3835         unsigned long src_i;
3836
3837         if (src_offset + len > dst->len) {
3838                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3839                        "len %lu len %lu\n", src_offset, len, dst->len);
3840                 BUG_ON(1);
3841         }
3842         if (dst_offset + len > dst->len) {
3843                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3844                        "len %lu len %lu\n", dst_offset, len, dst->len);
3845                 BUG_ON(1);
3846         }
3847         if (dst_offset < src_offset) {
3848                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3849                 return;
3850         }
3851         while (len > 0) {
3852                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3853                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3854
3855                 dst_off_in_page = (start_offset + dst_end) &
3856                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3857                 src_off_in_page = (start_offset + src_end) &
3858                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3859
3860                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3861                 cur = min(cur, dst_off_in_page + 1);
3862                 move_pages(extent_buffer_page(dst, dst_i),
3863                            extent_buffer_page(dst, src_i),
3864                            dst_off_in_page - cur + 1,
3865                            src_off_in_page - cur + 1, cur);
3866
3867                 dst_end -= cur;
3868                 src_end -= cur;
3869                 len -= cur;
3870         }
3871 }
3872
3873 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3874 {
3875         struct extent_buffer *eb =
3876                         container_of(head, struct extent_buffer, rcu_head);
3877
3878         btrfs_release_extent_buffer(eb);
3879 }
3880
3881 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3882 {
3883         u64 start = page_offset(page);
3884         struct extent_buffer *eb;
3885         int ret = 1;
3886
3887         spin_lock(&tree->buffer_lock);
3888         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3889         if (!eb) {
3890                 spin_unlock(&tree->buffer_lock);
3891                 return ret;
3892         }
3893
3894         if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3895                 ret = 0;
3896                 goto out;
3897         }
3898
3899         /*
3900          * set @eb->refs to 0 if it is already 1, and then release the @eb.
3901          * Or go back.
3902          */
3903         if (atomic_cmpxchg(&eb->refs, 1, 0) != 1) {
3904                 ret = 0;
3905                 goto out;
3906         }
3907
3908         radix_tree_delete(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3909 out:
3910         spin_unlock(&tree->buffer_lock);
3911
3912         /* at this point we can safely release the extent buffer */
3913         if (atomic_read(&eb->refs) == 0)
3914                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3915         return ret;
3916 }