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 <linux/prefetch.h>
  14 #include <linux/cleancache.h>
  15 #include "extent_io.h"
  16 #include "extent_map.h"
  17 #include "compat.h"
  18 #include "ctree.h"
  19 #include "btrfs_inode.h"
  20
  21 static struct kmem_cache *extent_state_cache;
  22 static struct kmem_cache *extent_buffer_cache;
  23
  24 static LIST_HEAD(buffers);
  25 static LIST_HEAD(states);
  26
  27 #define LEAK_DEBUG 0
  28 #if LEAK_DEBUG
  29 static DEFINE_SPINLOCK(leak_lock);
  30 #endif
  31
  32 #define BUFFER_LRU_MAX 64
  33
  34 struct tree_entry {
  35         u64 start;
  36         u64 end;
  37         struct rb_node rb_node;
  38 };
  39
  40 struct extent_page_data {
  41         struct bio *bio;
  42         struct extent_io_tree *tree;
  43         get_extent_t *get_extent;
  44
  45         /* tells writepage not to lock the state bits for this range
  46          * it still does the unlocking
  47          */
  48         unsigned int extent_locked:1;
  49
  50         /* tells the submit_bio code to use a WRITE_SYNC */
  51         unsigned int sync_io:1;
  52 };
  53
  54 int __init extent_io_init(void)
  55 {
  56         extent_state_cache = kmem_cache_create("extent_state",
  57                         sizeof(struct extent_state), 0,
  58                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  59         if (!extent_state_cache)
  60                 return -ENOMEM;
  61
  62         extent_buffer_cache = kmem_cache_create("extent_buffers",
  63                         sizeof(struct extent_buffer), 0,
  64                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  65         if (!extent_buffer_cache)
  66                 goto free_state_cache;
  67         return 0;
  68
  69 free_state_cache:
  70         kmem_cache_destroy(extent_state_cache);
  71         return -ENOMEM;
  72 }
  73
  74 void extent_io_exit(void)
  75 {
  76         struct extent_state *state;
  77         struct extent_buffer *eb;
  78
  79         while (!list_empty(&states)) {
  80                 state = list_entry(states.next, struct extent_state, leak_list);
  81                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  82                        "state %lu in tree %p refs %d\n",
  83                        (unsigned long long)state->start,
  84                        (unsigned long long)state->end,
  85                        state->state, state->tree, atomic_read(&state->refs));
  86                 list_del(&state->leak_list);
  87                 kmem_cache_free(extent_state_cache, state);
  88
  89         }
  90
  91         while (!list_empty(&buffers)) {
  92                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  93                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  94                        "refs %d\n", (unsigned long long)eb->start,
  95                        eb->len, atomic_read(&eb->refs));
  96                 list_del(&eb->leak_list);
  97                 kmem_cache_free(extent_buffer_cache, eb);
  98         }
  99         if (extent_state_cache)
 100                 kmem_cache_destroy(extent_state_cache);
 101         if (extent_buffer_cache)
 102                 kmem_cache_destroy(extent_buffer_cache);
 103 }
 104
 105 void extent_io_tree_init(struct extent_io_tree *tree,
 106                          struct address_space *mapping)
 107 {
 108         tree->state = RB_ROOT;
 109         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 110         tree->ops = NULL;
 111         tree->dirty_bytes = 0;
 112         spin_lock_init(&tree->lock);
 113         spin_lock_init(&tree->buffer_lock);
 114         tree->mapping = mapping;
 115 }
 116
 117 static struct extent_state *alloc_extent_state(gfp_t mask)
 118 {
 119         struct extent_state *state;
 120 #if LEAK_DEBUG
 121         unsigned long flags;
 122 #endif
 123
 124         state = kmem_cache_alloc(extent_state_cache, mask);
 125         if (!state)
 126                 return state;
 127         state->state = 0;
 128         state->private = 0;
 129         state->tree = NULL;
 130 #if LEAK_DEBUG
 131         spin_lock_irqsave(&leak_lock, flags);
 132         list_add(&state->leak_list, &states);
 133         spin_unlock_irqrestore(&leak_lock, flags);
 134 #endif
 135         atomic_set(&state->refs, 1);
 136         init_waitqueue_head(&state->wq);
 137         return state;
 138 }
 139
 140 void free_extent_state(struct extent_state *state)
 141 {
 142         if (!state)
 143                 return;
 144         if (atomic_dec_and_test(&state->refs)) {
 145 #if LEAK_DEBUG
 146                 unsigned long flags;
 147 #endif
 148                 WARN_ON(state->tree);
 149 #if LEAK_DEBUG
 150                 spin_lock_irqsave(&leak_lock, flags);
 151                 list_del(&state->leak_list);
 152                 spin_unlock_irqrestore(&leak_lock, flags);
 153 #endif
 154                 kmem_cache_free(extent_state_cache, state);
 155         }
 156 }
 157
 158 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 159                                    struct rb_node *node)
 160 {
 161         struct rb_node **p = &root->rb_node;
 162         struct rb_node *parent = NULL;
 163         struct tree_entry *entry;
 164
 165         while (*p) {
 166                 parent = *p;
 167                 entry = rb_entry(parent, struct tree_entry, rb_node);
 168
 169                 if (offset < entry->start)
 170                         p = &(*p)->rb_left;
 171                 else if (offset > entry->end)
 172                         p = &(*p)->rb_right;
 173                 else
 174                         return parent;
 175         }
 176
 177         entry = rb_entry(node, struct tree_entry, rb_node);
 178         rb_link_node(node, parent, p);
 179         rb_insert_color(node, root);
 180         return NULL;
 181 }
 182
 183 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 184                                      struct rb_node **prev_ret,
 185                                      struct rb_node **next_ret)
 186 {
 187         struct rb_root *root = &tree->state;
 188         struct rb_node *n = root->rb_node;
 189         struct rb_node *prev = NULL;
 190         struct rb_node *orig_prev = NULL;
 191         struct tree_entry *entry;
 192         struct tree_entry *prev_entry = NULL;
 193
 194         while (n) {
 195                 entry = rb_entry(n, struct tree_entry, rb_node);
 196                 prev = n;
 197                 prev_entry = entry;
 198
 199                 if (offset < entry->start)
 200                         n = n->rb_left;
 201                 else if (offset > entry->end)
 202                         n = n->rb_right;
 203                 else
 204                         return n;
 205         }
 206
 207         if (prev_ret) {
 208                 orig_prev = prev;
 209                 while (prev && offset > prev_entry->end) {
 210                         prev = rb_next(prev);
 211                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 212                 }
 213                 *prev_ret = prev;
 214                 prev = orig_prev;
 215         }
 216
 217         if (next_ret) {
 218                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 219                 while (prev && offset < prev_entry->start) {
 220                         prev = rb_prev(prev);
 221                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 222                 }
 223                 *next_ret = prev;
 224         }
 225         return NULL;
 226 }
 227
 228 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 229                                           u64 offset)
 230 {
 231         struct rb_node *prev = NULL;
 232         struct rb_node *ret;
 233
 234         ret = __etree_search(tree, offset, &prev, NULL);
 235         if (!ret)
 236                 return prev;
 237         return ret;
 238 }
 239
 240 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 241                      struct extent_state *other)
 242 {
 243         if (tree->ops && tree->ops->merge_extent_hook)
 244                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 245                                              other);
 246 }
 247
 248 /*
 249  * utility function to look for merge candidates inside a given range.
 250  * Any extents with matching state are merged together into a single
 251  * extent in the tree.  Extents with EXTENT_IO in their state field
 252  * are not merged because the end_io handlers need to be able to do
 253  * operations on them without sleeping (or doing allocations/splits).
 254  *
 255  * This should be called with the tree lock held.
 256  */
 257 static int merge_state(struct extent_io_tree *tree,
 258                        struct extent_state *state)
 259 {
 260         struct extent_state *other;
 261         struct rb_node *other_node;
 262
 263         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 264                 return 0;
 265
 266         other_node = rb_prev(&state->rb_node);
 267         if (other_node) {
 268                 other = rb_entry(other_node, struct extent_state, rb_node);
 269                 if (other->end == state->start - 1 &&
 270                     other->state == state->state) {
 271                         merge_cb(tree, state, other);
 272                         state->start = other->start;
 273                         other->tree = NULL;
 274                         rb_erase(&other->rb_node, &tree->state);
 275                         free_extent_state(other);
 276                 }
 277         }
 278         other_node = rb_next(&state->rb_node);
 279         if (other_node) {
 280                 other = rb_entry(other_node, struct extent_state, rb_node);
 281                 if (other->start == state->end + 1 &&
 282                     other->state == state->state) {
 283                         merge_cb(tree, state, other);
 284                         state->end = other->end;
 285                         other->tree = NULL;
 286                         rb_erase(&other->rb_node, &tree->state);
 287                         free_extent_state(other);
 288                 }
 289         }
 290
 291         return 0;
 292 }
 293
 294 static int set_state_cb(struct extent_io_tree *tree,
 295                          struct extent_state *state, int *bits)
 296 {
 297         if (tree->ops && tree->ops->set_bit_hook) {
 298                 return tree->ops->set_bit_hook(tree->mapping->host,
 299                                                state, bits);
 300         }
 301
 302         return 0;
 303 }
 304
 305 static void clear_state_cb(struct extent_io_tree *tree,
 306                            struct extent_state *state, int *bits)
 307 {
 308         if (tree->ops && tree->ops->clear_bit_hook)
 309                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 310 }
 311
 312 /*
 313  * insert an extent_state struct into the tree.  'bits' are set on the
 314  * struct before it is inserted.
 315  *
 316  * This may return -EEXIST if the extent is already there, in which case the
 317  * state struct is freed.
 318  *
 319  * The tree lock is not taken internally.  This is a utility function and
 320  * probably isn't what you want to call (see set/clear_extent_bit).
 321  */
 322 static int insert_state(struct extent_io_tree *tree,
 323                         struct extent_state *state, u64 start, u64 end,
 324                         int *bits)
 325 {
 326         struct rb_node *node;
 327         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 328         int ret;
 329
 330         if (end < start) {
 331                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
 332                        (unsigned long long)end,
 333                        (unsigned long long)start);
 334                 WARN_ON(1);
 335         }
 336         state->start = start;
 337         state->end = end;
 338         ret = set_state_cb(tree, state, bits);
 339         if (ret)
 340                 return ret;
 341
 342         if (bits_to_set & EXTENT_DIRTY)
 343                 tree->dirty_bytes += end - start + 1;
 344         state->state |= bits_to_set;
 345         node = tree_insert(&tree->state, end, &state->rb_node);
 346         if (node) {
 347                 struct extent_state *found;
 348                 found = rb_entry(node, struct extent_state, rb_node);
 349                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 350                        "%llu %llu\n", (unsigned long long)found->start,
 351                        (unsigned long long)found->end,
 352                        (unsigned long long)start, (unsigned long long)end);
 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 static struct extent_state *
 443 alloc_extent_state_atomic(struct extent_state *prealloc)
 444 {
 445         if (!prealloc)
 446                 prealloc = alloc_extent_state(GFP_ATOMIC);
 447
 448         return prealloc;
 449 }
 450
 451 /*
 452  * clear some bits on a range in the tree.  This may require splitting
 453  * or inserting elements in the tree, so the gfp mask is used to
 454  * indicate which allocations or sleeping are allowed.
 455  *
 456  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 457  * the given range from the tree regardless of state (ie for truncate).
 458  *
 459  * the range [start, end] is inclusive.
 460  *
 461  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 462  * bits were already set, or zero if none of the bits were already set.
 463  */
 464 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 465                      int bits, int wake, int delete,
 466                      struct extent_state **cached_state,
 467                      gfp_t mask)
 468 {
 469         struct extent_state *state;
 470         struct extent_state *cached;
 471         struct extent_state *prealloc = NULL;
 472         struct rb_node *next_node;
 473         struct rb_node *node;
 474         u64 last_end;
 475         int err;
 476         int set = 0;
 477         int clear = 0;
 478
 479         if (delete)
 480                 bits |= ~EXTENT_CTLBITS;
 481         bits |= EXTENT_FIRST_DELALLOC;
 482
 483         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 484                 clear = 1;
 485 again:
 486         if (!prealloc && (mask & __GFP_WAIT)) {
 487                 prealloc = alloc_extent_state(mask);
 488                 if (!prealloc)
 489                         return -ENOMEM;
 490         }
 491
 492         spin_lock(&tree->lock);
 493         if (cached_state) {
 494                 cached = *cached_state;
 495
 496                 if (clear) {
 497                         *cached_state = NULL;
 498                         cached_state = NULL;
 499                 }
 500
 501                 if (cached && cached->tree && cached->start <= start &&
 502                     cached->end > start) {
 503                         if (clear)
 504                                 atomic_dec(&cached->refs);
 505                         state = cached;
 506                         goto hit_next;
 507                 }
 508                 if (clear)
 509                         free_extent_state(cached);
 510         }
 511         /*
 512          * this search will find the extents that end after
 513          * our range starts
 514          */
 515         node = tree_search(tree, start);
 516         if (!node)
 517                 goto out;
 518         state = rb_entry(node, struct extent_state, rb_node);
 519 hit_next:
 520         if (state->start > end)
 521                 goto out;
 522         WARN_ON(state->end < start);
 523         last_end = state->end;
 524
 525         /*
 526          *     | ---- desired range ---- |
 527          *  | state | or
 528          *  | ------------- state -------------- |
 529          *
 530          * We need to split the extent we found, and may flip
 531          * bits on second half.
 532          *
 533          * If the extent we found extends past our range, we
 534          * just split and search again.  It'll get split again
 535          * the next time though.
 536          *
 537          * If the extent we found is inside our range, we clear
 538          * the desired bit on it.
 539          */
 540
 541         if (state->start < start) {
 542                 prealloc = alloc_extent_state_atomic(prealloc);
 543                 BUG_ON(!prealloc);
 544                 err = split_state(tree, state, prealloc, start);
 545                 BUG_ON(err == -EEXIST);
 546                 prealloc = NULL;
 547                 if (err)
 548                         goto out;
 549                 if (state->end <= end) {
 550                         set |= clear_state_bit(tree, state, &bits, wake);
 551                         if (last_end == (u64)-1)
 552                                 goto out;
 553                         start = last_end + 1;
 554                 }
 555                 goto search_again;
 556         }
 557         /*
 558          * | ---- desired range ---- |
 559          *                        | state |
 560          * We need to split the extent, and clear the bit
 561          * on the first half
 562          */
 563         if (state->start <= end && state->end > end) {
 564                 prealloc = alloc_extent_state_atomic(prealloc);
 565                 BUG_ON(!prealloc);
 566                 err = split_state(tree, state, prealloc, end + 1);
 567                 BUG_ON(err == -EEXIST);
 568                 if (wake)
 569                         wake_up(&state->wq);
 570
 571                 set |= clear_state_bit(tree, prealloc, &bits, wake);
 572
 573                 prealloc = NULL;
 574                 goto out;
 575         }
 576
 577         if (state->end < end && prealloc && !need_resched())
 578                 next_node = rb_next(&state->rb_node);
 579         else
 580                 next_node = NULL;
 581
 582         set |= clear_state_bit(tree, state, &bits, wake);
 583         if (last_end == (u64)-1)
 584                 goto out;
 585         start = last_end + 1;
 586         if (start <= end && next_node) {
 587                 state = rb_entry(next_node, struct extent_state,
 588                                  rb_node);
 589                 if (state->start == start)
 590                         goto hit_next;
 591         }
 592         goto search_again;
 593
 594 out:
 595         spin_unlock(&tree->lock);
 596         if (prealloc)
 597                 free_extent_state(prealloc);
 598
 599         return set;
 600
 601 search_again:
 602         if (start > end)
 603                 goto out;
 604         spin_unlock(&tree->lock);
 605         if (mask & __GFP_WAIT)
 606                 cond_resched();
 607         goto again;
 608 }
 609
 610 static int wait_on_state(struct extent_io_tree *tree,
 611                          struct extent_state *state)
 612                 __releases(tree->lock)
 613                 __acquires(tree->lock)
 614 {
 615         DEFINE_WAIT(wait);
 616         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 617         spin_unlock(&tree->lock);
 618         schedule();
 619         spin_lock(&tree->lock);
 620         finish_wait(&state->wq, &wait);
 621         return 0;
 622 }
 623
 624 /*
 625  * waits for one or more bits to clear on a range in the state tree.
 626  * The range [start, end] is inclusive.
 627  * The tree lock is taken by this function
 628  */
 629 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 630 {
 631         struct extent_state *state;
 632         struct rb_node *node;
 633
 634         spin_lock(&tree->lock);
 635 again:
 636         while (1) {
 637                 /*
 638                  * this search will find all the extents that end after
 639                  * our range starts
 640                  */
 641                 node = tree_search(tree, start);
 642                 if (!node)
 643                         break;
 644
 645                 state = rb_entry(node, struct extent_state, rb_node);
 646
 647                 if (state->start > end)
 648                         goto out;
 649
 650                 if (state->state & bits) {
 651                         start = state->start;
 652                         atomic_inc(&state->refs);
 653                         wait_on_state(tree, state);
 654                         free_extent_state(state);
 655                         goto again;
 656                 }
 657                 start = state->end + 1;
 658
 659                 if (start > end)
 660                         break;
 661
 662                 if (need_resched()) {
 663                         spin_unlock(&tree->lock);
 664                         cond_resched();
 665                         spin_lock(&tree->lock);
 666                 }
 667         }
 668 out:
 669         spin_unlock(&tree->lock);
 670         return 0;
 671 }
 672
 673 static int set_state_bits(struct extent_io_tree *tree,
 674                            struct extent_state *state,
 675                            int *bits)
 676 {
 677         int ret;
 678         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 679
 680         ret = set_state_cb(tree, state, bits);
 681         if (ret)
 682                 return ret;
 683         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 684                 u64 range = state->end - state->start + 1;
 685                 tree->dirty_bytes += range;
 686         }
 687         state->state |= bits_to_set;
 688
 689         return 0;
 690 }
 691
 692 static void cache_state(struct extent_state *state,
 693                         struct extent_state **cached_ptr)
 694 {
 695         if (cached_ptr && !(*cached_ptr)) {
 696                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 697                         *cached_ptr = state;
 698                         atomic_inc(&state->refs);
 699                 }
 700         }
 701 }
 702
 703 static void uncache_state(struct extent_state **cached_ptr)
 704 {
 705         if (cached_ptr && (*cached_ptr)) {
 706                 struct extent_state *state = *cached_ptr;
 707                 *cached_ptr = NULL;
 708                 free_extent_state(state);
 709         }
 710 }
 711
 712 /*
 713  * set some bits on a range in the tree.  This may require allocations or
 714  * sleeping, so the gfp mask is used to indicate what is allowed.
 715  *
 716  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 717  * part of the range already has the desired bits set.  The start of the
 718  * existing range is returned in failed_start in this case.
 719  *
 720  * [start, end] is inclusive This takes the tree lock.
 721  */
 722
 723 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 724                    int bits, int exclusive_bits, u64 *failed_start,
 725                    struct extent_state **cached_state, gfp_t mask)
 726 {
 727         struct extent_state *state;
 728         struct extent_state *prealloc = NULL;
 729         struct rb_node *node;
 730         int err = 0;
 731         u64 last_start;
 732         u64 last_end;
 733
 734         bits |= EXTENT_FIRST_DELALLOC;
 735 again:
 736         if (!prealloc && (mask & __GFP_WAIT)) {
 737                 prealloc = alloc_extent_state(mask);
 738                 BUG_ON(!prealloc);
 739         }
 740
 741         spin_lock(&tree->lock);
 742         if (cached_state && *cached_state) {
 743                 state = *cached_state;
 744                 if (state->start <= start && state->end > start &&
 745                     state->tree) {
 746                         node = &state->rb_node;
 747                         goto hit_next;
 748                 }
 749         }
 750         /*
 751          * this search will find all the extents that end after
 752          * our range starts.
 753          */
 754         node = tree_search(tree, start);
 755         if (!node) {
 756                 prealloc = alloc_extent_state_atomic(prealloc);
 757                 BUG_ON(!prealloc);
 758                 err = insert_state(tree, prealloc, start, end, &bits);
 759                 prealloc = NULL;
 760                 BUG_ON(err == -EEXIST);
 761                 goto out;
 762         }
 763         state = rb_entry(node, struct extent_state, rb_node);
 764 hit_next:
 765         last_start = state->start;
 766         last_end = state->end;
 767
 768         /*
 769          * | ---- desired range ---- |
 770          * | state |
 771          *
 772          * Just lock what we found and keep going
 773          */
 774         if (state->start == start && state->end <= end) {
 775                 struct rb_node *next_node;
 776                 if (state->state & exclusive_bits) {
 777                         *failed_start = state->start;
 778                         err = -EEXIST;
 779                         goto out;
 780                 }
 781
 782                 err = set_state_bits(tree, state, &bits);
 783                 if (err)
 784                         goto out;
 785
 786                 cache_state(state, cached_state);
 787                 merge_state(tree, state);
 788                 if (last_end == (u64)-1)
 789                         goto out;
 790
 791                 start = last_end + 1;
 792                 next_node = rb_next(&state->rb_node);
 793                 if (next_node && start < end && prealloc && !need_resched()) {
 794                         state = rb_entry(next_node, struct extent_state,
 795                                          rb_node);
 796                         if (state->start == start)
 797                                 goto hit_next;
 798                 }
 799                 goto search_again;
 800         }
 801
 802         /*
 803          *     | ---- desired range ---- |
 804          * | state |
 805          *   or
 806          * | ------------- state -------------- |
 807          *
 808          * We need to split the extent we found, and may flip bits on
 809          * second half.
 810          *
 811          * If the extent we found extends past our
 812          * range, we just split and search again.  It'll get split
 813          * again the next time though.
 814          *
 815          * If the extent we found is inside our range, we set the
 816          * desired bit on it.
 817          */
 818         if (state->start < start) {
 819                 if (state->state & exclusive_bits) {
 820                         *failed_start = start;
 821                         err = -EEXIST;
 822                         goto out;
 823                 }
 824
 825                 prealloc = alloc_extent_state_atomic(prealloc);
 826                 BUG_ON(!prealloc);
 827                 err = split_state(tree, state, prealloc, start);
 828                 BUG_ON(err == -EEXIST);
 829                 prealloc = NULL;
 830                 if (err)
 831                         goto out;
 832                 if (state->end <= end) {
 833                         err = set_state_bits(tree, state, &bits);
 834                         if (err)
 835                                 goto out;
 836                         cache_state(state, cached_state);
 837                         merge_state(tree, state);
 838                         if (last_end == (u64)-1)
 839                                 goto out;
 840                         start = last_end + 1;
 841                 }
 842                 goto search_again;
 843         }
 844         /*
 845          * | ---- desired range ---- |
 846          *     | state | or               | state |
 847          *
 848          * There's a hole, we need to insert something in it and
 849          * ignore the extent we found.
 850          */
 851         if (state->start > start) {
 852                 u64 this_end;
 853                 if (end < last_start)
 854                         this_end = end;
 855                 else
 856                         this_end = last_start - 1;
 857
 858                 prealloc = alloc_extent_state_atomic(prealloc);
 859                 BUG_ON(!prealloc);
 860
 861                 /*
 862                  * Avoid to free 'prealloc' if it can be merged with
 863                  * the later extent.
 864                  */
 865                 err = insert_state(tree, prealloc, start, this_end,
 866                                    &bits);
 867                 BUG_ON(err == -EEXIST);
 868                 if (err) {
 869                         free_extent_state(prealloc);
 870                         prealloc = NULL;
 871                         goto out;
 872                 }
 873                 cache_state(prealloc, cached_state);
 874                 prealloc = NULL;
 875                 start = this_end + 1;
 876                 goto search_again;
 877         }
 878         /*
 879          * | ---- desired range ---- |
 880          *                        | state |
 881          * We need to split the extent, and set the bit
 882          * on the first half
 883          */
 884         if (state->start <= end && state->end > end) {
 885                 if (state->state & exclusive_bits) {
 886                         *failed_start = start;
 887                         err = -EEXIST;
 888                         goto out;
 889                 }
 890
 891                 prealloc = alloc_extent_state_atomic(prealloc);
 892                 BUG_ON(!prealloc);
 893                 err = split_state(tree, state, prealloc, end + 1);
 894                 BUG_ON(err == -EEXIST);
 895
 896                 err = set_state_bits(tree, prealloc, &bits);
 897                 if (err) {
 898                         prealloc = NULL;
 899                         goto out;
 900                 }
 901                 cache_state(prealloc, cached_state);
 902                 merge_state(tree, prealloc);
 903                 prealloc = NULL;
 904                 goto out;
 905         }
 906
 907         goto search_again;
 908
 909 out:
 910         spin_unlock(&tree->lock);
 911         if (prealloc)
 912                 free_extent_state(prealloc);
 913
 914         return err;
 915
 916 search_again:
 917         if (start > end)
 918                 goto out;
 919         spin_unlock(&tree->lock);
 920         if (mask & __GFP_WAIT)
 921                 cond_resched();
 922         goto again;
 923 }
 924
 925 /* wrappers around set/clear extent bit */
 926 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 927                      gfp_t mask)
 928 {
 929         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 930                               NULL, mask);
 931 }
 932
 933 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 934                     int bits, gfp_t mask)
 935 {
 936         return set_extent_bit(tree, start, end, bits, 0, NULL,
 937                               NULL, mask);
 938 }
 939
 940 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 941                       int bits, gfp_t mask)
 942 {
 943         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
 944 }
 945
 946 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
 947                         struct extent_state **cached_state, gfp_t mask)
 948 {
 949         return set_extent_bit(tree, start, end,
 950                               EXTENT_DELALLOC | EXTENT_DIRTY | EXTENT_UPTODATE,
 951                               0, NULL, cached_state, mask);
 952 }
 953
 954 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 955                        gfp_t mask)
 956 {
 957         return clear_extent_bit(tree, start, end,
 958                                 EXTENT_DIRTY | EXTENT_DELALLOC |
 959                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
 960 }
 961
 962 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 963                      gfp_t mask)
 964 {
 965         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 966                               NULL, mask);
 967 }
 968
 969 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 970                         struct extent_state **cached_state, gfp_t mask)
 971 {
 972         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
 973                               NULL, cached_state, mask);
 974 }
 975
 976 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
 977                                  u64 end, struct extent_state **cached_state,
 978                                  gfp_t mask)
 979 {
 980         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
 981                                 cached_state, mask);
 982 }
 983
 984 /*
 985  * either insert or lock state struct between start and end use mask to tell
 986  * us if waiting is desired.
 987  */
 988 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 989                      int bits, struct extent_state **cached_state, gfp_t mask)
 990 {
 991         int err;
 992         u64 failed_start;
 993         while (1) {
 994                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
 995                                      EXTENT_LOCKED, &failed_start,
 996                                      cached_state, mask);
 997                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
 998                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
 999                         start = failed_start;
1000                 } else {
1001                         break;
1002                 }
1003                 WARN_ON(start > end);
1004         }
1005         return err;
1006 }
1007
1008 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1009 {
1010         return lock_extent_bits(tree, start, end, 0, NULL, mask);
1011 }
1012
1013 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1014                     gfp_t mask)
1015 {
1016         int err;
1017         u64 failed_start;
1018
1019         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1020                              &failed_start, NULL, mask);
1021         if (err == -EEXIST) {
1022                 if (failed_start > start)
1023                         clear_extent_bit(tree, start, failed_start - 1,
1024                                          EXTENT_LOCKED, 1, 0, NULL, mask);
1025                 return 0;
1026         }
1027         return 1;
1028 }
1029
1030 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1031                          struct extent_state **cached, gfp_t mask)
1032 {
1033         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1034                                 mask);
1035 }
1036
1037 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1038 {
1039         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1040                                 mask);
1041 }
1042
1043 /*
1044  * helper function to set both pages and extents in the tree writeback
1045  */
1046 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1047 {
1048         unsigned long index = start >> PAGE_CACHE_SHIFT;
1049         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1050         struct page *page;
1051
1052         while (index <= end_index) {
1053                 page = find_get_page(tree->mapping, index);
1054                 BUG_ON(!page);
1055                 set_page_writeback(page);
1056                 page_cache_release(page);
1057                 index++;
1058         }
1059         return 0;
1060 }
1061
1062 /*
1063  * find the first offset in the io tree with 'bits' set. zero is
1064  * returned if we find something, and *start_ret and *end_ret are
1065  * set to reflect the state struct that was found.
1066  *
1067  * If nothing was found, 1 is returned, < 0 on error
1068  */
1069 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1070                           u64 *start_ret, u64 *end_ret, int bits)
1071 {
1072         struct rb_node *node;
1073         struct extent_state *state;
1074         int ret = 1;
1075
1076         spin_lock(&tree->lock);
1077         /*
1078          * this search will find all the extents that end after
1079          * our range starts.
1080          */
1081         node = tree_search(tree, start);
1082         if (!node)
1083                 goto out;
1084
1085         while (1) {
1086                 state = rb_entry(node, struct extent_state, rb_node);
1087                 if (state->end >= start && (state->state & bits)) {
1088                         *start_ret = state->start;
1089                         *end_ret = state->end;
1090                         ret = 0;
1091                         break;
1092                 }
1093                 node = rb_next(node);
1094                 if (!node)
1095                         break;
1096         }
1097 out:
1098         spin_unlock(&tree->lock);
1099         return ret;
1100 }
1101
1102 /* find the first state struct with 'bits' set after 'start', and
1103  * return it.  tree->lock must be held.  NULL will returned if
1104  * nothing was found after 'start'
1105  */
1106 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1107                                                  u64 start, int bits)
1108 {
1109         struct rb_node *node;
1110         struct extent_state *state;
1111
1112         /*
1113          * this search will find all the extents that end after
1114          * our range starts.
1115          */
1116         node = tree_search(tree, start);
1117         if (!node)
1118                 goto out;
1119
1120         while (1) {
1121                 state = rb_entry(node, struct extent_state, rb_node);
1122                 if (state->end >= start && (state->state & bits))
1123                         return state;
1124
1125                 node = rb_next(node);
1126                 if (!node)
1127                         break;
1128         }
1129 out:
1130         return NULL;
1131 }
1132
1133 /*
1134  * find a contiguous range of bytes in the file marked as delalloc, not
1135  * more than 'max_bytes'.  start and end are used to return the range,
1136  *
1137  * 1 is returned if we find something, 0 if nothing was in the tree
1138  */
1139 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1140                                         u64 *start, u64 *end, u64 max_bytes,
1141                                         struct extent_state **cached_state)
1142 {
1143         struct rb_node *node;
1144         struct extent_state *state;
1145         u64 cur_start = *start;
1146         u64 found = 0;
1147         u64 total_bytes = 0;
1148
1149         spin_lock(&tree->lock);
1150
1151         /*
1152          * this search will find all the extents that end after
1153          * our range starts.
1154          */
1155         node = tree_search(tree, cur_start);
1156         if (!node) {
1157                 if (!found)
1158                         *end = (u64)-1;
1159                 goto out;
1160         }
1161
1162         while (1) {
1163                 state = rb_entry(node, struct extent_state, rb_node);
1164                 if (found && (state->start != cur_start ||
1165                               (state->state & EXTENT_BOUNDARY))) {
1166                         goto out;
1167                 }
1168                 if (!(state->state & EXTENT_DELALLOC)) {
1169                         if (!found)
1170                                 *end = state->end;
1171                         goto out;
1172                 }
1173                 if (!found) {
1174                         *start = state->start;
1175                         *cached_state = state;
1176                         atomic_inc(&state->refs);
1177                 }
1178                 found++;
1179                 *end = state->end;
1180                 cur_start = state->end + 1;
1181                 node = rb_next(node);
1182                 if (!node)
1183                         break;
1184                 total_bytes += state->end - state->start + 1;
1185                 if (total_bytes >= max_bytes)
1186                         break;
1187         }
1188 out:
1189         spin_unlock(&tree->lock);
1190         return found;
1191 }
1192
1193 static noinline int __unlock_for_delalloc(struct inode *inode,
1194                                           struct page *locked_page,
1195                                           u64 start, u64 end)
1196 {
1197         int ret;
1198         struct page *pages[16];
1199         unsigned long index = start >> PAGE_CACHE_SHIFT;
1200         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1201         unsigned long nr_pages = end_index - index + 1;
1202         int i;
1203
1204         if (index == locked_page->index && end_index == index)
1205                 return 0;
1206
1207         while (nr_pages > 0) {
1208                 ret = find_get_pages_contig(inode->i_mapping, index,
1209                                      min_t(unsigned long, nr_pages,
1210                                      ARRAY_SIZE(pages)), pages);
1211                 for (i = 0; i < ret; i++) {
1212                         if (pages[i] != locked_page)
1213                                 unlock_page(pages[i]);
1214                         page_cache_release(pages[i]);
1215                 }
1216                 nr_pages -= ret;
1217                 index += ret;
1218                 cond_resched();
1219         }
1220         return 0;
1221 }
1222
1223 static noinline int lock_delalloc_pages(struct inode *inode,
1224                                         struct page *locked_page,
1225                                         u64 delalloc_start,
1226                                         u64 delalloc_end)
1227 {
1228         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1229         unsigned long start_index = index;
1230         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1231         unsigned long pages_locked = 0;
1232         struct page *pages[16];
1233         unsigned long nrpages;
1234         int ret;
1235         int i;
1236
1237         /* the caller is responsible for locking the start index */
1238         if (index == locked_page->index && index == end_index)
1239                 return 0;
1240
1241         /* skip the page at the start index */
1242         nrpages = end_index - index + 1;
1243         while (nrpages > 0) {
1244                 ret = find_get_pages_contig(inode->i_mapping, index,
1245                                      min_t(unsigned long,
1246                                      nrpages, ARRAY_SIZE(pages)), pages);
1247                 if (ret == 0) {
1248                         ret = -EAGAIN;
1249                         goto done;
1250                 }
1251                 /* now we have an array of pages, lock them all */
1252                 for (i = 0; i < ret; i++) {
1253                         /*
1254                          * the caller is taking responsibility for
1255                          * locked_page
1256                          */
1257                         if (pages[i] != locked_page) {
1258                                 lock_page(pages[i]);
1259                                 if (!PageDirty(pages[i]) ||
1260                                     pages[i]->mapping != inode->i_mapping) {
1261                                         ret = -EAGAIN;
1262                                         unlock_page(pages[i]);
1263                                         page_cache_release(pages[i]);
1264                                         goto done;
1265                                 }
1266                         }
1267                         page_cache_release(pages[i]);
1268                         pages_locked++;
1269                 }
1270                 nrpages -= ret;
1271                 index += ret;
1272                 cond_resched();
1273         }
1274         ret = 0;
1275 done:
1276         if (ret && pages_locked) {
1277                 __unlock_for_delalloc(inode, locked_page,
1278                               delalloc_start,
1279                               ((u64)(start_index + pages_locked - 1)) <<
1280                               PAGE_CACHE_SHIFT);
1281         }
1282         return ret;
1283 }
1284
1285 /*
1286  * find a contiguous range of bytes in the file marked as delalloc, not
1287  * more than 'max_bytes'.  start and end are used to return the range,
1288  *
1289  * 1 is returned if we find something, 0 if nothing was in the tree
1290  */
1291 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1292                                              struct extent_io_tree *tree,
1293                                              struct page *locked_page,
1294                                              u64 *start, u64 *end,
1295                                              u64 max_bytes)
1296 {
1297         u64 delalloc_start;
1298         u64 delalloc_end;
1299         u64 found;
1300         struct extent_state *cached_state = NULL;
1301         int ret;
1302         int loops = 0;
1303
1304 again:
1305         /* step one, find a bunch of delalloc bytes starting at start */
1306         delalloc_start = *start;
1307         delalloc_end = 0;
1308         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1309                                     max_bytes, &cached_state);
1310         if (!found || delalloc_end <= *start) {
1311                 *start = delalloc_start;
1312                 *end = delalloc_end;
1313                 free_extent_state(cached_state);
1314                 return found;
1315         }
1316
1317         /*
1318          * start comes from the offset of locked_page.  We have to lock
1319          * pages in order, so we can't process delalloc bytes before
1320          * locked_page
1321          */
1322         if (delalloc_start < *start)
1323                 delalloc_start = *start;
1324
1325         /*
1326          * make sure to limit the number of pages we try to lock down
1327          * if we're looping.
1328          */
1329         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1330                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1331
1332         /* step two, lock all the pages after the page that has start */
1333         ret = lock_delalloc_pages(inode, locked_page,
1334                                   delalloc_start, delalloc_end);
1335         if (ret == -EAGAIN) {
1336                 /* some of the pages are gone, lets avoid looping by
1337                  * shortening the size of the delalloc range we're searching
1338                  */
1339                 free_extent_state(cached_state);
1340                 if (!loops) {
1341                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1342                         max_bytes = PAGE_CACHE_SIZE - offset;
1343                         loops = 1;
1344                         goto again;
1345                 } else {
1346                         found = 0;
1347                         goto out_failed;
1348                 }
1349         }
1350         BUG_ON(ret);
1351
1352         /* step three, lock the state bits for the whole range */
1353         lock_extent_bits(tree, delalloc_start, delalloc_end,
1354                          0, &cached_state, GFP_NOFS);
1355
1356         /* then test to make sure it is all still delalloc */
1357         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1358                              EXTENT_DELALLOC, 1, cached_state);
1359         if (!ret) {
1360                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1361                                      &cached_state, GFP_NOFS);
1362                 __unlock_for_delalloc(inode, locked_page,
1363                               delalloc_start, delalloc_end);
1364                 cond_resched();
1365                 goto again;
1366         }
1367         free_extent_state(cached_state);
1368         *start = delalloc_start;
1369         *end = delalloc_end;
1370 out_failed:
1371         return found;
1372 }
1373
1374 int extent_clear_unlock_delalloc(struct inode *inode,
1375                                 struct extent_io_tree *tree,
1376                                 u64 start, u64 end, struct page *locked_page,
1377                                 unsigned long op)
1378 {
1379         int ret;
1380         struct page *pages[16];
1381         unsigned long index = start >> PAGE_CACHE_SHIFT;
1382         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1383         unsigned long nr_pages = end_index - index + 1;
1384         int i;
1385         int clear_bits = 0;
1386
1387         if (op & EXTENT_CLEAR_UNLOCK)
1388                 clear_bits |= EXTENT_LOCKED;
1389         if (op & EXTENT_CLEAR_DIRTY)
1390                 clear_bits |= EXTENT_DIRTY;
1391
1392         if (op & EXTENT_CLEAR_DELALLOC)
1393                 clear_bits |= EXTENT_DELALLOC;
1394
1395         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1396         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1397                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1398                     EXTENT_SET_PRIVATE2)))
1399                 return 0;
1400
1401         while (nr_pages > 0) {
1402                 ret = find_get_pages_contig(inode->i_mapping, index,
1403                                      min_t(unsigned long,
1404                                      nr_pages, ARRAY_SIZE(pages)), pages);
1405                 for (i = 0; i < ret; i++) {
1406
1407                         if (op & EXTENT_SET_PRIVATE2)
1408                                 SetPagePrivate2(pages[i]);
1409
1410                         if (pages[i] == locked_page) {
1411                                 page_cache_release(pages[i]);
1412                                 continue;
1413                         }
1414                         if (op & EXTENT_CLEAR_DIRTY)
1415                                 clear_page_dirty_for_io(pages[i]);
1416                         if (op & EXTENT_SET_WRITEBACK)
1417                                 set_page_writeback(pages[i]);
1418                         if (op & EXTENT_END_WRITEBACK)
1419                                 end_page_writeback(pages[i]);
1420                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1421                                 unlock_page(pages[i]);
1422                         page_cache_release(pages[i]);
1423                 }
1424                 nr_pages -= ret;
1425                 index += ret;
1426                 cond_resched();
1427         }
1428         return 0;
1429 }
1430
1431 /*
1432  * count the number of bytes in the tree that have a given bit(s)
1433  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1434  * cached.  The total number found is returned.
1435  */
1436 u64 count_range_bits(struct extent_io_tree *tree,
1437                      u64 *start, u64 search_end, u64 max_bytes,
1438                      unsigned long bits, int contig)
1439 {
1440         struct rb_node *node;
1441         struct extent_state *state;
1442         u64 cur_start = *start;
1443         u64 total_bytes = 0;
1444         u64 last = 0;
1445         int found = 0;
1446
1447         if (search_end <= cur_start) {
1448                 WARN_ON(1);
1449                 return 0;
1450         }
1451
1452         spin_lock(&tree->lock);
1453         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1454                 total_bytes = tree->dirty_bytes;
1455                 goto out;
1456         }
1457         /*
1458          * this search will find all the extents that end after
1459          * our range starts.
1460          */
1461         node = tree_search(tree, cur_start);
1462         if (!node)
1463                 goto out;
1464
1465         while (1) {
1466                 state = rb_entry(node, struct extent_state, rb_node);
1467                 if (state->start > search_end)
1468                         break;
1469                 if (contig && found && state->start > last + 1)
1470                         break;
1471                 if (state->end >= cur_start && (state->state & bits) == bits) {
1472                         total_bytes += min(search_end, state->end) + 1 -
1473                                        max(cur_start, state->start);
1474                         if (total_bytes >= max_bytes)
1475                                 break;
1476                         if (!found) {
1477                                 *start = max(cur_start, state->start);
1478                                 found = 1;
1479                         }
1480                         last = state->end;
1481                 } else if (contig && found) {
1482                         break;
1483                 }
1484                 node = rb_next(node);
1485                 if (!node)
1486                         break;
1487         }
1488 out:
1489         spin_unlock(&tree->lock);
1490         return total_bytes;
1491 }
1492
1493 /*
1494  * set the private field for a given byte offset in the tree.  If there isn't
1495  * an extent_state there already, this does nothing.
1496  */
1497 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1498 {
1499         struct rb_node *node;
1500         struct extent_state *state;
1501         int ret = 0;
1502
1503         spin_lock(&tree->lock);
1504         /*
1505          * this search will find all the extents that end after
1506          * our range starts.
1507          */
1508         node = tree_search(tree, start);
1509         if (!node) {
1510                 ret = -ENOENT;
1511                 goto out;
1512         }
1513         state = rb_entry(node, struct extent_state, rb_node);
1514         if (state->start != start) {
1515                 ret = -ENOENT;
1516                 goto out;
1517         }
1518         state->private = private;
1519 out:
1520         spin_unlock(&tree->lock);
1521         return ret;
1522 }
1523
1524 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1525 {
1526         struct rb_node *node;
1527         struct extent_state *state;
1528         int ret = 0;
1529
1530         spin_lock(&tree->lock);
1531         /*
1532          * this search will find all the extents that end after
1533          * our range starts.
1534          */
1535         node = tree_search(tree, start);
1536         if (!node) {
1537                 ret = -ENOENT;
1538                 goto out;
1539         }
1540         state = rb_entry(node, struct extent_state, rb_node);
1541         if (state->start != start) {
1542                 ret = -ENOENT;
1543                 goto out;
1544         }
1545         *private = state->private;
1546 out:
1547         spin_unlock(&tree->lock);
1548         return ret;
1549 }
1550
1551 /*
1552  * searches a range in the state tree for a given mask.
1553  * If 'filled' == 1, this returns 1 only if every extent in the tree
1554  * has the bits set.  Otherwise, 1 is returned if any bit in the
1555  * range is found set.
1556  */
1557 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1558                    int bits, int filled, struct extent_state *cached)
1559 {
1560         struct extent_state *state = NULL;
1561         struct rb_node *node;
1562         int bitset = 0;
1563
1564         spin_lock(&tree->lock);
1565         if (cached && cached->tree && cached->start <= start &&
1566             cached->end > start)
1567                 node = &cached->rb_node;
1568         else
1569                 node = tree_search(tree, start);
1570         while (node && start <= end) {
1571                 state = rb_entry(node, struct extent_state, rb_node);
1572
1573                 if (filled && state->start > start) {
1574                         bitset = 0;
1575                         break;
1576                 }
1577
1578                 if (state->start > end)
1579                         break;
1580
1581                 if (state->state & bits) {
1582                         bitset = 1;
1583                         if (!filled)
1584                                 break;
1585                 } else if (filled) {
1586                         bitset = 0;
1587                         break;
1588                 }
1589
1590                 if (state->end == (u64)-1)
1591                         break;
1592
1593                 start = state->end + 1;
1594                 if (start > end)
1595                         break;
1596                 node = rb_next(node);
1597                 if (!node) {
1598                         if (filled)
1599                                 bitset = 0;
1600                         break;
1601                 }
1602         }
1603         spin_unlock(&tree->lock);
1604         return bitset;
1605 }
1606
1607 /*
1608  * helper function to set a given page up to date if all the
1609  * extents in the tree for that page are up to date
1610  */
1611 static int check_page_uptodate(struct extent_io_tree *tree,
1612                                struct page *page)
1613 {
1614         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1615         u64 end = start + PAGE_CACHE_SIZE - 1;
1616         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1617                 SetPageUptodate(page);
1618         return 0;
1619 }
1620
1621 /*
1622  * helper function to unlock a page if all the extents in the tree
1623  * for that page are unlocked
1624  */
1625 static int check_page_locked(struct extent_io_tree *tree,
1626                              struct page *page)
1627 {
1628         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1629         u64 end = start + PAGE_CACHE_SIZE - 1;
1630         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1631                 unlock_page(page);
1632         return 0;
1633 }
1634
1635 /*
1636  * helper function to end page writeback if all the extents
1637  * in the tree for that page are done with writeback
1638  */
1639 static int check_page_writeback(struct extent_io_tree *tree,
1640                              struct page *page)
1641 {
1642         end_page_writeback(page);
1643         return 0;
1644 }
1645
1646 /* lots and lots of room for performance fixes in the end_bio funcs */
1647
1648 /*
1649  * after a writepage IO is done, we need to:
1650  * clear the uptodate bits on error
1651  * clear the writeback bits in the extent tree for this IO
1652  * end_page_writeback if the page has no more pending IO
1653  *
1654  * Scheduling is not allowed, so the extent state tree is expected
1655  * to have one and only one object corresponding to this IO.
1656  */
1657 static void end_bio_extent_writepage(struct bio *bio, int err)
1658 {
1659         int uptodate = err == 0;
1660         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1661         struct extent_io_tree *tree;
1662         u64 start;
1663         u64 end;
1664         int whole_page;
1665         int ret;
1666
1667         do {
1668                 struct page *page = bvec->bv_page;
1669                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1670
1671                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1672                          bvec->bv_offset;
1673                 end = start + bvec->bv_len - 1;
1674
1675                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1676                         whole_page = 1;
1677                 else
1678                         whole_page = 0;
1679
1680                 if (--bvec >= bio->bi_io_vec)
1681                         prefetchw(&bvec->bv_page->flags);
1682                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1683                         ret = tree->ops->writepage_end_io_hook(page, start,
1684                                                        end, NULL, uptodate);
1685                         if (ret)
1686                                 uptodate = 0;
1687                 }
1688
1689                 if (!uptodate && tree->ops &&
1690                     tree->ops->writepage_io_failed_hook) {
1691                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1692                                                          start, end, NULL);
1693                         if (ret == 0) {
1694                                 uptodate = (err == 0);
1695                                 continue;
1696                         }
1697                 }
1698
1699                 if (!uptodate) {
1700                         clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
1701                         ClearPageUptodate(page);
1702                         SetPageError(page);
1703                 }
1704
1705                 if (whole_page)
1706                         end_page_writeback(page);
1707                 else
1708                         check_page_writeback(tree, page);
1709         } while (bvec >= bio->bi_io_vec);
1710
1711         bio_put(bio);
1712 }
1713
1714 /*
1715  * after a readpage IO is done, we need to:
1716  * clear the uptodate bits on error
1717  * set the uptodate bits if things worked
1718  * set the page up to date if all extents in the tree are uptodate
1719  * clear the lock bit in the extent tree
1720  * unlock the page if there are no other extents locked for it
1721  *
1722  * Scheduling is not allowed, so the extent state tree is expected
1723  * to have one and only one object corresponding to this IO.
1724  */
1725 static void end_bio_extent_readpage(struct bio *bio, int err)
1726 {
1727         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1728         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
1729         struct bio_vec *bvec = bio->bi_io_vec;
1730         struct extent_io_tree *tree;
1731         u64 start;
1732         u64 end;
1733         int whole_page;
1734         int ret;
1735
1736         if (err)
1737                 uptodate = 0;
1738
1739         do {
1740                 struct page *page = bvec->bv_page;
1741                 struct extent_state *cached = NULL;
1742                 struct extent_state *state;
1743
1744                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1745
1746                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1747                         bvec->bv_offset;
1748                 end = start + bvec->bv_len - 1;
1749
1750                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1751                         whole_page = 1;
1752                 else
1753                         whole_page = 0;
1754
1755                 if (++bvec <= bvec_end)
1756                         prefetchw(&bvec->bv_page->flags);
1757
1758                 spin_lock(&tree->lock);
1759                 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
1760                 if (state && state->start == start) {
1761                         /*
1762                          * take a reference on the state, unlock will drop
1763                          * the ref
1764                          */
1765                         cache_state(state, &cached);
1766                 }
1767                 spin_unlock(&tree->lock);
1768
1769                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1770                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1771                                                               state);
1772                         if (ret)
1773                                 uptodate = 0;
1774                 }
1775                 if (!uptodate && tree->ops &&
1776                     tree->ops->readpage_io_failed_hook) {
1777                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1778                                                          start, end, NULL);
1779                         if (ret == 0) {
1780                                 uptodate =
1781                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1782                                 if (err)
1783                                         uptodate = 0;
1784                                 uncache_state(&cached);
1785                                 continue;
1786                         }
1787                 }
1788
1789                 if (uptodate) {
1790                         set_extent_uptodate(tree, start, end, &cached,
1791                                             GFP_ATOMIC);
1792                 }
1793                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
1794
1795                 if (whole_page) {
1796                         if (uptodate) {
1797                                 SetPageUptodate(page);
1798                         } else {
1799                                 ClearPageUptodate(page);
1800                                 SetPageError(page);
1801                         }
1802                         unlock_page(page);
1803                 } else {
1804                         if (uptodate) {
1805                                 check_page_uptodate(tree, page);
1806                         } else {
1807                                 ClearPageUptodate(page);
1808                                 SetPageError(page);
1809                         }
1810                         check_page_locked(tree, page);
1811                 }
1812         } while (bvec <= bvec_end);
1813
1814         bio_put(bio);
1815 }
1816
1817 struct bio *
1818 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1819                 gfp_t gfp_flags)
1820 {
1821         struct bio *bio;
1822
1823         bio = bio_alloc(gfp_flags, nr_vecs);
1824
1825         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1826                 while (!bio && (nr_vecs /= 2))
1827                         bio = bio_alloc(gfp_flags, nr_vecs);
1828         }
1829
1830         if (bio) {
1831                 bio->bi_size = 0;
1832                 bio->bi_bdev = bdev;
1833                 bio->bi_sector = first_sector;
1834         }
1835         return bio;
1836 }
1837
1838 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1839                           unsigned long bio_flags)
1840 {
1841         int ret = 0;
1842         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1843         struct page *page = bvec->bv_page;
1844         struct extent_io_tree *tree = bio->bi_private;
1845         u64 start;
1846
1847         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1848
1849         bio->bi_private = NULL;
1850
1851         bio_get(bio);
1852
1853         if (tree->ops && tree->ops->submit_bio_hook)
1854                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1855                                            mirror_num, bio_flags, start);
1856         else
1857                 submit_bio(rw, bio);
1858         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1859                 ret = -EOPNOTSUPP;
1860         bio_put(bio);
1861         return ret;
1862 }
1863
1864 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1865                               struct page *page, sector_t sector,
1866                               size_t size, unsigned long offset,
1867                               struct block_device *bdev,
1868                               struct bio **bio_ret,
1869                               unsigned long max_pages,
1870                               bio_end_io_t end_io_func,
1871                               int mirror_num,
1872                               unsigned long prev_bio_flags,
1873                               unsigned long bio_flags)
1874 {
1875         int ret = 0;
1876         struct bio *bio;
1877         int nr;
1878         int contig = 0;
1879         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1880         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1881         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1882
1883         if (bio_ret && *bio_ret) {
1884                 bio = *bio_ret;
1885                 if (old_compressed)
1886                         contig = bio->bi_sector == sector;
1887                 else
1888                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
1889                                 sector;
1890
1891                 if (prev_bio_flags != bio_flags || !contig ||
1892                     (tree->ops && tree->ops->merge_bio_hook &&
1893                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
1894                                                bio_flags)) ||
1895                     bio_add_page(bio, page, page_size, offset) < page_size) {
1896                         ret = submit_one_bio(rw, bio, mirror_num,
1897                                              prev_bio_flags);
1898                         bio = NULL;
1899                 } else {
1900                         return 0;
1901                 }
1902         }
1903         if (this_compressed)
1904                 nr = BIO_MAX_PAGES;
1905         else
1906                 nr = bio_get_nr_vecs(bdev);
1907
1908         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1909         if (!bio)
1910                 return -ENOMEM;
1911
1912         bio_add_page(bio, page, page_size, offset);
1913         bio->bi_end_io = end_io_func;
1914         bio->bi_private = tree;
1915
1916         if (bio_ret)
1917                 *bio_ret = bio;
1918         else
1919                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1920
1921         return ret;
1922 }
1923
1924 void set_page_extent_mapped(struct page *page)
1925 {
1926         if (!PagePrivate(page)) {
1927                 SetPagePrivate(page);
1928                 page_cache_get(page);
1929                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1930         }
1931 }
1932
1933 static void set_page_extent_head(struct page *page, unsigned long len)
1934 {
1935         WARN_ON(!PagePrivate(page));
1936         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1937 }
1938
1939 /*
1940  * basic readpage implementation.  Locked extent state structs are inserted
1941  * into the tree that are removed when the IO is done (by the end_io
1942  * handlers)
1943  */
1944 static int __extent_read_full_page(struct extent_io_tree *tree,
1945                                    struct page *page,
1946                                    get_extent_t *get_extent,
1947                                    struct bio **bio, int mirror_num,
1948                                    unsigned long *bio_flags)
1949 {
1950         struct inode *inode = page->mapping->host;
1951         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1952         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1953         u64 end;
1954         u64 cur = start;
1955         u64 extent_offset;
1956         u64 last_byte = i_size_read(inode);
1957         u64 block_start;
1958         u64 cur_end;
1959         sector_t sector;
1960         struct extent_map *em;
1961         struct block_device *bdev;
1962         struct btrfs_ordered_extent *ordered;
1963         int ret;
1964         int nr = 0;
1965         size_t pg_offset = 0;
1966         size_t iosize;
1967         size_t disk_io_size;
1968         size_t blocksize = inode->i_sb->s_blocksize;
1969         unsigned long this_bio_flag = 0;
1970
1971         set_page_extent_mapped(page);
1972
1973         if (!PageUptodate(page)) {
1974                 if (cleancache_get_page(page) == 0) {
1975                         BUG_ON(blocksize != PAGE_SIZE);
1976                         goto out;
1977                 }
1978         }
1979
1980         end = page_end;
1981         while (1) {
1982                 lock_extent(tree, start, end, GFP_NOFS);
1983                 ordered = btrfs_lookup_ordered_extent(inode, start);
1984                 if (!ordered)
1985                         break;
1986                 unlock_extent(tree, start, end, GFP_NOFS);
1987                 btrfs_start_ordered_extent(inode, ordered, 1);
1988                 btrfs_put_ordered_extent(ordered);
1989         }
1990
1991         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
1992                 char *userpage;
1993                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
1994
1995                 if (zero_offset) {
1996                         iosize = PAGE_CACHE_SIZE - zero_offset;
1997                         userpage = kmap_atomic(page, KM_USER0);
1998                         memset(userpage + zero_offset, 0, iosize);
1999                         flush_dcache_page(page);
2000                         kunmap_atomic(userpage, KM_USER0);
2001                 }
2002         }
2003         while (cur <= end) {
2004                 if (cur >= last_byte) {
2005                         char *userpage;
2006                         struct extent_state *cached = NULL;
2007
2008                         iosize = PAGE_CACHE_SIZE - pg_offset;
2009                         userpage = kmap_atomic(page, KM_USER0);
2010                         memset(userpage + pg_offset, 0, iosize);
2011                         flush_dcache_page(page);
2012                         kunmap_atomic(userpage, KM_USER0);
2013                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2014                                             &cached, GFP_NOFS);
2015                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2016                                              &cached, GFP_NOFS);
2017                         break;
2018                 }
2019                 em = get_extent(inode, page, pg_offset, cur,
2020                                 end - cur + 1, 0);
2021                 if (IS_ERR_OR_NULL(em)) {
2022                         SetPageError(page);
2023                         unlock_extent(tree, cur, end, GFP_NOFS);
2024                         break;
2025                 }
2026                 extent_offset = cur - em->start;
2027                 BUG_ON(extent_map_end(em) <= cur);
2028                 BUG_ON(end < cur);
2029
2030                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2031                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2032                         extent_set_compress_type(&this_bio_flag,
2033                                                  em->compress_type);
2034                 }
2035
2036                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2037                 cur_end = min(extent_map_end(em) - 1, end);
2038                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2039                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2040                         disk_io_size = em->block_len;
2041                         sector = em->block_start >> 9;
2042                 } else {
2043                         sector = (em->block_start + extent_offset) >> 9;
2044                         disk_io_size = iosize;
2045                 }
2046                 bdev = em->bdev;
2047                 block_start = em->block_start;
2048                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2049                         block_start = EXTENT_MAP_HOLE;
2050                 free_extent_map(em);
2051                 em = NULL;
2052
2053                 /* we've found a hole, just zero and go on */
2054                 if (block_start == EXTENT_MAP_HOLE) {
2055                         char *userpage;
2056                         struct extent_state *cached = NULL;
2057
2058                         userpage = kmap_atomic(page, KM_USER0);
2059                         memset(userpage + pg_offset, 0, iosize);
2060                         flush_dcache_page(page);
2061                         kunmap_atomic(userpage, KM_USER0);
2062
2063                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2064                                             &cached, GFP_NOFS);
2065                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2066                                              &cached, GFP_NOFS);
2067                         cur = cur + iosize;
2068                         pg_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                         pg_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                         pg_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, pg_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                 pg_offset += iosize;
2112         }
2113 out:
2114         if (!nr) {
2115                 if (!PageError(page))
2116                         SetPageUptodate(page);
2117                 unlock_page(page);
2118         }
2119         return 0;
2120 }
2121
2122 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2123                             get_extent_t *get_extent)
2124 {
2125         struct bio *bio = NULL;
2126         unsigned long bio_flags = 0;
2127         int ret;
2128
2129         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2130                                       &bio_flags);
2131         if (bio)
2132                 ret = submit_one_bio(READ, bio, 0, bio_flags);
2133         return ret;
2134 }
2135
2136 static noinline void update_nr_written(struct page *page,
2137                                       struct writeback_control *wbc,
2138                                       unsigned long nr_written)
2139 {
2140         wbc->nr_to_write -= nr_written;
2141         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2142             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2143                 page->mapping->writeback_index = page->index + nr_written;
2144 }
2145
2146 /*
2147  * the writepage semantics are similar to regular writepage.  extent
2148  * records are inserted to lock ranges in the tree, and as dirty areas
2149  * are found, they are marked writeback.  Then the lock bits are removed
2150  * and the end_io handler clears the writeback ranges
2151  */
2152 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2153                               void *data)
2154 {
2155         struct inode *inode = page->mapping->host;
2156         struct extent_page_data *epd = data;
2157         struct extent_io_tree *tree = epd->tree;
2158         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2159         u64 delalloc_start;
2160         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2161         u64 end;
2162         u64 cur = start;
2163         u64 extent_offset;
2164         u64 last_byte = i_size_read(inode);
2165         u64 block_start;
2166         u64 iosize;
2167         sector_t sector;
2168         struct extent_state *cached_state = NULL;
2169         struct extent_map *em;
2170         struct block_device *bdev;
2171         int ret;
2172         int nr = 0;
2173         size_t pg_offset = 0;
2174         size_t blocksize;
2175         loff_t i_size = i_size_read(inode);
2176         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2177         u64 nr_delalloc;
2178         u64 delalloc_end;
2179         int page_started;
2180         int compressed;
2181         int write_flags;
2182         unsigned long nr_written = 0;
2183
2184         if (wbc->sync_mode == WB_SYNC_ALL)
2185                 write_flags = WRITE_SYNC;
2186         else
2187                 write_flags = WRITE;
2188
2189         trace___extent_writepage(page, inode, wbc);
2190
2191         WARN_ON(!PageLocked(page));
2192         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2193         if (page->index > end_index ||
2194            (page->index == end_index && !pg_offset)) {
2195                 page->mapping->a_ops->invalidatepage(page, 0);
2196                 unlock_page(page);
2197                 return 0;
2198         }
2199
2200         if (page->index == end_index) {
2201                 char *userpage;
2202
2203                 userpage = kmap_atomic(page, KM_USER0);
2204                 memset(userpage + pg_offset, 0,
2205                        PAGE_CACHE_SIZE - pg_offset);
2206                 kunmap_atomic(userpage, KM_USER0);
2207                 flush_dcache_page(page);
2208         }
2209         pg_offset = 0;
2210
2211         set_page_extent_mapped(page);
2212
2213         delalloc_start = start;
2214         delalloc_end = 0;
2215         page_started = 0;
2216         if (!epd->extent_locked) {
2217                 u64 delalloc_to_write = 0;
2218                 /*
2219                  * make sure the wbc mapping index is at least updated
2220                  * to this page.
2221                  */
2222                 update_nr_written(page, wbc, 0);
2223
2224                 while (delalloc_end < page_end) {
2225                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2226                                                        page,
2227                                                        &delalloc_start,
2228                                                        &delalloc_end,
2229                                                        128 * 1024 * 1024);
2230                         if (nr_delalloc == 0) {
2231                                 delalloc_start = delalloc_end + 1;
2232                                 continue;
2233                         }
2234                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2235                                                  delalloc_end, &page_started,
2236                                                  &nr_written);
2237                         /*
2238                          * delalloc_end is already one less than the total
2239                          * length, so we don't subtract one from
2240                          * PAGE_CACHE_SIZE
2241                          */
2242                         delalloc_to_write += (delalloc_end - delalloc_start +
2243                                               PAGE_CACHE_SIZE) >>
2244                                               PAGE_CACHE_SHIFT;
2245                         delalloc_start = delalloc_end + 1;
2246                 }
2247                 if (wbc->nr_to_write < delalloc_to_write) {
2248                         int thresh = 8192;
2249
2250                         if (delalloc_to_write < thresh * 2)
2251                                 thresh = delalloc_to_write;
2252                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2253                                                  thresh);
2254                 }
2255
2256                 /* did the fill delalloc function already unlock and start
2257                  * the IO?
2258                  */
2259                 if (page_started) {
2260                         ret = 0;
2261                         /*
2262                          * we've unlocked the page, so we can't update
2263                          * the mapping's writeback index, just update
2264                          * nr_to_write.
2265                          */
2266                         wbc->nr_to_write -= nr_written;
2267                         goto done_unlocked;
2268                 }
2269         }
2270         if (tree->ops && tree->ops->writepage_start_hook) {
2271                 ret = tree->ops->writepage_start_hook(page, start,
2272                                                       page_end);
2273                 if (ret == -EAGAIN) {
2274                         redirty_page_for_writepage(wbc, page);
2275                         update_nr_written(page, wbc, nr_written);
2276                         unlock_page(page);
2277                         ret = 0;
2278                         goto done_unlocked;
2279                 }
2280         }
2281
2282         /*
2283          * we don't want to touch the inode after unlocking the page,
2284          * so we update the mapping writeback index now
2285          */
2286         update_nr_written(page, wbc, nr_written + 1);
2287
2288         end = page_end;
2289         if (last_byte <= start) {
2290                 if (tree->ops && tree->ops->writepage_end_io_hook)
2291                         tree->ops->writepage_end_io_hook(page, start,
2292                                                          page_end, NULL, 1);
2293                 goto done;
2294         }
2295
2296         blocksize = inode->i_sb->s_blocksize;
2297
2298         while (cur <= end) {
2299                 if (cur >= last_byte) {
2300                         if (tree->ops && tree->ops->writepage_end_io_hook)
2301                                 tree->ops->writepage_end_io_hook(page, cur,
2302                                                          page_end, NULL, 1);
2303                         break;
2304                 }
2305                 em = epd->get_extent(inode, page, pg_offset, cur,
2306                                      end - cur + 1, 1);
2307                 if (IS_ERR_OR_NULL(em)) {
2308                         SetPageError(page);
2309                         break;
2310                 }
2311
2312                 extent_offset = cur - em->start;
2313                 BUG_ON(extent_map_end(em) <= cur);
2314                 BUG_ON(end < cur);
2315                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2316                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2317                 sector = (em->block_start + extent_offset) >> 9;
2318                 bdev = em->bdev;
2319                 block_start = em->block_start;
2320                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2321                 free_extent_map(em);
2322                 em = NULL;
2323
2324                 /*
2325                  * compressed and inline extents are written through other
2326                  * paths in the FS
2327                  */
2328                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2329                     block_start == EXTENT_MAP_INLINE) {
2330                         /*
2331                          * end_io notification does not happen here for
2332                          * compressed extents
2333                          */
2334                         if (!compressed && tree->ops &&
2335                             tree->ops->writepage_end_io_hook)
2336                                 tree->ops->writepage_end_io_hook(page, cur,
2337                                                          cur + iosize - 1,
2338                                                          NULL, 1);
2339                         else if (compressed) {
2340                                 /* we don't want to end_page_writeback on
2341                                  * a compressed extent.  this happens
2342                                  * elsewhere
2343                                  */
2344                                 nr++;
2345                         }
2346
2347                         cur += iosize;
2348                         pg_offset += iosize;
2349                         continue;
2350                 }
2351                 /* leave this out until we have a page_mkwrite call */
2352                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2353                                    EXTENT_DIRTY, 0, NULL)) {
2354                         cur = cur + iosize;
2355                         pg_offset += iosize;
2356                         continue;
2357                 }
2358
2359                 if (tree->ops && tree->ops->writepage_io_hook) {
2360                         ret = tree->ops->writepage_io_hook(page, cur,
2361                                                 cur + iosize - 1);
2362                 } else {
2363                         ret = 0;
2364                 }
2365                 if (ret) {
2366                         SetPageError(page);
2367                 } else {
2368                         unsigned long max_nr = end_index + 1;
2369
2370                         set_range_writeback(tree, cur, cur + iosize - 1);
2371                         if (!PageWriteback(page)) {
2372                                 printk(KERN_ERR "btrfs warning page %lu not "
2373                                        "writeback, cur %llu end %llu\n",
2374                                        page->index, (unsigned long long)cur,
2375                                        (unsigned long long)end);
2376                         }
2377
2378                         ret = submit_extent_page(write_flags, tree, page,
2379                                                  sector, iosize, pg_offset,
2380                                                  bdev, &epd->bio, max_nr,
2381                                                  end_bio_extent_writepage,
2382                                                  0, 0, 0);
2383                         if (ret)
2384                                 SetPageError(page);
2385                 }
2386                 cur = cur + iosize;
2387                 pg_offset += iosize;
2388                 nr++;
2389         }
2390 done:
2391         if (nr == 0) {
2392                 /* make sure the mapping tag for page dirty gets cleared */
2393                 set_page_writeback(page);
2394                 end_page_writeback(page);
2395         }
2396         unlock_page(page);
2397
2398 done_unlocked:
2399
2400         /* drop our reference on any cached states */
2401         free_extent_state(cached_state);
2402         return 0;
2403 }
2404
2405 /**
2406  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2407  * @mapping: address space structure to write
2408  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2409  * @writepage: function called for each page
2410  * @data: data passed to writepage function
2411  *
2412  * If a page is already under I/O, write_cache_pages() skips it, even
2413  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2414  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2415  * and msync() need to guarantee that all the data which was dirty at the time
2416  * the call was made get new I/O started against them.  If wbc->sync_mode is
2417  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2418  * existing IO to complete.
2419  */
2420 static int extent_write_cache_pages(struct extent_io_tree *tree,
2421                              struct address_space *mapping,
2422                              struct writeback_control *wbc,
2423                              writepage_t writepage, void *data,
2424                              void (*flush_fn)(void *))
2425 {
2426         int ret = 0;
2427         int done = 0;
2428         int nr_to_write_done = 0;
2429         struct pagevec pvec;
2430         int nr_pages;
2431         pgoff_t index;
2432         pgoff_t end;            /* Inclusive */
2433         int scanned = 0;
2434         int tag;
2435
2436         pagevec_init(&pvec, 0);
2437         if (wbc->range_cyclic) {
2438                 index = mapping->writeback_index; /* Start from prev offset */
2439                 end = -1;
2440         } else {
2441                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2442                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2443                 scanned = 1;
2444         }
2445         if (wbc->sync_mode == WB_SYNC_ALL)
2446                 tag = PAGECACHE_TAG_TOWRITE;
2447         else
2448                 tag = PAGECACHE_TAG_DIRTY;
2449 retry:
2450         if (wbc->sync_mode == WB_SYNC_ALL)
2451                 tag_pages_for_writeback(mapping, index, end);
2452         while (!done && !nr_to_write_done && (index <= end) &&
2453                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2454                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2455                 unsigned i;
2456
2457                 scanned = 1;
2458                 for (i = 0; i < nr_pages; i++) {
2459                         struct page *page = pvec.pages[i];
2460
2461                         /*
2462                          * At this point we hold neither mapping->tree_lock nor
2463                          * lock on the page itself: the page may be truncated or
2464                          * invalidated (changing page->mapping to NULL), or even
2465                          * swizzled back from swapper_space to tmpfs file
2466                          * mapping
2467                          */
2468                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2469                                 tree->ops->write_cache_pages_lock_hook(page);
2470                         else
2471                                 lock_page(page);
2472
2473                         if (unlikely(page->mapping != mapping)) {
2474                                 unlock_page(page);
2475                                 continue;
2476                         }
2477
2478                         if (!wbc->range_cyclic && page->index > end) {
2479                                 done = 1;
2480                                 unlock_page(page);
2481                                 continue;
2482                         }
2483
2484                         if (wbc->sync_mode != WB_SYNC_NONE) {
2485                                 if (PageWriteback(page))
2486                                         flush_fn(data);
2487                                 wait_on_page_writeback(page);
2488                         }
2489
2490                         if (PageWriteback(page) ||
2491                             !clear_page_dirty_for_io(page)) {
2492                                 unlock_page(page);
2493                                 continue;
2494                         }
2495
2496                         ret = (*writepage)(page, wbc, data);
2497
2498                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2499                                 unlock_page(page);
2500                                 ret = 0;
2501                         }
2502                         if (ret)
2503                                 done = 1;
2504
2505                         /*
2506                          * the filesystem may choose to bump up nr_to_write.
2507                          * We have to make sure to honor the new nr_to_write
2508                          * at any time
2509                          */
2510                         nr_to_write_done = wbc->nr_to_write <= 0;
2511                 }
2512                 pagevec_release(&pvec);
2513                 cond_resched();
2514         }
2515         if (!scanned && !done) {
2516                 /*
2517                  * We hit the last page and there is more work to be done: wrap
2518                  * back to the start of the file
2519                  */
2520                 scanned = 1;
2521                 index = 0;
2522                 goto retry;
2523         }
2524         return ret;
2525 }
2526
2527 static void flush_epd_write_bio(struct extent_page_data *epd)
2528 {
2529         if (epd->bio) {
2530                 if (epd->sync_io)
2531                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2532                 else
2533                         submit_one_bio(WRITE, epd->bio, 0, 0);
2534                 epd->bio = NULL;
2535         }
2536 }
2537
2538 static noinline void flush_write_bio(void *data)
2539 {
2540         struct extent_page_data *epd = data;
2541         flush_epd_write_bio(epd);
2542 }
2543
2544 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2545                           get_extent_t *get_extent,
2546                           struct writeback_control *wbc)
2547 {
2548         int ret;
2549         struct address_space *mapping = page->mapping;
2550         struct extent_page_data epd = {
2551                 .bio = NULL,
2552                 .tree = tree,
2553                 .get_extent = get_extent,
2554                 .extent_locked = 0,
2555                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2556         };
2557         struct writeback_control wbc_writepages = {
2558                 .sync_mode      = wbc->sync_mode,
2559                 .nr_to_write    = 64,
2560                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2561                 .range_end      = (loff_t)-1,
2562         };
2563
2564         ret = __extent_writepage(page, wbc, &epd);
2565
2566         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2567                                  __extent_writepage, &epd, flush_write_bio);
2568         flush_epd_write_bio(&epd);
2569         return ret;
2570 }
2571
2572 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2573                               u64 start, u64 end, get_extent_t *get_extent,
2574                               int mode)
2575 {
2576         int ret = 0;
2577         struct address_space *mapping = inode->i_mapping;
2578         struct page *page;
2579         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2580                 PAGE_CACHE_SHIFT;
2581
2582         struct extent_page_data epd = {
2583                 .bio = NULL,
2584                 .tree = tree,
2585                 .get_extent = get_extent,
2586                 .extent_locked = 1,
2587                 .sync_io = mode == WB_SYNC_ALL,
2588         };
2589         struct writeback_control wbc_writepages = {
2590                 .sync_mode      = mode,
2591                 .nr_to_write    = nr_pages * 2,
2592                 .range_start    = start,
2593                 .range_end      = end + 1,
2594         };
2595
2596         while (start <= end) {
2597                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2598                 if (clear_page_dirty_for_io(page))
2599                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2600                 else {
2601                         if (tree->ops && tree->ops->writepage_end_io_hook)
2602                                 tree->ops->writepage_end_io_hook(page, start,
2603                                                  start + PAGE_CACHE_SIZE - 1,
2604                                                  NULL, 1);
2605                         unlock_page(page);
2606                 }
2607                 page_cache_release(page);
2608                 start += PAGE_CACHE_SIZE;
2609         }
2610
2611         flush_epd_write_bio(&epd);
2612         return ret;
2613 }
2614
2615 int extent_writepages(struct extent_io_tree *tree,
2616                       struct address_space *mapping,
2617                       get_extent_t *get_extent,
2618                       struct writeback_control *wbc)
2619 {
2620         int ret = 0;
2621         struct extent_page_data epd = {
2622                 .bio = NULL,
2623                 .tree = tree,
2624                 .get_extent = get_extent,
2625                 .extent_locked = 0,
2626                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2627         };
2628
2629         ret = extent_write_cache_pages(tree, mapping, wbc,
2630                                        __extent_writepage, &epd,
2631                                        flush_write_bio);
2632         flush_epd_write_bio(&epd);
2633         return ret;
2634 }
2635
2636 int extent_readpages(struct extent_io_tree *tree,
2637                      struct address_space *mapping,
2638                      struct list_head *pages, unsigned nr_pages,
2639                      get_extent_t get_extent)
2640 {
2641         struct bio *bio = NULL;
2642         unsigned page_idx;
2643         unsigned long bio_flags = 0;
2644
2645         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2646                 struct page *page = list_entry(pages->prev, struct page, lru);
2647
2648                 prefetchw(&page->flags);
2649                 list_del(&page->lru);
2650                 if (!add_to_page_cache_lru(page, mapping,
2651                                         page->index, GFP_NOFS)) {
2652                         __extent_read_full_page(tree, page, get_extent,
2653                                                 &bio, 0, &bio_flags);
2654                 }
2655                 page_cache_release(page);
2656         }
2657         BUG_ON(!list_empty(pages));
2658         if (bio)
2659                 submit_one_bio(READ, bio, 0, bio_flags);
2660         return 0;
2661 }
2662
2663 /*
2664  * basic invalidatepage code, this waits on any locked or writeback
2665  * ranges corresponding to the page, and then deletes any extent state
2666  * records from the tree
2667  */
2668 int extent_invalidatepage(struct extent_io_tree *tree,
2669                           struct page *page, unsigned long offset)
2670 {
2671         struct extent_state *cached_state = NULL;
2672         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2673         u64 end = start + PAGE_CACHE_SIZE - 1;
2674         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2675
2676         start += (offset + blocksize - 1) & ~(blocksize - 1);
2677         if (start > end)
2678                 return 0;
2679
2680         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
2681         wait_on_page_writeback(page);
2682         clear_extent_bit(tree, start, end,
2683                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2684                          EXTENT_DO_ACCOUNTING,
2685                          1, 1, &cached_state, GFP_NOFS);
2686         return 0;
2687 }
2688
2689 /*
2690  * a helper for releasepage, this tests for areas of the page that
2691  * are locked or under IO and drops the related state bits if it is safe
2692  * to drop the page.
2693  */
2694 int try_release_extent_state(struct extent_map_tree *map,
2695                              struct extent_io_tree *tree, struct page *page,
2696                              gfp_t mask)
2697 {
2698         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2699         u64 end = start + PAGE_CACHE_SIZE - 1;
2700         int ret = 1;
2701
2702         if (test_range_bit(tree, start, end,
2703                            EXTENT_IOBITS, 0, NULL))
2704                 ret = 0;
2705         else {
2706                 if ((mask & GFP_NOFS) == GFP_NOFS)
2707                         mask = GFP_NOFS;
2708                 /*
2709                  * at this point we can safely clear everything except the
2710                  * locked bit and the nodatasum bit
2711                  */
2712                 ret = clear_extent_bit(tree, start, end,
2713                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2714                                  0, 0, NULL, mask);
2715
2716                 /* if clear_extent_bit failed for enomem reasons,
2717                  * we can't allow the release to continue.
2718                  */
2719                 if (ret < 0)
2720                         ret = 0;
2721                 else
2722                         ret = 1;
2723         }
2724         return ret;
2725 }
2726
2727 /*
2728  * a helper for releasepage.  As long as there are no locked extents
2729  * in the range corresponding to the page, both state records and extent
2730  * map records are removed
2731  */
2732 int try_release_extent_mapping(struct extent_map_tree *map,
2733                                struct extent_io_tree *tree, struct page *page,
2734                                gfp_t mask)
2735 {
2736         struct extent_map *em;
2737         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2738         u64 end = start + PAGE_CACHE_SIZE - 1;
2739
2740         if ((mask & __GFP_WAIT) &&
2741             page->mapping->host->i_size > 16 * 1024 * 1024) {
2742                 u64 len;
2743                 while (start <= end) {
2744                         len = end - start + 1;
2745                         write_lock(&map->lock);
2746                         em = lookup_extent_mapping(map, start, len);
2747                         if (IS_ERR_OR_NULL(em)) {
2748                                 write_unlock(&map->lock);
2749                                 break;
2750                         }
2751                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2752                             em->start != start) {
2753                                 write_unlock(&map->lock);
2754                                 free_extent_map(em);
2755                                 break;
2756                         }
2757                         if (!test_range_bit(tree, em->start,
2758                                             extent_map_end(em) - 1,
2759                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
2760                                             0, NULL)) {
2761                                 remove_extent_mapping(map, em);
2762                                 /* once for the rb tree */
2763                                 free_extent_map(em);
2764                         }
2765                         start = extent_map_end(em);
2766                         write_unlock(&map->lock);
2767
2768                         /* once for us */
2769                         free_extent_map(em);
2770                 }
2771         }
2772         return try_release_extent_state(map, tree, page, mask);
2773 }
2774
2775 /*
2776  * helper function for fiemap, which doesn't want to see any holes.
2777  * This maps until we find something past 'last'
2778  */
2779 static struct extent_map *get_extent_skip_holes(struct inode *inode,
2780                                                 u64 offset,
2781                                                 u64 last,
2782                                                 get_extent_t *get_extent)
2783 {
2784         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
2785         struct extent_map *em;
2786         u64 len;
2787
2788         if (offset >= last)
2789                 return NULL;
2790
2791         while(1) {
2792                 len = last - offset;
2793                 if (len == 0)
2794                         break;
2795                 len = (len + sectorsize - 1) & ~(sectorsize - 1);
2796                 em = get_extent(inode, NULL, 0, offset, len, 0);
2797                 if (IS_ERR_OR_NULL(em))
2798                         return em;
2799
2800                 /* if this isn't a hole return it */
2801                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
2802                     em->block_start != EXTENT_MAP_HOLE) {
2803                         return em;
2804                 }
2805
2806                 /* this is a hole, advance to the next extent */
2807                 offset = extent_map_end(em);
2808                 free_extent_map(em);
2809                 if (offset >= last)
2810                         break;
2811         }
2812         return NULL;
2813 }
2814
2815 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2816                 __u64 start, __u64 len, get_extent_t *get_extent)
2817 {
2818         int ret = 0;
2819         u64 off = start;
2820         u64 max = start + len;
2821         u32 flags = 0;
2822         u32 found_type;
2823         u64 last;
2824         u64 last_for_get_extent = 0;
2825         u64 disko = 0;
2826         u64 isize = i_size_read(inode);
2827         struct btrfs_key found_key;
2828         struct extent_map *em = NULL;
2829         struct extent_state *cached_state = NULL;
2830         struct btrfs_path *path;
2831         struct btrfs_file_extent_item *item;
2832         int end = 0;
2833         u64 em_start = 0;
2834         u64 em_len = 0;
2835         u64 em_end = 0;
2836         unsigned long emflags;
2837
2838         if (len == 0)
2839                 return -EINVAL;
2840
2841         path = btrfs_alloc_path();
2842         if (!path)
2843                 return -ENOMEM;
2844         path->leave_spinning = 1;
2845
2846         /*
2847          * lookup the last file extent.  We're not using i_size here
2848          * because there might be preallocation past i_size
2849          */
2850         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
2851                                        path, btrfs_ino(inode), -1, 0);
2852         if (ret < 0) {
2853                 btrfs_free_path(path);
2854                 return ret;
2855         }
2856         WARN_ON(!ret);
2857         path->slots[0]--;
2858         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2859                               struct btrfs_file_extent_item);
2860         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
2861         found_type = btrfs_key_type(&found_key);
2862
2863         /* No extents, but there might be delalloc bits */
2864         if (found_key.objectid != btrfs_ino(inode) ||
2865             found_type != BTRFS_EXTENT_DATA_KEY) {
2866                 /* have to trust i_size as the end */
2867                 last = (u64)-1;
2868                 last_for_get_extent = isize;
2869         } else {
2870                 /*
2871                  * remember the start of the last extent.  There are a
2872                  * bunch of different factors that go into the length of the
2873                  * extent, so its much less complex to remember where it started
2874                  */
2875                 last = found_key.offset;
2876                 last_for_get_extent = last + 1;
2877         }
2878         btrfs_free_path(path);
2879
2880         /*
2881          * we might have some extents allocated but more delalloc past those
2882          * extents.  so, we trust isize unless the start of the last extent is
2883          * beyond isize
2884          */
2885         if (last < isize) {
2886                 last = (u64)-1;
2887                 last_for_get_extent = isize;
2888         }
2889
2890         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
2891                          &cached_state, GFP_NOFS);
2892
2893         em = get_extent_skip_holes(inode, off, last_for_get_extent,
2894                                    get_extent);
2895         if (!em)
2896                 goto out;
2897         if (IS_ERR(em)) {
2898                 ret = PTR_ERR(em);
2899                 goto out;
2900         }
2901
2902         while (!end) {
2903                 u64 offset_in_extent;
2904
2905                 /* break if the extent we found is outside the range */
2906                 if (em->start >= max || extent_map_end(em) < off)
2907                         break;
2908
2909                 /*
2910                  * get_extent may return an extent that starts before our
2911                  * requested range.  We have to make sure the ranges
2912                  * we return to fiemap always move forward and don't
2913                  * overlap, so adjust the offsets here
2914                  */
2915                 em_start = max(em->start, off);
2916
2917                 /*
2918                  * record the offset from the start of the extent
2919                  * for adjusting the disk offset below
2920                  */
2921                 offset_in_extent = em_start - em->start;
2922                 em_end = extent_map_end(em);
2923                 em_len = em_end - em_start;
2924                 emflags = em->flags;
2925                 disko = 0;
2926                 flags = 0;
2927
2928                 /*
2929                  * bump off for our next call to get_extent
2930                  */
2931                 off = extent_map_end(em);
2932                 if (off >= max)
2933                         end = 1;
2934
2935                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
2936                         end = 1;
2937                         flags |= FIEMAP_EXTENT_LAST;
2938                 } else if (em->block_start == EXTENT_MAP_INLINE) {
2939                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
2940                                   FIEMAP_EXTENT_NOT_ALIGNED);
2941                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
2942                         flags |= (FIEMAP_EXTENT_DELALLOC |
2943                                   FIEMAP_EXTENT_UNKNOWN);
2944                 } else {
2945                         disko = em->block_start + offset_in_extent;
2946                 }
2947                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2948                         flags |= FIEMAP_EXTENT_ENCODED;
2949
2950                 free_extent_map(em);
2951                 em = NULL;
2952                 if ((em_start >= last) || em_len == (u64)-1 ||
2953                    (last == (u64)-1 && isize <= em_end)) {
2954                         flags |= FIEMAP_EXTENT_LAST;
2955                         end = 1;
2956                 }
2957
2958                 /* now scan forward to see if this is really the last extent. */
2959                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
2960                                            get_extent);
2961                 if (IS_ERR(em)) {
2962                         ret = PTR_ERR(em);
2963                         goto out;
2964                 }
2965                 if (!em) {
2966                         flags |= FIEMAP_EXTENT_LAST;
2967                         end = 1;
2968                 }
2969                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
2970                                               em_len, flags);
2971                 if (ret)
2972                         goto out_free;
2973         }
2974 out_free:
2975         free_extent_map(em);
2976 out:
2977         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
2978                              &cached_state, GFP_NOFS);
2979         return ret;
2980 }
2981
2982 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2983                                               unsigned long i)
2984 {
2985         struct page *p;
2986         struct address_space *mapping;
2987
2988         if (i == 0)
2989                 return eb->first_page;
2990         i += eb->start >> PAGE_CACHE_SHIFT;
2991         mapping = eb->first_page->mapping;
2992         if (!mapping)
2993                 return NULL;
2994
2995         /*
2996          * extent_buffer_page is only called after pinning the page
2997          * by increasing the reference count.  So we know the page must
2998          * be in the radix tree.
2999          */
3000         rcu_read_lock();
3001         p = radix_tree_lookup(&mapping->page_tree, i);
3002         rcu_read_unlock();
3003
3004         return p;
3005 }
3006
3007 static inline unsigned long num_extent_pages(u64 start, u64 len)
3008 {
3009         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3010                 (start >> PAGE_CACHE_SHIFT);
3011 }
3012
3013 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3014                                                    u64 start,
3015                                                    unsigned long len,
3016                                                    gfp_t mask)
3017 {
3018         struct extent_buffer *eb = NULL;
3019 #if LEAK_DEBUG
3020         unsigned long flags;
3021 #endif
3022
3023         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3024         if (eb == NULL)
3025                 return NULL;
3026         eb->start = start;
3027         eb->len = len;
3028         rwlock_init(&eb->lock);
3029         atomic_set(&eb->write_locks, 0);
3030         atomic_set(&eb->read_locks, 0);
3031         atomic_set(&eb->blocking_readers, 0);
3032         atomic_set(&eb->blocking_writers, 0);
3033         atomic_set(&eb->spinning_readers, 0);
3034         atomic_set(&eb->spinning_writers, 0);
3035         init_waitqueue_head(&eb->write_lock_wq);
3036         init_waitqueue_head(&eb->read_lock_wq);
3037
3038 #if LEAK_DEBUG
3039         spin_lock_irqsave(&leak_lock, flags);
3040         list_add(&eb->leak_list, &buffers);
3041         spin_unlock_irqrestore(&leak_lock, flags);
3042 #endif
3043         atomic_set(&eb->refs, 1);
3044
3045         return eb;
3046 }
3047
3048 static void __free_extent_buffer(struct extent_buffer *eb)
3049 {
3050 #if LEAK_DEBUG
3051         unsigned long flags;
3052         spin_lock_irqsave(&leak_lock, flags);
3053         list_del(&eb->leak_list);
3054         spin_unlock_irqrestore(&leak_lock, flags);
3055 #endif
3056         kmem_cache_free(extent_buffer_cache, eb);
3057 }
3058
3059 /*
3060  * Helper for releasing extent buffer page.
3061  */
3062 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3063                                                 unsigned long start_idx)
3064 {
3065         unsigned long index;
3066         struct page *page;
3067
3068         if (!eb->first_page)
3069                 return;
3070
3071         index = num_extent_pages(eb->start, eb->len);
3072         if (start_idx >= index)
3073                 return;
3074
3075         do {
3076                 index--;
3077                 page = extent_buffer_page(eb, index);
3078                 if (page)
3079                         page_cache_release(page);
3080         } while (index != start_idx);
3081 }
3082
3083 /*
3084  * Helper for releasing the extent buffer.
3085  */
3086 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3087 {
3088         btrfs_release_extent_buffer_page(eb, 0);
3089         __free_extent_buffer(eb);
3090 }
3091
3092 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3093                                           u64 start, unsigned long len,
3094                                           struct page *page0)
3095 {
3096         unsigned long num_pages = num_extent_pages(start, len);
3097         unsigned long i;
3098         unsigned long index = start >> PAGE_CACHE_SHIFT;
3099         struct extent_buffer *eb;
3100         struct extent_buffer *exists = NULL;
3101         struct page *p;
3102         struct address_space *mapping = tree->mapping;
3103         int uptodate = 1;
3104         int ret;
3105
3106         rcu_read_lock();
3107         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3108         if (eb && atomic_inc_not_zero(&eb->refs)) {
3109                 rcu_read_unlock();
3110                 mark_page_accessed(eb->first_page);
3111                 return eb;
3112         }
3113         rcu_read_unlock();
3114
3115         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
3116         if (!eb)
3117                 return NULL;
3118
3119         if (page0) {
3120                 eb->first_page = page0;
3121                 i = 1;
3122                 index++;
3123                 page_cache_get(page0);
3124                 mark_page_accessed(page0);
3125                 set_page_extent_mapped(page0);
3126                 set_page_extent_head(page0, len);
3127                 uptodate = PageUptodate(page0);
3128         } else {
3129                 i = 0;
3130         }
3131         for (; i < num_pages; i++, index++) {
3132                 p = find_or_create_page(mapping, index, GFP_NOFS);
3133                 if (!p) {
3134                         WARN_ON(1);
3135                         goto free_eb;
3136                 }
3137                 set_page_extent_mapped(p);
3138                 mark_page_accessed(p);
3139                 if (i == 0) {
3140                         eb->first_page = p;
3141                         set_page_extent_head(p, len);
3142                 } else {
3143                         set_page_private(p, EXTENT_PAGE_PRIVATE);
3144                 }
3145                 if (!PageUptodate(p))
3146                         uptodate = 0;
3147
3148                 /*
3149                  * see below about how we avoid a nasty race with release page
3150                  * and why we unlock later
3151                  */
3152                 if (i != 0)
3153                         unlock_page(p);
3154         }
3155         if (uptodate)
3156                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3157
3158         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
3159         if (ret)
3160                 goto free_eb;
3161
3162         spin_lock(&tree->buffer_lock);
3163         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
3164         if (ret == -EEXIST) {
3165                 exists = radix_tree_lookup(&tree->buffer,
3166                                                 start >> PAGE_CACHE_SHIFT);
3167                 /* add one reference for the caller */
3168                 atomic_inc(&exists->refs);
3169                 spin_unlock(&tree->buffer_lock);
3170                 radix_tree_preload_end();
3171                 goto free_eb;
3172         }
3173         /* add one reference for the tree */
3174         atomic_inc(&eb->refs);
3175         spin_unlock(&tree->buffer_lock);
3176         radix_tree_preload_end();
3177
3178         /*
3179          * there is a race where release page may have
3180          * tried to find this extent buffer in the radix
3181          * but failed.  It will tell the VM it is safe to
3182          * reclaim the, and it will clear the page private bit.
3183          * We must make sure to set the page private bit properly
3184          * after the extent buffer is in the radix tree so
3185          * it doesn't get lost
3186          */
3187         set_page_extent_mapped(eb->first_page);
3188         set_page_extent_head(eb->first_page, eb->len);
3189         if (!page0)
3190                 unlock_page(eb->first_page);
3191         return eb;
3192
3193 free_eb:
3194         if (eb->first_page && !page0)
3195                 unlock_page(eb->first_page);
3196
3197         if (!atomic_dec_and_test(&eb->refs))
3198                 return exists;
3199         btrfs_release_extent_buffer(eb);
3200         return exists;
3201 }
3202
3203 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3204                                          u64 start, unsigned long len)
3205 {
3206         struct extent_buffer *eb;
3207
3208         rcu_read_lock();
3209         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3210         if (eb && atomic_inc_not_zero(&eb->refs)) {
3211                 rcu_read_unlock();
3212                 mark_page_accessed(eb->first_page);
3213                 return eb;
3214         }
3215         rcu_read_unlock();
3216
3217         return NULL;
3218 }
3219
3220 void free_extent_buffer(struct extent_buffer *eb)
3221 {
3222         if (!eb)
3223                 return;
3224
3225         if (!atomic_dec_and_test(&eb->refs))
3226                 return;
3227
3228         WARN_ON(1);
3229 }
3230
3231 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3232                               struct extent_buffer *eb)
3233 {
3234         unsigned long i;
3235         unsigned long num_pages;
3236         struct page *page;
3237
3238         num_pages = num_extent_pages(eb->start, eb->len);
3239
3240         for (i = 0; i < num_pages; i++) {
3241                 page = extent_buffer_page(eb, i);
3242                 if (!PageDirty(page))
3243                         continue;
3244
3245                 lock_page(page);
3246                 WARN_ON(!PagePrivate(page));
3247
3248                 set_page_extent_mapped(page);
3249                 if (i == 0)
3250                         set_page_extent_head(page, eb->len);
3251
3252                 clear_page_dirty_for_io(page);
3253                 spin_lock_irq(&page->mapping->tree_lock);
3254                 if (!PageDirty(page)) {
3255                         radix_tree_tag_clear(&page->mapping->page_tree,
3256                                                 page_index(page),
3257                                                 PAGECACHE_TAG_DIRTY);
3258                 }
3259                 spin_unlock_irq(&page->mapping->tree_lock);
3260                 unlock_page(page);
3261         }
3262         return 0;
3263 }
3264
3265 int set_extent_buffer_dirty(struct extent_io_tree *tree,
3266                              struct extent_buffer *eb)
3267 {
3268         unsigned long i;
3269         unsigned long num_pages;
3270         int was_dirty = 0;
3271
3272         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3273         num_pages = num_extent_pages(eb->start, eb->len);
3274         for (i = 0; i < num_pages; i++)
3275                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3276         return was_dirty;
3277 }
3278
3279 static int __eb_straddles_pages(u64 start, u64 len)
3280 {
3281         if (len < PAGE_CACHE_SIZE)
3282                 return 1;
3283         if (start & (PAGE_CACHE_SIZE - 1))
3284                 return 1;
3285         if ((start + len) & (PAGE_CACHE_SIZE - 1))
3286                 return 1;
3287         return 0;
3288 }
3289
3290 static int eb_straddles_pages(struct extent_buffer *eb)
3291 {
3292         return __eb_straddles_pages(eb->start, eb->len);
3293 }
3294
3295 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3296                                 struct extent_buffer *eb,
3297                                 struct extent_state **cached_state)
3298 {
3299         unsigned long i;
3300         struct page *page;
3301         unsigned long num_pages;
3302
3303         num_pages = num_extent_pages(eb->start, eb->len);
3304         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3305
3306         if (eb_straddles_pages(eb)) {
3307                 clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3308                                       cached_state, GFP_NOFS);
3309         }
3310         for (i = 0; i < num_pages; i++) {
3311                 page = extent_buffer_page(eb, i);
3312                 if (page)
3313                         ClearPageUptodate(page);
3314         }
3315         return 0;
3316 }
3317
3318 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3319                                 struct extent_buffer *eb)
3320 {
3321         unsigned long i;
3322         struct page *page;
3323         unsigned long num_pages;
3324
3325         num_pages = num_extent_pages(eb->start, eb->len);
3326
3327         if (eb_straddles_pages(eb)) {
3328                 set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3329                                     NULL, GFP_NOFS);
3330         }
3331         for (i = 0; i < num_pages; i++) {
3332                 page = extent_buffer_page(eb, i);
3333                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3334                     ((i == num_pages - 1) &&
3335                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3336                         check_page_uptodate(tree, page);
3337                         continue;
3338                 }
3339                 SetPageUptodate(page);
3340         }
3341         return 0;
3342 }
3343
3344 int extent_range_uptodate(struct extent_io_tree *tree,
3345                           u64 start, u64 end)
3346 {
3347         struct page *page;
3348         int ret;
3349         int pg_uptodate = 1;
3350         int uptodate;
3351         unsigned long index;
3352
3353         if (__eb_straddles_pages(start, end - start + 1)) {
3354                 ret = test_range_bit(tree, start, end,
3355                                      EXTENT_UPTODATE, 1, NULL);
3356                 if (ret)
3357                         return 1;
3358         }
3359         while (start <= end) {
3360                 index = start >> PAGE_CACHE_SHIFT;
3361                 page = find_get_page(tree->mapping, index);
3362                 uptodate = PageUptodate(page);
3363                 page_cache_release(page);
3364                 if (!uptodate) {
3365                         pg_uptodate = 0;
3366                         break;
3367                 }
3368                 start += PAGE_CACHE_SIZE;
3369         }
3370         return pg_uptodate;
3371 }
3372
3373 int extent_buffer_uptodate(struct extent_io_tree *tree,
3374                            struct extent_buffer *eb,
3375                            struct extent_state *cached_state)
3376 {
3377         int ret = 0;
3378         unsigned long num_pages;
3379         unsigned long i;
3380         struct page *page;
3381         int pg_uptodate = 1;
3382
3383         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3384                 return 1;
3385
3386         if (eb_straddles_pages(eb)) {
3387                 ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3388                                    EXTENT_UPTODATE, 1, cached_state);
3389                 if (ret)
3390                         return ret;
3391         }
3392
3393         num_pages = num_extent_pages(eb->start, eb->len);
3394         for (i = 0; i < num_pages; i++) {
3395                 page = extent_buffer_page(eb, i);
3396                 if (!PageUptodate(page)) {
3397                         pg_uptodate = 0;
3398                         break;
3399                 }
3400         }
3401         return pg_uptodate;
3402 }
3403
3404 int read_extent_buffer_pages(struct extent_io_tree *tree,
3405                              struct extent_buffer *eb,
3406                              u64 start, int wait,
3407                              get_extent_t *get_extent, int mirror_num)
3408 {
3409         unsigned long i;
3410         unsigned long start_i;
3411         struct page *page;
3412         int err;
3413         int ret = 0;
3414         int locked_pages = 0;
3415         int all_uptodate = 1;
3416         int inc_all_pages = 0;
3417         unsigned long num_pages;
3418         struct bio *bio = NULL;
3419         unsigned long bio_flags = 0;
3420
3421         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3422                 return 0;
3423
3424         if (eb_straddles_pages(eb)) {
3425                 if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3426                                    EXTENT_UPTODATE, 1, NULL)) {
3427                         return 0;
3428                 }
3429         }
3430
3431         if (start) {
3432                 WARN_ON(start < eb->start);
3433                 start_i = (start >> PAGE_CACHE_SHIFT) -
3434                         (eb->start >> PAGE_CACHE_SHIFT);
3435         } else {
3436                 start_i = 0;
3437         }
3438
3439         num_pages = num_extent_pages(eb->start, eb->len);
3440         for (i = start_i; i < num_pages; i++) {
3441                 page = extent_buffer_page(eb, i);
3442                 if (!wait) {
3443                         if (!trylock_page(page))
3444                                 goto unlock_exit;
3445                 } else {
3446                         lock_page(page);
3447                 }
3448                 locked_pages++;
3449                 if (!PageUptodate(page))
3450                         all_uptodate = 0;
3451         }
3452         if (all_uptodate) {
3453                 if (start_i == 0)
3454                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3455                 goto unlock_exit;
3456         }
3457
3458         for (i = start_i; i < num_pages; i++) {
3459                 page = extent_buffer_page(eb, i);
3460
3461                 WARN_ON(!PagePrivate(page));
3462
3463                 set_page_extent_mapped(page);
3464                 if (i == 0)
3465                         set_page_extent_head(page, eb->len);
3466
3467                 if (inc_all_pages)
3468                         page_cache_get(page);
3469                 if (!PageUptodate(page)) {
3470                         if (start_i == 0)
3471                                 inc_all_pages = 1;
3472                         ClearPageError(page);
3473                         err = __extent_read_full_page(tree, page,
3474                                                       get_extent, &bio,
3475                                                       mirror_num, &bio_flags);
3476                         if (err)
3477                                 ret = err;
3478                 } else {
3479                         unlock_page(page);
3480                 }
3481         }
3482
3483         if (bio)
3484                 submit_one_bio(READ, bio, mirror_num, bio_flags);
3485
3486         if (ret || !wait)
3487                 return ret;
3488
3489         for (i = start_i; i < num_pages; i++) {
3490                 page = extent_buffer_page(eb, i);
3491                 wait_on_page_locked(page);
3492                 if (!PageUptodate(page))
3493                         ret = -EIO;
3494         }
3495
3496         if (!ret)
3497                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3498         return ret;
3499
3500 unlock_exit:
3501         i = start_i;
3502         while (locked_pages > 0) {
3503                 page = extent_buffer_page(eb, i);
3504                 i++;
3505                 unlock_page(page);
3506                 locked_pages--;
3507         }
3508         return ret;
3509 }
3510
3511 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3512                         unsigned long start,
3513                         unsigned long len)
3514 {
3515         size_t cur;
3516         size_t offset;
3517         struct page *page;
3518         char *kaddr;
3519         char *dst = (char *)dstv;
3520         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3521         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3522
3523         WARN_ON(start > eb->len);
3524         WARN_ON(start + len > eb->start + eb->len);
3525
3526         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3527
3528         while (len > 0) {
3529                 page = extent_buffer_page(eb, i);
3530
3531                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3532                 kaddr = page_address(page);
3533                 memcpy(dst, kaddr + offset, cur);
3534
3535                 dst += cur;
3536                 len -= cur;
3537                 offset = 0;
3538                 i++;
3539         }
3540 }
3541
3542 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3543                                unsigned long min_len, char **map,
3544                                unsigned long *map_start,
3545                                unsigned long *map_len)
3546 {
3547         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3548         char *kaddr;
3549         struct page *p;
3550         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3551         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3552         unsigned long end_i = (start_offset + start + min_len - 1) >>
3553                 PAGE_CACHE_SHIFT;
3554
3555         if (i != end_i)
3556                 return -EINVAL;
3557
3558         if (i == 0) {
3559                 offset = start_offset;
3560                 *map_start = 0;
3561         } else {
3562                 offset = 0;
3563                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3564         }
3565
3566         if (start + min_len > eb->len) {
3567                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3568                        "wanted %lu %lu\n", (unsigned long long)eb->start,
3569                        eb->len, start, min_len);
3570                 WARN_ON(1);
3571                 return -EINVAL;
3572         }
3573
3574         p = extent_buffer_page(eb, i);
3575         kaddr = page_address(p);
3576         *map = kaddr + offset;
3577         *map_len = PAGE_CACHE_SIZE - offset;
3578         return 0;
3579 }
3580
3581 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3582                           unsigned long start,
3583                           unsigned long len)
3584 {
3585         size_t cur;
3586         size_t offset;
3587         struct page *page;
3588         char *kaddr;
3589         char *ptr = (char *)ptrv;
3590         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3591         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3592         int ret = 0;
3593
3594         WARN_ON(start > eb->len);
3595         WARN_ON(start + len > eb->start + eb->len);
3596
3597         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3598
3599         while (len > 0) {
3600                 page = extent_buffer_page(eb, i);
3601
3602                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3603
3604                 kaddr = page_address(page);
3605                 ret = memcmp(ptr, kaddr + offset, cur);
3606                 if (ret)
3607                         break;
3608
3609                 ptr += cur;
3610                 len -= cur;
3611                 offset = 0;
3612                 i++;
3613         }
3614         return ret;
3615 }
3616
3617 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3618                          unsigned long start, unsigned long len)
3619 {
3620         size_t cur;
3621         size_t offset;
3622         struct page *page;
3623         char *kaddr;
3624         char *src = (char *)srcv;
3625         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3626         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3627
3628         WARN_ON(start > eb->len);
3629         WARN_ON(start + len > eb->start + eb->len);
3630
3631         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3632
3633         while (len > 0) {
3634                 page = extent_buffer_page(eb, i);
3635                 WARN_ON(!PageUptodate(page));
3636
3637                 cur = min(len, PAGE_CACHE_SIZE - offset);
3638                 kaddr = page_address(page);
3639                 memcpy(kaddr + offset, src, cur);
3640
3641                 src += cur;
3642                 len -= cur;
3643                 offset = 0;
3644                 i++;
3645         }
3646 }
3647
3648 void memset_extent_buffer(struct extent_buffer *eb, char c,
3649                           unsigned long start, unsigned long len)
3650 {
3651         size_t cur;
3652         size_t offset;
3653         struct page *page;
3654         char *kaddr;
3655         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3656         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3657
3658         WARN_ON(start > eb->len);
3659         WARN_ON(start + len > eb->start + eb->len);
3660
3661         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3662
3663         while (len > 0) {
3664                 page = extent_buffer_page(eb, i);
3665                 WARN_ON(!PageUptodate(page));
3666
3667                 cur = min(len, PAGE_CACHE_SIZE - offset);
3668                 kaddr = page_address(page);
3669                 memset(kaddr + offset, c, cur);
3670
3671                 len -= cur;
3672                 offset = 0;
3673                 i++;
3674         }
3675 }
3676
3677 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3678                         unsigned long dst_offset, unsigned long src_offset,
3679                         unsigned long len)
3680 {
3681         u64 dst_len = dst->len;
3682         size_t cur;
3683         size_t offset;
3684         struct page *page;
3685         char *kaddr;
3686         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3687         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3688
3689         WARN_ON(src->len != dst_len);
3690
3691         offset = (start_offset + dst_offset) &
3692                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3693
3694         while (len > 0) {
3695                 page = extent_buffer_page(dst, i);
3696                 WARN_ON(!PageUptodate(page));
3697
3698                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3699
3700                 kaddr = page_address(page);
3701                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3702
3703                 src_offset += cur;
3704                 len -= cur;
3705                 offset = 0;
3706                 i++;
3707         }
3708 }
3709
3710 static void move_pages(struct page *dst_page, struct page *src_page,
3711                        unsigned long dst_off, unsigned long src_off,
3712                        unsigned long len)
3713 {
3714         char *dst_kaddr = page_address(dst_page);
3715         if (dst_page == src_page) {
3716                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3717         } else {
3718                 char *src_kaddr = page_address(src_page);
3719                 char *p = dst_kaddr + dst_off + len;
3720                 char *s = src_kaddr + src_off + len;
3721
3722                 while (len--)
3723                         *--p = *--s;
3724         }
3725 }
3726
3727 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
3728 {
3729         unsigned long distance = (src > dst) ? src - dst : dst - src;
3730         return distance < len;
3731 }
3732
3733 static void copy_pages(struct page *dst_page, struct page *src_page,
3734                        unsigned long dst_off, unsigned long src_off,
3735                        unsigned long len)
3736 {
3737         char *dst_kaddr = page_address(dst_page);
3738         char *src_kaddr;
3739
3740         if (dst_page != src_page) {
3741                 src_kaddr = page_address(src_page);
3742         } else {
3743                 src_kaddr = dst_kaddr;
3744                 BUG_ON(areas_overlap(src_off, dst_off, len));
3745         }
3746
3747         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3748 }
3749
3750 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3751                            unsigned long src_offset, unsigned long len)
3752 {
3753         size_t cur;
3754         size_t dst_off_in_page;
3755         size_t src_off_in_page;
3756         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3757         unsigned long dst_i;
3758         unsigned long src_i;
3759
3760         if (src_offset + len > dst->len) {
3761                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3762                        "len %lu dst len %lu\n", src_offset, len, dst->len);
3763                 BUG_ON(1);
3764         }
3765         if (dst_offset + len > dst->len) {
3766                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3767                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
3768                 BUG_ON(1);
3769         }
3770
3771         while (len > 0) {
3772                 dst_off_in_page = (start_offset + dst_offset) &
3773                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3774                 src_off_in_page = (start_offset + src_offset) &
3775                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3776
3777                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3778                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3779
3780                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3781                                                src_off_in_page));
3782                 cur = min_t(unsigned long, cur,
3783                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3784
3785                 copy_pages(extent_buffer_page(dst, dst_i),
3786                            extent_buffer_page(dst, src_i),
3787                            dst_off_in_page, src_off_in_page, cur);
3788
3789                 src_offset += cur;
3790                 dst_offset += cur;
3791                 len -= cur;
3792         }
3793 }
3794
3795 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3796                            unsigned long src_offset, unsigned long len)
3797 {
3798         size_t cur;
3799         size_t dst_off_in_page;
3800         size_t src_off_in_page;
3801         unsigned long dst_end = dst_offset + len - 1;
3802         unsigned long src_end = src_offset + len - 1;
3803         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3804         unsigned long dst_i;
3805         unsigned long src_i;
3806
3807         if (src_offset + len > dst->len) {
3808                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3809                        "len %lu len %lu\n", src_offset, len, dst->len);
3810                 BUG_ON(1);
3811         }
3812         if (dst_offset + len > dst->len) {
3813                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3814                        "len %lu len %lu\n", dst_offset, len, dst->len);
3815                 BUG_ON(1);
3816         }
3817         if (!areas_overlap(src_offset, dst_offset, len)) {
3818                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3819                 return;
3820         }
3821         while (len > 0) {
3822                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3823                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3824
3825                 dst_off_in_page = (start_offset + dst_end) &
3826                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3827                 src_off_in_page = (start_offset + src_end) &
3828                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3829
3830                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3831                 cur = min(cur, dst_off_in_page + 1);
3832                 move_pages(extent_buffer_page(dst, dst_i),
3833                            extent_buffer_page(dst, src_i),
3834                            dst_off_in_page - cur + 1,
3835                            src_off_in_page - cur + 1, cur);
3836
3837                 dst_end -= cur;
3838                 src_end -= cur;
3839                 len -= cur;
3840         }
3841 }
3842
3843 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3844 {
3845         struct extent_buffer *eb =
3846                         container_of(head, struct extent_buffer, rcu_head);
3847
3848         btrfs_release_extent_buffer(eb);
3849 }
3850
3851 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3852 {
3853         u64 start = page_offset(page);
3854         struct extent_buffer *eb;
3855         int ret = 1;
3856
3857         spin_lock(&tree->buffer_lock);
3858         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3859         if (!eb) {
3860                 spin_unlock(&tree->buffer_lock);
3861                 return ret;
3862         }
3863
3864         if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3865                 ret = 0;
3866                 goto out;
3867         }
3868
3869         /*
3870          * set @eb->refs to 0 if it is already 1, and then release the @eb.
3871          * Or go back.
3872          */
3873         if (atomic_cmpxchg(&eb->refs, 1, 0) != 1) {
3874                 ret = 0;
3875                 goto out;
3876         }
3877
3878         radix_tree_delete(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3879 out:
3880         spin_unlock(&tree->buffer_lock);
3881
3882         /* at this point we can safely release the extent buffer */
3883         if (atomic_read(&eb->refs) == 0)
3884                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3885         return ret;
3886 }