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                 ret = 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         if (!bio)
1924                 return -ENOMEM;
1925
1926         bio_add_page(bio, page, page_size, offset);
1927         bio->bi_end_io = end_io_func;
1928         bio->bi_private = tree;
1929
1930         if (bio_ret)
1931                 *bio_ret = bio;
1932         else
1933                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1934
1935         return ret;
1936 }
1937
1938 void set_page_extent_mapped(struct page *page)
1939 {
1940         if (!PagePrivate(page)) {
1941                 SetPagePrivate(page);
1942                 page_cache_get(page);
1943                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1944         }
1945 }
1946
1947 static void set_page_extent_head(struct page *page, unsigned long len)
1948 {
1949         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1950 }
1951
1952 /*
1953  * basic readpage implementation.  Locked extent state structs are inserted
1954  * into the tree that are removed when the IO is done (by the end_io
1955  * handlers)
1956  */
1957 static int __extent_read_full_page(struct extent_io_tree *tree,
1958                                    struct page *page,
1959                                    get_extent_t *get_extent,
1960                                    struct bio **bio, int mirror_num,
1961                                    unsigned long *bio_flags)
1962 {
1963         struct inode *inode = page->mapping->host;
1964         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1965         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1966         u64 end;
1967         u64 cur = start;
1968         u64 extent_offset;
1969         u64 last_byte = i_size_read(inode);
1970         u64 block_start;
1971         u64 cur_end;
1972         sector_t sector;
1973         struct extent_map *em;
1974         struct block_device *bdev;
1975         struct btrfs_ordered_extent *ordered;
1976         int ret;
1977         int nr = 0;
1978         size_t page_offset = 0;
1979         size_t iosize;
1980         size_t disk_io_size;
1981         size_t blocksize = inode->i_sb->s_blocksize;
1982         unsigned long this_bio_flag = 0;
1983
1984         set_page_extent_mapped(page);
1985
1986         end = page_end;
1987         while (1) {
1988                 lock_extent(tree, start, end, GFP_NOFS);
1989                 ordered = btrfs_lookup_ordered_extent(inode, start);
1990                 if (!ordered)
1991                         break;
1992                 unlock_extent(tree, start, end, GFP_NOFS);
1993                 btrfs_start_ordered_extent(inode, ordered, 1);
1994                 btrfs_put_ordered_extent(ordered);
1995         }
1996
1997         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
1998                 char *userpage;
1999                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2000
2001                 if (zero_offset) {
2002                         iosize = PAGE_CACHE_SIZE - zero_offset;
2003                         userpage = kmap_atomic(page, KM_USER0);
2004                         memset(userpage + zero_offset, 0, iosize);
2005                         flush_dcache_page(page);
2006                         kunmap_atomic(userpage, KM_USER0);
2007                 }
2008         }
2009         while (cur <= end) {
2010                 if (cur >= last_byte) {
2011                         char *userpage;
2012                         iosize = PAGE_CACHE_SIZE - page_offset;
2013                         userpage = kmap_atomic(page, KM_USER0);
2014                         memset(userpage + page_offset, 0, iosize);
2015                         flush_dcache_page(page);
2016                         kunmap_atomic(userpage, KM_USER0);
2017                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2018                                             GFP_NOFS);
2019                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2020                         break;
2021                 }
2022                 em = get_extent(inode, page, page_offset, cur,
2023                                 end - cur + 1, 0);
2024                 if (IS_ERR(em) || !em) {
2025                         SetPageError(page);
2026                         unlock_extent(tree, cur, end, GFP_NOFS);
2027                         break;
2028                 }
2029                 extent_offset = cur - em->start;
2030                 BUG_ON(extent_map_end(em) <= cur);
2031                 BUG_ON(end < cur);
2032
2033                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2034                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2035                         extent_set_compress_type(&this_bio_flag,
2036                                                  em->compress_type);
2037                 }
2038
2039                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2040                 cur_end = min(extent_map_end(em) - 1, end);
2041                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2042                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2043                         disk_io_size = em->block_len;
2044                         sector = em->block_start >> 9;
2045                 } else {
2046                         sector = (em->block_start + extent_offset) >> 9;
2047                         disk_io_size = iosize;
2048                 }
2049                 bdev = em->bdev;
2050                 block_start = em->block_start;
2051                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2052                         block_start = EXTENT_MAP_HOLE;
2053                 free_extent_map(em);
2054                 em = NULL;
2055
2056                 /* we've found a hole, just zero and go on */
2057                 if (block_start == EXTENT_MAP_HOLE) {
2058                         char *userpage;
2059                         userpage = kmap_atomic(page, KM_USER0);
2060                         memset(userpage + page_offset, 0, iosize);
2061                         flush_dcache_page(page);
2062                         kunmap_atomic(userpage, KM_USER0);
2063
2064                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2065                                             GFP_NOFS);
2066                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2067                         cur = cur + iosize;
2068                         page_offset += iosize;
2069                         continue;
2070                 }
2071                 /* the get_extent function already copied into the page */
2072                 if (test_range_bit(tree, cur, cur_end,
2073                                    EXTENT_UPTODATE, 1, NULL)) {
2074                         check_page_uptodate(tree, page);
2075                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2076                         cur = cur + iosize;
2077                         page_offset += iosize;
2078                         continue;
2079                 }
2080                 /* we have an inline extent but it didn't get marked up
2081                  * to date.  Error out
2082                  */
2083                 if (block_start == EXTENT_MAP_INLINE) {
2084                         SetPageError(page);
2085                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2086                         cur = cur + iosize;
2087                         page_offset += iosize;
2088                         continue;
2089                 }
2090
2091                 ret = 0;
2092                 if (tree->ops && tree->ops->readpage_io_hook) {
2093                         ret = tree->ops->readpage_io_hook(page, cur,
2094                                                           cur + iosize - 1);
2095                 }
2096                 if (!ret) {
2097                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2098                         pnr -= page->index;
2099                         ret = submit_extent_page(READ, tree, page,
2100                                          sector, disk_io_size, page_offset,
2101                                          bdev, bio, pnr,
2102                                          end_bio_extent_readpage, mirror_num,
2103                                          *bio_flags,
2104                                          this_bio_flag);
2105                         nr++;
2106                         *bio_flags = this_bio_flag;
2107                 }
2108                 if (ret)
2109                         SetPageError(page);
2110                 cur = cur + iosize;
2111                 page_offset += iosize;
2112         }
2113         if (!nr) {
2114                 if (!PageError(page))
2115                         SetPageUptodate(page);
2116                 unlock_page(page);
2117         }
2118         return 0;
2119 }
2120
2121 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2122                             get_extent_t *get_extent)
2123 {
2124         struct bio *bio = NULL;
2125         unsigned long bio_flags = 0;
2126         int ret;
2127
2128         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2129                                       &bio_flags);
2130         if (bio)
2131                 ret = submit_one_bio(READ, bio, 0, bio_flags);
2132         return ret;
2133 }
2134
2135 static noinline void update_nr_written(struct page *page,
2136                                       struct writeback_control *wbc,
2137                                       unsigned long nr_written)
2138 {
2139         wbc->nr_to_write -= nr_written;
2140         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2141             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2142                 page->mapping->writeback_index = page->index + nr_written;
2143 }
2144
2145 /*
2146  * the writepage semantics are similar to regular writepage.  extent
2147  * records are inserted to lock ranges in the tree, and as dirty areas
2148  * are found, they are marked writeback.  Then the lock bits are removed
2149  * and the end_io handler clears the writeback ranges
2150  */
2151 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2152                               void *data)
2153 {
2154         struct inode *inode = page->mapping->host;
2155         struct extent_page_data *epd = data;
2156         struct extent_io_tree *tree = epd->tree;
2157         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2158         u64 delalloc_start;
2159         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2160         u64 end;
2161         u64 cur = start;
2162         u64 extent_offset;
2163         u64 last_byte = i_size_read(inode);
2164         u64 block_start;
2165         u64 iosize;
2166         sector_t sector;
2167         struct extent_state *cached_state = NULL;
2168         struct extent_map *em;
2169         struct block_device *bdev;
2170         int ret;
2171         int nr = 0;
2172         size_t pg_offset = 0;
2173         size_t blocksize;
2174         loff_t i_size = i_size_read(inode);
2175         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2176         u64 nr_delalloc;
2177         u64 delalloc_end;
2178         int page_started;
2179         int compressed;
2180         int write_flags;
2181         unsigned long nr_written = 0;
2182
2183         if (wbc->sync_mode == WB_SYNC_ALL)
2184                 write_flags = WRITE_SYNC_PLUG;
2185         else
2186                 write_flags = WRITE;
2187
2188         WARN_ON(!PageLocked(page));
2189         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2190         if (page->index > end_index ||
2191            (page->index == end_index && !pg_offset)) {
2192                 page->mapping->a_ops->invalidatepage(page, 0);
2193                 unlock_page(page);
2194                 return 0;
2195         }
2196
2197         if (page->index == end_index) {
2198                 char *userpage;
2199
2200                 userpage = kmap_atomic(page, KM_USER0);
2201                 memset(userpage + pg_offset, 0,
2202                        PAGE_CACHE_SIZE - pg_offset);
2203                 kunmap_atomic(userpage, KM_USER0);
2204                 flush_dcache_page(page);
2205         }
2206         pg_offset = 0;
2207
2208         set_page_extent_mapped(page);
2209
2210         delalloc_start = start;
2211         delalloc_end = 0;
2212         page_started = 0;
2213         if (!epd->extent_locked) {
2214                 u64 delalloc_to_write = 0;
2215                 /*
2216                  * make sure the wbc mapping index is at least updated
2217                  * to this page.
2218                  */
2219                 update_nr_written(page, wbc, 0);
2220
2221                 while (delalloc_end < page_end) {
2222                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2223                                                        page,
2224                                                        &delalloc_start,
2225                                                        &delalloc_end,
2226                                                        128 * 1024 * 1024);
2227                         if (nr_delalloc == 0) {
2228                                 delalloc_start = delalloc_end + 1;
2229                                 continue;
2230                         }
2231                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2232                                                  delalloc_end, &page_started,
2233                                                  &nr_written);
2234                         /*
2235                          * delalloc_end is already one less than the total
2236                          * length, so we don't subtract one from
2237                          * PAGE_CACHE_SIZE
2238                          */
2239                         delalloc_to_write += (delalloc_end - delalloc_start +
2240                                               PAGE_CACHE_SIZE) >>
2241                                               PAGE_CACHE_SHIFT;
2242                         delalloc_start = delalloc_end + 1;
2243                 }
2244                 if (wbc->nr_to_write < delalloc_to_write) {
2245                         int thresh = 8192;
2246
2247                         if (delalloc_to_write < thresh * 2)
2248                                 thresh = delalloc_to_write;
2249                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2250                                                  thresh);
2251                 }
2252
2253                 /* did the fill delalloc function already unlock and start
2254                  * the IO?
2255                  */
2256                 if (page_started) {
2257                         ret = 0;
2258                         /*
2259                          * we've unlocked the page, so we can't update
2260                          * the mapping's writeback index, just update
2261                          * nr_to_write.
2262                          */
2263                         wbc->nr_to_write -= nr_written;
2264                         goto done_unlocked;
2265                 }
2266         }
2267         if (tree->ops && tree->ops->writepage_start_hook) {
2268                 ret = tree->ops->writepage_start_hook(page, start,
2269                                                       page_end);
2270                 if (ret == -EAGAIN) {
2271                         redirty_page_for_writepage(wbc, page);
2272                         update_nr_written(page, wbc, nr_written);
2273                         unlock_page(page);
2274                         ret = 0;
2275                         goto done_unlocked;
2276                 }
2277         }
2278
2279         /*
2280          * we don't want to touch the inode after unlocking the page,
2281          * so we update the mapping writeback index now
2282          */
2283         update_nr_written(page, wbc, nr_written + 1);
2284
2285         end = page_end;
2286         if (last_byte <= start) {
2287                 if (tree->ops && tree->ops->writepage_end_io_hook)
2288                         tree->ops->writepage_end_io_hook(page, start,
2289                                                          page_end, NULL, 1);
2290                 goto done;
2291         }
2292
2293         blocksize = inode->i_sb->s_blocksize;
2294
2295         while (cur <= end) {
2296                 if (cur >= last_byte) {
2297                         if (tree->ops && tree->ops->writepage_end_io_hook)
2298                                 tree->ops->writepage_end_io_hook(page, cur,
2299                                                          page_end, NULL, 1);
2300                         break;
2301                 }
2302                 em = epd->get_extent(inode, page, pg_offset, cur,
2303                                      end - cur + 1, 1);
2304                 if (IS_ERR(em) || !em) {
2305                         SetPageError(page);
2306                         break;
2307                 }
2308
2309                 extent_offset = cur - em->start;
2310                 BUG_ON(extent_map_end(em) <= cur);
2311                 BUG_ON(end < cur);
2312                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2313                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2314                 sector = (em->block_start + extent_offset) >> 9;
2315                 bdev = em->bdev;
2316                 block_start = em->block_start;
2317                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2318                 free_extent_map(em);
2319                 em = NULL;
2320
2321                 /*
2322                  * compressed and inline extents are written through other
2323                  * paths in the FS
2324                  */
2325                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2326                     block_start == EXTENT_MAP_INLINE) {
2327                         /*
2328                          * end_io notification does not happen here for
2329                          * compressed extents
2330                          */
2331                         if (!compressed && tree->ops &&
2332                             tree->ops->writepage_end_io_hook)
2333                                 tree->ops->writepage_end_io_hook(page, cur,
2334                                                          cur + iosize - 1,
2335                                                          NULL, 1);
2336                         else if (compressed) {
2337                                 /* we don't want to end_page_writeback on
2338                                  * a compressed extent.  this happens
2339                                  * elsewhere
2340                                  */
2341                                 nr++;
2342                         }
2343
2344                         cur += iosize;
2345                         pg_offset += iosize;
2346                         continue;
2347                 }
2348                 /* leave this out until we have a page_mkwrite call */
2349                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2350                                    EXTENT_DIRTY, 0, NULL)) {
2351                         cur = cur + iosize;
2352                         pg_offset += iosize;
2353                         continue;
2354                 }
2355
2356                 if (tree->ops && tree->ops->writepage_io_hook) {
2357                         ret = tree->ops->writepage_io_hook(page, cur,
2358                                                 cur + iosize - 1);
2359                 } else {
2360                         ret = 0;
2361                 }
2362                 if (ret) {
2363                         SetPageError(page);
2364                 } else {
2365                         unsigned long max_nr = end_index + 1;
2366
2367                         set_range_writeback(tree, cur, cur + iosize - 1);
2368                         if (!PageWriteback(page)) {
2369                                 printk(KERN_ERR "btrfs warning page %lu not "
2370                                        "writeback, cur %llu end %llu\n",
2371                                        page->index, (unsigned long long)cur,
2372                                        (unsigned long long)end);
2373                         }
2374
2375                         ret = submit_extent_page(write_flags, tree, page,
2376                                                  sector, iosize, pg_offset,
2377                                                  bdev, &epd->bio, max_nr,
2378                                                  end_bio_extent_writepage,
2379                                                  0, 0, 0);
2380                         if (ret)
2381                                 SetPageError(page);
2382                 }
2383                 cur = cur + iosize;
2384                 pg_offset += iosize;
2385                 nr++;
2386         }
2387 done:
2388         if (nr == 0) {
2389                 /* make sure the mapping tag for page dirty gets cleared */
2390                 set_page_writeback(page);
2391                 end_page_writeback(page);
2392         }
2393         unlock_page(page);
2394
2395 done_unlocked:
2396
2397         /* drop our reference on any cached states */
2398         free_extent_state(cached_state);
2399         return 0;
2400 }
2401
2402 /**
2403  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2404  * @mapping: address space structure to write
2405  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2406  * @writepage: function called for each page
2407  * @data: data passed to writepage function
2408  *
2409  * If a page is already under I/O, write_cache_pages() skips it, even
2410  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2411  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2412  * and msync() need to guarantee that all the data which was dirty at the time
2413  * the call was made get new I/O started against them.  If wbc->sync_mode is
2414  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2415  * existing IO to complete.
2416  */
2417 static int extent_write_cache_pages(struct extent_io_tree *tree,
2418                              struct address_space *mapping,
2419                              struct writeback_control *wbc,
2420                              writepage_t writepage, void *data,
2421                              void (*flush_fn)(void *))
2422 {
2423         int ret = 0;
2424         int done = 0;
2425         int nr_to_write_done = 0;
2426         struct pagevec pvec;
2427         int nr_pages;
2428         pgoff_t index;
2429         pgoff_t end;            /* Inclusive */
2430         int scanned = 0;
2431
2432         pagevec_init(&pvec, 0);
2433         if (wbc->range_cyclic) {
2434                 index = mapping->writeback_index; /* Start from prev offset */
2435                 end = -1;
2436         } else {
2437                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2438                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2439                 scanned = 1;
2440         }
2441 retry:
2442         while (!done && !nr_to_write_done && (index <= end) &&
2443                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2444                               PAGECACHE_TAG_DIRTY, min(end - index,
2445                                   (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2446                 unsigned i;
2447
2448                 scanned = 1;
2449                 for (i = 0; i < nr_pages; i++) {
2450                         struct page *page = pvec.pages[i];
2451
2452                         /*
2453                          * At this point we hold neither mapping->tree_lock nor
2454                          * lock on the page itself: the page may be truncated or
2455                          * invalidated (changing page->mapping to NULL), or even
2456                          * swizzled back from swapper_space to tmpfs file
2457                          * mapping
2458                          */
2459                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2460                                 tree->ops->write_cache_pages_lock_hook(page);
2461                         else
2462                                 lock_page(page);
2463
2464                         if (unlikely(page->mapping != mapping)) {
2465                                 unlock_page(page);
2466                                 continue;
2467                         }
2468
2469                         if (!wbc->range_cyclic && page->index > end) {
2470                                 done = 1;
2471                                 unlock_page(page);
2472                                 continue;
2473                         }
2474
2475                         if (wbc->sync_mode != WB_SYNC_NONE) {
2476                                 if (PageWriteback(page))
2477                                         flush_fn(data);
2478                                 wait_on_page_writeback(page);
2479                         }
2480
2481                         if (PageWriteback(page) ||
2482                             !clear_page_dirty_for_io(page)) {
2483                                 unlock_page(page);
2484                                 continue;
2485                         }
2486
2487                         ret = (*writepage)(page, wbc, data);
2488
2489                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2490                                 unlock_page(page);
2491                                 ret = 0;
2492                         }
2493                         if (ret)
2494                                 done = 1;
2495
2496                         /*
2497                          * the filesystem may choose to bump up nr_to_write.
2498                          * We have to make sure to honor the new nr_to_write
2499                          * at any time
2500                          */
2501                         nr_to_write_done = wbc->nr_to_write <= 0;
2502                 }
2503                 pagevec_release(&pvec);
2504                 cond_resched();
2505         }
2506         if (!scanned && !done) {
2507                 /*
2508                  * We hit the last page and there is more work to be done: wrap
2509                  * back to the start of the file
2510                  */
2511                 scanned = 1;
2512                 index = 0;
2513                 goto retry;
2514         }
2515         return ret;
2516 }
2517
2518 static void flush_epd_write_bio(struct extent_page_data *epd)
2519 {
2520         if (epd->bio) {
2521                 if (epd->sync_io)
2522                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2523                 else
2524                         submit_one_bio(WRITE, epd->bio, 0, 0);
2525                 epd->bio = NULL;
2526         }
2527 }
2528
2529 static noinline void flush_write_bio(void *data)
2530 {
2531         struct extent_page_data *epd = data;
2532         flush_epd_write_bio(epd);
2533 }
2534
2535 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2536                           get_extent_t *get_extent,
2537                           struct writeback_control *wbc)
2538 {
2539         int ret;
2540         struct address_space *mapping = page->mapping;
2541         struct extent_page_data epd = {
2542                 .bio = NULL,
2543                 .tree = tree,
2544                 .get_extent = get_extent,
2545                 .extent_locked = 0,
2546                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2547         };
2548         struct writeback_control wbc_writepages = {
2549                 .sync_mode      = wbc->sync_mode,
2550                 .older_than_this = NULL,
2551                 .nr_to_write    = 64,
2552                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2553                 .range_end      = (loff_t)-1,
2554         };
2555
2556         ret = __extent_writepage(page, wbc, &epd);
2557
2558         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2559                                  __extent_writepage, &epd, flush_write_bio);
2560         flush_epd_write_bio(&epd);
2561         return ret;
2562 }
2563
2564 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2565                               u64 start, u64 end, get_extent_t *get_extent,
2566                               int mode)
2567 {
2568         int ret = 0;
2569         struct address_space *mapping = inode->i_mapping;
2570         struct page *page;
2571         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2572                 PAGE_CACHE_SHIFT;
2573
2574         struct extent_page_data epd = {
2575                 .bio = NULL,
2576                 .tree = tree,
2577                 .get_extent = get_extent,
2578                 .extent_locked = 1,
2579                 .sync_io = mode == WB_SYNC_ALL,
2580         };
2581         struct writeback_control wbc_writepages = {
2582                 .sync_mode      = mode,
2583                 .older_than_this = NULL,
2584                 .nr_to_write    = nr_pages * 2,
2585                 .range_start    = start,
2586                 .range_end      = end + 1,
2587         };
2588
2589         while (start <= end) {
2590                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2591                 if (clear_page_dirty_for_io(page))
2592                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2593                 else {
2594                         if (tree->ops && tree->ops->writepage_end_io_hook)
2595                                 tree->ops->writepage_end_io_hook(page, start,
2596                                                  start + PAGE_CACHE_SIZE - 1,
2597                                                  NULL, 1);
2598                         unlock_page(page);
2599                 }
2600                 page_cache_release(page);
2601                 start += PAGE_CACHE_SIZE;
2602         }
2603
2604         flush_epd_write_bio(&epd);
2605         return ret;
2606 }
2607
2608 int extent_writepages(struct extent_io_tree *tree,
2609                       struct address_space *mapping,
2610                       get_extent_t *get_extent,
2611                       struct writeback_control *wbc)
2612 {
2613         int ret = 0;
2614         struct extent_page_data epd = {
2615                 .bio = NULL,
2616                 .tree = tree,
2617                 .get_extent = get_extent,
2618                 .extent_locked = 0,
2619                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2620         };
2621
2622         ret = extent_write_cache_pages(tree, mapping, wbc,
2623                                        __extent_writepage, &epd,
2624                                        flush_write_bio);
2625         flush_epd_write_bio(&epd);
2626         return ret;
2627 }
2628
2629 int extent_readpages(struct extent_io_tree *tree,
2630                      struct address_space *mapping,
2631                      struct list_head *pages, unsigned nr_pages,
2632                      get_extent_t get_extent)
2633 {
2634         struct bio *bio = NULL;
2635         unsigned page_idx;
2636         unsigned long bio_flags = 0;
2637
2638         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2639                 struct page *page = list_entry(pages->prev, struct page, lru);
2640
2641                 prefetchw(&page->flags);
2642                 list_del(&page->lru);
2643                 if (!add_to_page_cache_lru(page, mapping,
2644                                         page->index, GFP_KERNEL)) {
2645                         __extent_read_full_page(tree, page, get_extent,
2646                                                 &bio, 0, &bio_flags);
2647                 }
2648                 page_cache_release(page);
2649         }
2650         BUG_ON(!list_empty(pages));
2651         if (bio)
2652                 submit_one_bio(READ, bio, 0, bio_flags);
2653         return 0;
2654 }
2655
2656 /*
2657  * basic invalidatepage code, this waits on any locked or writeback
2658  * ranges corresponding to the page, and then deletes any extent state
2659  * records from the tree
2660  */
2661 int extent_invalidatepage(struct extent_io_tree *tree,
2662                           struct page *page, unsigned long offset)
2663 {
2664         struct extent_state *cached_state = NULL;
2665         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2666         u64 end = start + PAGE_CACHE_SIZE - 1;
2667         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2668
2669         start += (offset + blocksize - 1) & ~(blocksize - 1);
2670         if (start > end)
2671                 return 0;
2672
2673         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
2674         wait_on_page_writeback(page);
2675         clear_extent_bit(tree, start, end,
2676                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2677                          EXTENT_DO_ACCOUNTING,
2678                          1, 1, &cached_state, GFP_NOFS);
2679         return 0;
2680 }
2681
2682 /*
2683  * simple commit_write call, set_range_dirty is used to mark both
2684  * the pages and the extent records as dirty
2685  */
2686 int extent_commit_write(struct extent_io_tree *tree,
2687                         struct inode *inode, struct page *page,
2688                         unsigned from, unsigned to)
2689 {
2690         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2691
2692         set_page_extent_mapped(page);
2693         set_page_dirty(page);
2694
2695         if (pos > inode->i_size) {
2696                 i_size_write(inode, pos);
2697                 mark_inode_dirty(inode);
2698         }
2699         return 0;
2700 }
2701
2702 int extent_prepare_write(struct extent_io_tree *tree,
2703                          struct inode *inode, struct page *page,
2704                          unsigned from, unsigned to, get_extent_t *get_extent)
2705 {
2706         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2707         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2708         u64 block_start;
2709         u64 orig_block_start;
2710         u64 block_end;
2711         u64 cur_end;
2712         struct extent_map *em;
2713         unsigned blocksize = 1 << inode->i_blkbits;
2714         size_t page_offset = 0;
2715         size_t block_off_start;
2716         size_t block_off_end;
2717         int err = 0;
2718         int iocount = 0;
2719         int ret = 0;
2720         int isnew;
2721
2722         set_page_extent_mapped(page);
2723
2724         block_start = (page_start + from) & ~((u64)blocksize - 1);
2725         block_end = (page_start + to - 1) | (blocksize - 1);
2726         orig_block_start = block_start;
2727
2728         lock_extent(tree, page_start, page_end, GFP_NOFS);
2729         while (block_start <= block_end) {
2730                 em = get_extent(inode, page, page_offset, block_start,
2731                                 block_end - block_start + 1, 1);
2732                 if (IS_ERR(em) || !em)
2733                         goto err;
2734
2735                 cur_end = min(block_end, extent_map_end(em) - 1);
2736                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2737                 block_off_end = block_off_start + blocksize;
2738                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2739
2740                 if (!PageUptodate(page) && isnew &&
2741                     (block_off_end > to || block_off_start < from)) {
2742                         void *kaddr;
2743
2744                         kaddr = kmap_atomic(page, KM_USER0);
2745                         if (block_off_end > to)
2746                                 memset(kaddr + to, 0, block_off_end - to);
2747                         if (block_off_start < from)
2748                                 memset(kaddr + block_off_start, 0,
2749                                        from - block_off_start);
2750                         flush_dcache_page(page);
2751                         kunmap_atomic(kaddr, KM_USER0);
2752                 }
2753                 if ((em->block_start != EXTENT_MAP_HOLE &&
2754                      em->block_start != EXTENT_MAP_INLINE) &&
2755                     !isnew && !PageUptodate(page) &&
2756                     (block_off_end > to || block_off_start < from) &&
2757                     !test_range_bit(tree, block_start, cur_end,
2758                                     EXTENT_UPTODATE, 1, NULL)) {
2759                         u64 sector;
2760                         u64 extent_offset = block_start - em->start;
2761                         size_t iosize;
2762                         sector = (em->block_start + extent_offset) >> 9;
2763                         iosize = (cur_end - block_start + blocksize) &
2764                                 ~((u64)blocksize - 1);
2765                         /*
2766                          * we've already got the extent locked, but we
2767                          * need to split the state such that our end_bio
2768                          * handler can clear the lock.
2769                          */
2770                         set_extent_bit(tree, block_start,
2771                                        block_start + iosize - 1,
2772                                        EXTENT_LOCKED, 0, NULL, NULL, GFP_NOFS);
2773                         ret = submit_extent_page(READ, tree, page,
2774                                          sector, iosize, page_offset, em->bdev,
2775                                          NULL, 1,
2776                                          end_bio_extent_preparewrite, 0,
2777                                          0, 0);
2778                         if (ret && !err)
2779                                 err = ret;
2780                         iocount++;
2781                         block_start = block_start + iosize;
2782                 } else {
2783                         set_extent_uptodate(tree, block_start, cur_end,
2784                                             GFP_NOFS);
2785                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2786                         block_start = cur_end + 1;
2787                 }
2788                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2789                 free_extent_map(em);
2790         }
2791         if (iocount) {
2792                 wait_extent_bit(tree, orig_block_start,
2793                                 block_end, EXTENT_LOCKED);
2794         }
2795         check_page_uptodate(tree, page);
2796 err:
2797         /* FIXME, zero out newly allocated blocks on error */
2798         return err;
2799 }
2800
2801 /*
2802  * a helper for releasepage, this tests for areas of the page that
2803  * are locked or under IO and drops the related state bits if it is safe
2804  * to drop the page.
2805  */
2806 int try_release_extent_state(struct extent_map_tree *map,
2807                              struct extent_io_tree *tree, struct page *page,
2808                              gfp_t mask)
2809 {
2810         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2811         u64 end = start + PAGE_CACHE_SIZE - 1;
2812         int ret = 1;
2813
2814         if (test_range_bit(tree, start, end,
2815                            EXTENT_IOBITS, 0, NULL))
2816                 ret = 0;
2817         else {
2818                 if ((mask & GFP_NOFS) == GFP_NOFS)
2819                         mask = GFP_NOFS;
2820                 /*
2821                  * at this point we can safely clear everything except the
2822                  * locked bit and the nodatasum bit
2823                  */
2824                 clear_extent_bit(tree, start, end,
2825                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2826                                  0, 0, NULL, mask);
2827         }
2828         return ret;
2829 }
2830
2831 /*
2832  * a helper for releasepage.  As long as there are no locked extents
2833  * in the range corresponding to the page, both state records and extent
2834  * map records are removed
2835  */
2836 int try_release_extent_mapping(struct extent_map_tree *map,
2837                                struct extent_io_tree *tree, struct page *page,
2838                                gfp_t mask)
2839 {
2840         struct extent_map *em;
2841         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2842         u64 end = start + PAGE_CACHE_SIZE - 1;
2843
2844         if ((mask & __GFP_WAIT) &&
2845             page->mapping->host->i_size > 16 * 1024 * 1024) {
2846                 u64 len;
2847                 while (start <= end) {
2848                         len = end - start + 1;
2849                         write_lock(&map->lock);
2850                         em = lookup_extent_mapping(map, start, len);
2851                         if (!em || IS_ERR(em)) {
2852                                 write_unlock(&map->lock);
2853                                 break;
2854                         }
2855                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2856                             em->start != start) {
2857                                 write_unlock(&map->lock);
2858                                 free_extent_map(em);
2859                                 break;
2860                         }
2861                         if (!test_range_bit(tree, em->start,
2862                                             extent_map_end(em) - 1,
2863                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
2864                                             0, NULL)) {
2865                                 remove_extent_mapping(map, em);
2866                                 /* once for the rb tree */
2867                                 free_extent_map(em);
2868                         }
2869                         start = extent_map_end(em);
2870                         write_unlock(&map->lock);
2871
2872                         /* once for us */
2873                         free_extent_map(em);
2874                 }
2875         }
2876         return try_release_extent_state(map, tree, page, mask);
2877 }
2878
2879 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2880                 get_extent_t *get_extent)
2881 {
2882         struct inode *inode = mapping->host;
2883         struct extent_state *cached_state = NULL;
2884         u64 start = iblock << inode->i_blkbits;
2885         sector_t sector = 0;
2886         size_t blksize = (1 << inode->i_blkbits);
2887         struct extent_map *em;
2888
2889         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2890                          0, &cached_state, GFP_NOFS);
2891         em = get_extent(inode, NULL, 0, start, blksize, 0);
2892         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start,
2893                              start + blksize - 1, &cached_state, GFP_NOFS);
2894         if (!em || IS_ERR(em))
2895                 return 0;
2896
2897         if (em->block_start > EXTENT_MAP_LAST_BYTE)
2898                 goto out;
2899
2900         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2901 out:
2902         free_extent_map(em);
2903         return sector;
2904 }
2905
2906 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2907                 __u64 start, __u64 len, get_extent_t *get_extent)
2908 {
2909         int ret = 0;
2910         u64 off = start;
2911         u64 max = start + len;
2912         u32 flags = 0;
2913         u32 found_type;
2914         u64 last;
2915         u64 disko = 0;
2916         struct btrfs_key found_key;
2917         struct extent_map *em = NULL;
2918         struct extent_state *cached_state = NULL;
2919         struct btrfs_path *path;
2920         struct btrfs_file_extent_item *item;
2921         int end = 0;
2922         u64 em_start = 0, em_len = 0;
2923         unsigned long emflags;
2924         int hole = 0;
2925
2926         if (len == 0)
2927                 return -EINVAL;
2928
2929         path = btrfs_alloc_path();
2930         if (!path)
2931                 return -ENOMEM;
2932         path->leave_spinning = 1;
2933
2934         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
2935                                        path, inode->i_ino, -1, 0);
2936         if (ret < 0) {
2937                 btrfs_free_path(path);
2938                 return ret;
2939         }
2940         WARN_ON(!ret);
2941         path->slots[0]--;
2942         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2943                               struct btrfs_file_extent_item);
2944         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
2945         found_type = btrfs_key_type(&found_key);
2946
2947         /* No extents, just return */
2948         if (found_key.objectid != inode->i_ino ||
2949             found_type != BTRFS_EXTENT_DATA_KEY) {
2950                 btrfs_free_path(path);
2951                 return 0;
2952         }
2953         last = found_key.offset;
2954         btrfs_free_path(path);
2955
2956         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
2957                          &cached_state, GFP_NOFS);
2958         em = get_extent(inode, NULL, 0, off, max - off, 0);
2959         if (!em)
2960                 goto out;
2961         if (IS_ERR(em)) {
2962                 ret = PTR_ERR(em);
2963                 goto out;
2964         }
2965
2966         while (!end) {
2967                 hole = 0;
2968                 off = em->start + em->len;
2969                 if (off >= max)
2970                         end = 1;
2971
2972                 if (em->block_start == EXTENT_MAP_HOLE) {
2973                         hole = 1;
2974                         goto next;
2975                 }
2976
2977                 em_start = em->start;
2978                 em_len = em->len;
2979
2980                 disko = 0;
2981                 flags = 0;
2982
2983                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
2984                         end = 1;
2985                         flags |= FIEMAP_EXTENT_LAST;
2986                 } else if (em->block_start == EXTENT_MAP_INLINE) {
2987                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
2988                                   FIEMAP_EXTENT_NOT_ALIGNED);
2989                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
2990                         flags |= (FIEMAP_EXTENT_DELALLOC |
2991                                   FIEMAP_EXTENT_UNKNOWN);
2992                 } else {
2993                         disko = em->block_start;
2994                 }
2995                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2996                         flags |= FIEMAP_EXTENT_ENCODED;
2997
2998 next:
2999                 emflags = em->flags;
3000                 free_extent_map(em);
3001                 em = NULL;
3002                 if (!end) {
3003                         em = get_extent(inode, NULL, 0, off, max - off, 0);
3004                         if (!em)
3005                                 goto out;
3006                         if (IS_ERR(em)) {
3007                                 ret = PTR_ERR(em);
3008                                 goto out;
3009                         }
3010                         emflags = em->flags;
3011                 }
3012
3013                 if (test_bit(EXTENT_FLAG_VACANCY, &emflags)) {
3014                         flags |= FIEMAP_EXTENT_LAST;
3015                         end = 1;
3016                 }
3017
3018                 if (em_start == last) {
3019                         flags |= FIEMAP_EXTENT_LAST;
3020                         end = 1;
3021                 }
3022
3023                 if (!hole) {
3024                         ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3025                                                 em_len, flags);
3026                         if (ret)
3027                                 goto out_free;
3028                 }
3029         }
3030 out_free:
3031         free_extent_map(em);
3032 out:
3033         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3034                              &cached_state, GFP_NOFS);
3035         return ret;
3036 }
3037
3038 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
3039                                               unsigned long i)
3040 {
3041         struct page *p;
3042         struct address_space *mapping;
3043
3044         if (i == 0)
3045                 return eb->first_page;
3046         i += eb->start >> PAGE_CACHE_SHIFT;
3047         mapping = eb->first_page->mapping;
3048         if (!mapping)
3049                 return NULL;
3050
3051         /*
3052          * extent_buffer_page is only called after pinning the page