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/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26 static struct bio_set *btrfs_bioset;
  27
  28 #ifdef CONFIG_BTRFS_DEBUG
  29 static LIST_HEAD(buffers);
  30 static LIST_HEAD(states);
  31
  32 static DEFINE_SPINLOCK(leak_lock);
  33
  34 static inline
  35 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  36 {
  37         unsigned long flags;
  38
  39         spin_lock_irqsave(&leak_lock, flags);
  40         list_add(new, head);
  41         spin_unlock_irqrestore(&leak_lock, flags);
  42 }
  43
  44 static inline
  45 void btrfs_leak_debug_del(struct list_head *entry)
  46 {
  47         unsigned long flags;
  48
  49         spin_lock_irqsave(&leak_lock, flags);
  50         list_del(entry);
  51         spin_unlock_irqrestore(&leak_lock, flags);
  52 }
  53
  54 static inline
  55 void btrfs_leak_debug_check(void)
  56 {
  57         struct extent_state *state;
  58         struct extent_buffer *eb;
  59
  60         while (!list_empty(&states)) {
  61                 state = list_entry(states.next, struct extent_state, leak_list);
  62                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  63                        "state %lu in tree %p refs %d\n",
  64                        state->start, state->end, state->state, state->tree,
  65                        atomic_read(&state->refs));
  66                 list_del(&state->leak_list);
  67                 kmem_cache_free(extent_state_cache, state);
  68         }
  69
  70         while (!list_empty(&buffers)) {
  71                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  72                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  73                        "refs %d\n",
  74                        eb->start, eb->len, atomic_read(&eb->refs));
  75                 list_del(&eb->leak_list);
  76                 kmem_cache_free(extent_buffer_cache, eb);
  77         }
  78 }
  79
  80 #define btrfs_debug_check_extent_io_range(inode, start, end)            \
  81         __btrfs_debug_check_extent_io_range(__func__, (inode), (start), (end))
  82 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  83                 struct inode *inode, u64 start, u64 end)
  84 {
  85         u64 isize = i_size_read(inode);
  86
  87         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  88                 printk_ratelimited(KERN_DEBUG
  89                     "btrfs: %s: ino %llu isize %llu odd range [%llu,%llu]\n",
  90                                 caller, btrfs_ino(inode), isize, start, end);
  91         }
  92 }
  93 #else
  94 #define btrfs_leak_debug_add(new, head) do {} while (0)
  95 #define btrfs_leak_debug_del(entry)     do {} while (0)
  96 #define btrfs_leak_debug_check()        do {} while (0)
  97 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
  98 #endif
  99
 100 #define BUFFER_LRU_MAX 64
 101
 102 struct tree_entry {
 103         u64 start;
 104         u64 end;
 105         struct rb_node rb_node;
 106 };
 107
 108 struct extent_page_data {
 109         struct bio *bio;
 110         struct extent_io_tree *tree;
 111         get_extent_t *get_extent;
 112         unsigned long bio_flags;
 113
 114         /* tells writepage not to lock the state bits for this range
 115          * it still does the unlocking
 116          */
 117         unsigned int extent_locked:1;
 118
 119         /* tells the submit_bio code to use a WRITE_SYNC */
 120         unsigned int sync_io:1;
 121 };
 122
 123 static noinline void flush_write_bio(void *data);
 124 static inline struct btrfs_fs_info *
 125 tree_fs_info(struct extent_io_tree *tree)
 126 {
 127         return btrfs_sb(tree->mapping->host->i_sb);
 128 }
 129
 130 int __init extent_io_init(void)
 131 {
 132         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 133                         sizeof(struct extent_state), 0,
 134                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 135         if (!extent_state_cache)
 136                 return -ENOMEM;
 137
 138         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 139                         sizeof(struct extent_buffer), 0,
 140                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 141         if (!extent_buffer_cache)
 142                 goto free_state_cache;
 143
 144         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 145                                      offsetof(struct btrfs_io_bio, bio));
 146         if (!btrfs_bioset)
 147                 goto free_buffer_cache;
 148
 149         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 150                 goto free_bioset;
 151
 152         return 0;
 153
 154 free_bioset:
 155         bioset_free(btrfs_bioset);
 156         btrfs_bioset = NULL;
 157
 158 free_buffer_cache:
 159         kmem_cache_destroy(extent_buffer_cache);
 160         extent_buffer_cache = NULL;
 161
 162 free_state_cache:
 163         kmem_cache_destroy(extent_state_cache);
 164         extent_state_cache = NULL;
 165         return -ENOMEM;
 166 }
 167
 168 void extent_io_exit(void)
 169 {
 170         btrfs_leak_debug_check();
 171
 172         /*
 173          * Make sure all delayed rcu free are flushed before we
 174          * destroy caches.
 175          */
 176         rcu_barrier();
 177         if (extent_state_cache)
 178                 kmem_cache_destroy(extent_state_cache);
 179         if (extent_buffer_cache)
 180                 kmem_cache_destroy(extent_buffer_cache);
 181         if (btrfs_bioset)
 182                 bioset_free(btrfs_bioset);
 183 }
 184
 185 void extent_io_tree_init(struct extent_io_tree *tree,
 186                          struct address_space *mapping)
 187 {
 188         tree->state = RB_ROOT;
 189         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 190         tree->ops = NULL;
 191         tree->dirty_bytes = 0;
 192         spin_lock_init(&tree->lock);
 193         spin_lock_init(&tree->buffer_lock);
 194         tree->mapping = mapping;
 195 }
 196
 197 static struct extent_state *alloc_extent_state(gfp_t mask)
 198 {
 199         struct extent_state *state;
 200
 201         state = kmem_cache_alloc(extent_state_cache, mask);
 202         if (!state)
 203                 return state;
 204         state->state = 0;
 205         state->private = 0;
 206         state->tree = NULL;
 207         btrfs_leak_debug_add(&state->leak_list, &states);
 208         atomic_set(&state->refs, 1);
 209         init_waitqueue_head(&state->wq);
 210         trace_alloc_extent_state(state, mask, _RET_IP_);
 211         return state;
 212 }
 213
 214 void free_extent_state(struct extent_state *state)
 215 {
 216         if (!state)
 217                 return;
 218         if (atomic_dec_and_test(&state->refs)) {
 219                 WARN_ON(state->tree);
 220                 btrfs_leak_debug_del(&state->leak_list);
 221                 trace_free_extent_state(state, _RET_IP_);
 222                 kmem_cache_free(extent_state_cache, state);
 223         }
 224 }
 225
 226 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 227                                    struct rb_node *node)
 228 {
 229         struct rb_node **p = &root->rb_node;
 230         struct rb_node *parent = NULL;
 231         struct tree_entry *entry;
 232
 233         while (*p) {
 234                 parent = *p;
 235                 entry = rb_entry(parent, struct tree_entry, rb_node);
 236
 237                 if (offset < entry->start)
 238                         p = &(*p)->rb_left;
 239                 else if (offset > entry->end)
 240                         p = &(*p)->rb_right;
 241                 else
 242                         return parent;
 243         }
 244
 245         rb_link_node(node, parent, p);
 246         rb_insert_color(node, root);
 247         return NULL;
 248 }
 249
 250 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 251                                      struct rb_node **prev_ret,
 252                                      struct rb_node **next_ret)
 253 {
 254         struct rb_root *root = &tree->state;
 255         struct rb_node *n = root->rb_node;
 256         struct rb_node *prev = NULL;
 257         struct rb_node *orig_prev = NULL;
 258         struct tree_entry *entry;
 259         struct tree_entry *prev_entry = NULL;
 260
 261         while (n) {
 262                 entry = rb_entry(n, struct tree_entry, rb_node);
 263                 prev = n;
 264                 prev_entry = entry;
 265
 266                 if (offset < entry->start)
 267                         n = n->rb_left;
 268                 else if (offset > entry->end)
 269                         n = n->rb_right;
 270                 else
 271                         return n;
 272         }
 273
 274         if (prev_ret) {
 275                 orig_prev = prev;
 276                 while (prev && offset > prev_entry->end) {
 277                         prev = rb_next(prev);
 278                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 279                 }
 280                 *prev_ret = prev;
 281                 prev = orig_prev;
 282         }
 283
 284         if (next_ret) {
 285                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 286                 while (prev && offset < prev_entry->start) {
 287                         prev = rb_prev(prev);
 288                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 289                 }
 290                 *next_ret = prev;
 291         }
 292         return NULL;
 293 }
 294
 295 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 296                                           u64 offset)
 297 {
 298         struct rb_node *prev = NULL;
 299         struct rb_node *ret;
 300
 301         ret = __etree_search(tree, offset, &prev, NULL);
 302         if (!ret)
 303                 return prev;
 304         return ret;
 305 }
 306
 307 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 308                      struct extent_state *other)
 309 {
 310         if (tree->ops && tree->ops->merge_extent_hook)
 311                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 312                                              other);
 313 }
 314
 315 /*
 316  * utility function to look for merge candidates inside a given range.
 317  * Any extents with matching state are merged together into a single
 318  * extent in the tree.  Extents with EXTENT_IO in their state field
 319  * are not merged because the end_io handlers need to be able to do
 320  * operations on them without sleeping (or doing allocations/splits).
 321  *
 322  * This should be called with the tree lock held.
 323  */
 324 static void merge_state(struct extent_io_tree *tree,
 325                         struct extent_state *state)
 326 {
 327         struct extent_state *other;
 328         struct rb_node *other_node;
 329
 330         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 331                 return;
 332
 333         other_node = rb_prev(&state->rb_node);
 334         if (other_node) {
 335                 other = rb_entry(other_node, struct extent_state, rb_node);
 336                 if (other->end == state->start - 1 &&
 337                     other->state == state->state) {
 338                         merge_cb(tree, state, other);
 339                         state->start = other->start;
 340                         other->tree = NULL;
 341                         rb_erase(&other->rb_node, &tree->state);
 342                         free_extent_state(other);
 343                 }
 344         }
 345         other_node = rb_next(&state->rb_node);
 346         if (other_node) {
 347                 other = rb_entry(other_node, struct extent_state, rb_node);
 348                 if (other->start == state->end + 1 &&
 349                     other->state == state->state) {
 350                         merge_cb(tree, state, other);
 351                         state->end = other->end;
 352                         other->tree = NULL;
 353                         rb_erase(&other->rb_node, &tree->state);
 354                         free_extent_state(other);
 355                 }
 356         }
 357 }
 358
 359 static void set_state_cb(struct extent_io_tree *tree,
 360                          struct extent_state *state, unsigned long *bits)
 361 {
 362         if (tree->ops && tree->ops->set_bit_hook)
 363                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 364 }
 365
 366 static void clear_state_cb(struct extent_io_tree *tree,
 367                            struct extent_state *state, unsigned long *bits)
 368 {
 369         if (tree->ops && tree->ops->clear_bit_hook)
 370                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 371 }
 372
 373 static void set_state_bits(struct extent_io_tree *tree,
 374                            struct extent_state *state, unsigned long *bits);
 375
 376 /*
 377  * insert an extent_state struct into the tree.  'bits' are set on the
 378  * struct before it is inserted.
 379  *
 380  * This may return -EEXIST if the extent is already there, in which case the
 381  * state struct is freed.
 382  *
 383  * The tree lock is not taken internally.  This is a utility function and
 384  * probably isn't what you want to call (see set/clear_extent_bit).
 385  */
 386 static int insert_state(struct extent_io_tree *tree,
 387                         struct extent_state *state, u64 start, u64 end,
 388                         unsigned long *bits)
 389 {
 390         struct rb_node *node;
 391
 392         if (end < start)
 393                 WARN(1, KERN_ERR "btrfs end < start %llu %llu\n",
 394                        end, start);
 395         state->start = start;
 396         state->end = end;
 397
 398         set_state_bits(tree, state, bits);
 399
 400         node = tree_insert(&tree->state, end, &state->rb_node);
 401         if (node) {
 402                 struct extent_state *found;
 403                 found = rb_entry(node, struct extent_state, rb_node);
 404                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 405                        "%llu %llu\n",
 406                        found->start, found->end, start, end);
 407                 return -EEXIST;
 408         }
 409         state->tree = tree;
 410         merge_state(tree, state);
 411         return 0;
 412 }
 413
 414 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 415                      u64 split)
 416 {
 417         if (tree->ops && tree->ops->split_extent_hook)
 418                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 419 }
 420
 421 /*
 422  * split a given extent state struct in two, inserting the preallocated
 423  * struct 'prealloc' as the newly created second half.  'split' indicates an
 424  * offset inside 'orig' where it should be split.
 425  *
 426  * Before calling,
 427  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 428  * are two extent state structs in the tree:
 429  * prealloc: [orig->start, split - 1]
 430  * orig: [ split, orig->end ]
 431  *
 432  * The tree locks are not taken by this function. They need to be held
 433  * by the caller.
 434  */
 435 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 436                        struct extent_state *prealloc, u64 split)
 437 {
 438         struct rb_node *node;
 439
 440         split_cb(tree, orig, split);
 441
 442         prealloc->start = orig->start;
 443         prealloc->end = split - 1;
 444         prealloc->state = orig->state;
 445         orig->start = split;
 446
 447         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 448         if (node) {
 449                 free_extent_state(prealloc);
 450                 return -EEXIST;
 451         }
 452         prealloc->tree = tree;
 453         return 0;
 454 }
 455
 456 static struct extent_state *next_state(struct extent_state *state)
 457 {
 458         struct rb_node *next = rb_next(&state->rb_node);
 459         if (next)
 460                 return rb_entry(next, struct extent_state, rb_node);
 461         else
 462                 return NULL;
 463 }
 464
 465 /*
 466  * utility function to clear some bits in an extent state struct.
 467  * it will optionally wake up any one waiting on this state (wake == 1).
 468  *
 469  * If no bits are set on the state struct after clearing things, the
 470  * struct is freed and removed from the tree
 471  */
 472 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 473                                             struct extent_state *state,
 474                                             unsigned long *bits, int wake)
 475 {
 476         struct extent_state *next;
 477         unsigned long bits_to_clear = *bits & ~EXTENT_CTLBITS;
 478
 479         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 480                 u64 range = state->end - state->start + 1;
 481                 WARN_ON(range > tree->dirty_bytes);
 482                 tree->dirty_bytes -= range;
 483         }
 484         clear_state_cb(tree, state, bits);
 485         state->state &= ~bits_to_clear;
 486         if (wake)
 487                 wake_up(&state->wq);
 488         if (state->state == 0) {
 489                 next = next_state(state);
 490                 if (state->tree) {
 491                         rb_erase(&state->rb_node, &tree->state);
 492                         state->tree = NULL;
 493                         free_extent_state(state);
 494                 } else {
 495                         WARN_ON(1);
 496                 }
 497         } else {
 498                 merge_state(tree, state);
 499                 next = next_state(state);
 500         }
 501         return next;
 502 }
 503
 504 static struct extent_state *
 505 alloc_extent_state_atomic(struct extent_state *prealloc)
 506 {
 507         if (!prealloc)
 508                 prealloc = alloc_extent_state(GFP_ATOMIC);
 509
 510         return prealloc;
 511 }
 512
 513 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 514 {
 515         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 516                     "Extent tree was modified by another "
 517                     "thread while locked.");
 518 }
 519
 520 /*
 521  * clear some bits on a range in the tree.  This may require splitting
 522  * or inserting elements in the tree, so the gfp mask is used to
 523  * indicate which allocations or sleeping are allowed.
 524  *
 525  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 526  * the given range from the tree regardless of state (ie for truncate).
 527  *
 528  * the range [start, end] is inclusive.
 529  *
 530  * This takes the tree lock, and returns 0 on success and < 0 on error.
 531  */
 532 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 533                      unsigned long bits, int wake, int delete,
 534                      struct extent_state **cached_state,
 535                      gfp_t mask)
 536 {
 537         struct extent_state *state;
 538         struct extent_state *cached;
 539         struct extent_state *prealloc = NULL;
 540         struct rb_node *node;
 541         u64 last_end;
 542         int err;
 543         int clear = 0;
 544
 545         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
 546
 547         if (bits & EXTENT_DELALLOC)
 548                 bits |= EXTENT_NORESERVE;
 549
 550         if (delete)
 551                 bits |= ~EXTENT_CTLBITS;
 552         bits |= EXTENT_FIRST_DELALLOC;
 553
 554         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 555                 clear = 1;
 556 again:
 557         if (!prealloc && (mask & __GFP_WAIT)) {
 558                 prealloc = alloc_extent_state(mask);
 559                 if (!prealloc)
 560                         return -ENOMEM;
 561         }
 562
 563         spin_lock(&tree->lock);
 564         if (cached_state) {
 565                 cached = *cached_state;
 566
 567                 if (clear) {
 568                         *cached_state = NULL;
 569                         cached_state = NULL;
 570                 }
 571
 572                 if (cached && cached->tree && cached->start <= start &&
 573                     cached->end > start) {
 574                         if (clear)
 575                                 atomic_dec(&cached->refs);
 576                         state = cached;
 577                         goto hit_next;
 578                 }
 579                 if (clear)
 580                         free_extent_state(cached);
 581         }
 582         /*
 583          * this search will find the extents that end after
 584          * our range starts
 585          */
 586         node = tree_search(tree, start);
 587         if (!node)
 588                 goto out;
 589         state = rb_entry(node, struct extent_state, rb_node);
 590 hit_next:
 591         if (state->start > end)
 592                 goto out;
 593         WARN_ON(state->end < start);
 594         last_end = state->end;
 595
 596         /* the state doesn't have the wanted bits, go ahead */
 597         if (!(state->state & bits)) {
 598                 state = next_state(state);
 599                 goto next;
 600         }
 601
 602         /*
 603          *     | ---- desired range ---- |
 604          *  | state | or
 605          *  | ------------- state -------------- |
 606          *
 607          * We need to split the extent we found, and may flip
 608          * bits on second half.
 609          *
 610          * If the extent we found extends past our range, we
 611          * just split and search again.  It'll get split again
 612          * the next time though.
 613          *
 614          * If the extent we found is inside our range, we clear
 615          * the desired bit on it.
 616          */
 617
 618         if (state->start < start) {
 619                 prealloc = alloc_extent_state_atomic(prealloc);
 620                 BUG_ON(!prealloc);
 621                 err = split_state(tree, state, prealloc, start);
 622                 if (err)
 623                         extent_io_tree_panic(tree, err);
 624
 625                 prealloc = NULL;
 626                 if (err)
 627                         goto out;
 628                 if (state->end <= end) {
 629                         state = clear_state_bit(tree, state, &bits, wake);
 630                         goto next;
 631                 }
 632                 goto search_again;
 633         }
 634         /*
 635          * | ---- desired range ---- |
 636          *                        | state |
 637          * We need to split the extent, and clear the bit
 638          * on the first half
 639          */
 640         if (state->start <= end && state->end > end) {
 641                 prealloc = alloc_extent_state_atomic(prealloc);
 642                 BUG_ON(!prealloc);
 643                 err = split_state(tree, state, prealloc, end + 1);
 644                 if (err)
 645                         extent_io_tree_panic(tree, err);
 646
 647                 if (wake)
 648                         wake_up(&state->wq);
 649
 650                 clear_state_bit(tree, prealloc, &bits, wake);
 651
 652                 prealloc = NULL;
 653                 goto out;
 654         }
 655
 656         state = clear_state_bit(tree, state, &bits, wake);
 657 next:
 658         if (last_end == (u64)-1)
 659                 goto out;
 660         start = last_end + 1;
 661         if (start <= end && state && !need_resched())
 662                 goto hit_next;
 663         goto search_again;
 664
 665 out:
 666         spin_unlock(&tree->lock);
 667         if (prealloc)
 668                 free_extent_state(prealloc);
 669
 670         return 0;
 671
 672 search_again:
 673         if (start > end)
 674                 goto out;
 675         spin_unlock(&tree->lock);
 676         if (mask & __GFP_WAIT)
 677                 cond_resched();
 678         goto again;
 679 }
 680
 681 static void wait_on_state(struct extent_io_tree *tree,
 682                           struct extent_state *state)
 683                 __releases(tree->lock)
 684                 __acquires(tree->lock)
 685 {
 686         DEFINE_WAIT(wait);
 687         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 688         spin_unlock(&tree->lock);
 689         schedule();
 690         spin_lock(&tree->lock);
 691         finish_wait(&state->wq, &wait);
 692 }
 693
 694 /*
 695  * waits for one or more bits to clear on a range in the state tree.
 696  * The range [start, end] is inclusive.
 697  * The tree lock is taken by this function
 698  */
 699 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 700                             unsigned long bits)
 701 {
 702         struct extent_state *state;
 703         struct rb_node *node;
 704
 705         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
 706
 707         spin_lock(&tree->lock);
 708 again:
 709         while (1) {
 710                 /*
 711                  * this search will find all the extents that end after
 712                  * our range starts
 713                  */
 714                 node = tree_search(tree, start);
 715                 if (!node)
 716                         break;
 717
 718                 state = rb_entry(node, struct extent_state, rb_node);
 719
 720                 if (state->start > end)
 721                         goto out;
 722
 723                 if (state->state & bits) {
 724                         start = state->start;
 725                         atomic_inc(&state->refs);
 726                         wait_on_state(tree, state);
 727                         free_extent_state(state);
 728                         goto again;
 729                 }
 730                 start = state->end + 1;
 731
 732                 if (start > end)
 733                         break;
 734
 735                 cond_resched_lock(&tree->lock);
 736         }
 737 out:
 738         spin_unlock(&tree->lock);
 739 }
 740
 741 static void set_state_bits(struct extent_io_tree *tree,
 742                            struct extent_state *state,
 743                            unsigned long *bits)
 744 {
 745         unsigned long bits_to_set = *bits & ~EXTENT_CTLBITS;
 746
 747         set_state_cb(tree, state, bits);
 748         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 749                 u64 range = state->end - state->start + 1;
 750                 tree->dirty_bytes += range;
 751         }
 752         state->state |= bits_to_set;
 753 }
 754
 755 static void cache_state(struct extent_state *state,
 756                         struct extent_state **cached_ptr)
 757 {
 758         if (cached_ptr && !(*cached_ptr)) {
 759                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 760                         *cached_ptr = state;
 761                         atomic_inc(&state->refs);
 762                 }
 763         }
 764 }
 765
 766 /*
 767  * set some bits on a range in the tree.  This may require allocations or
 768  * sleeping, so the gfp mask is used to indicate what is allowed.
 769  *
 770  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 771  * part of the range already has the desired bits set.  The start of the
 772  * existing range is returned in failed_start in this case.
 773  *
 774  * [start, end] is inclusive This takes the tree lock.
 775  */
 776
 777 static int __must_check
 778 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 779                  unsigned long bits, unsigned long exclusive_bits,
 780                  u64 *failed_start, struct extent_state **cached_state,
 781                  gfp_t mask)
 782 {
 783         struct extent_state *state;
 784         struct extent_state *prealloc = NULL;
 785         struct rb_node *node;
 786         int err = 0;
 787         u64 last_start;
 788         u64 last_end;
 789
 790         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
 791
 792         bits |= EXTENT_FIRST_DELALLOC;
 793 again:
 794         if (!prealloc && (mask & __GFP_WAIT)) {
 795                 prealloc = alloc_extent_state(mask);
 796                 BUG_ON(!prealloc);
 797         }
 798
 799         spin_lock(&tree->lock);
 800         if (cached_state && *cached_state) {
 801                 state = *cached_state;
 802                 if (state->start <= start && state->end > start &&
 803                     state->tree) {
 804                         node = &state->rb_node;
 805                         goto hit_next;
 806                 }
 807         }
 808         /*
 809          * this search will find all the extents that end after
 810          * our range starts.
 811          */
 812         node = tree_search(tree, start);
 813         if (!node) {
 814                 prealloc = alloc_extent_state_atomic(prealloc);
 815                 BUG_ON(!prealloc);
 816                 err = insert_state(tree, prealloc, start, end, &bits);
 817                 if (err)
 818                         extent_io_tree_panic(tree, err);
 819
 820                 prealloc = NULL;
 821                 goto out;
 822         }
 823         state = rb_entry(node, struct extent_state, rb_node);
 824 hit_next:
 825         last_start = state->start;
 826         last_end = state->end;
 827
 828         /*
 829          * | ---- desired range ---- |
 830          * | state |
 831          *
 832          * Just lock what we found and keep going
 833          */
 834         if (state->start == start && state->end <= end) {
 835                 if (state->state & exclusive_bits) {
 836                         *failed_start = state->start;
 837                         err = -EEXIST;
 838                         goto out;
 839                 }
 840
 841                 set_state_bits(tree, state, &bits);
 842                 cache_state(state, cached_state);
 843                 merge_state(tree, state);
 844                 if (last_end == (u64)-1)
 845                         goto out;
 846                 start = last_end + 1;
 847                 state = next_state(state);
 848                 if (start < end && state && state->start == start &&
 849                     !need_resched())
 850                         goto hit_next;
 851                 goto search_again;
 852         }
 853
 854         /*
 855          *     | ---- desired range ---- |
 856          * | state |
 857          *   or
 858          * | ------------- state -------------- |
 859          *
 860          * We need to split the extent we found, and may flip bits on
 861          * second half.
 862          *
 863          * If the extent we found extends past our
 864          * range, we just split and search again.  It'll get split
 865          * again the next time though.
 866          *
 867          * If the extent we found is inside our range, we set the
 868          * desired bit on it.
 869          */
 870         if (state->start < start) {
 871                 if (state->state & exclusive_bits) {
 872                         *failed_start = start;
 873                         err = -EEXIST;
 874                         goto out;
 875                 }
 876
 877                 prealloc = alloc_extent_state_atomic(prealloc);
 878                 BUG_ON(!prealloc);
 879                 err = split_state(tree, state, prealloc, start);
 880                 if (err)
 881                         extent_io_tree_panic(tree, err);
 882
 883                 prealloc = NULL;
 884                 if (err)
 885                         goto out;
 886                 if (state->end <= end) {
 887                         set_state_bits(tree, state, &bits);
 888                         cache_state(state, cached_state);
 889                         merge_state(tree, state);
 890                         if (last_end == (u64)-1)
 891                                 goto out;
 892                         start = last_end + 1;
 893                         state = next_state(state);
 894                         if (start < end && state && state->start == start &&
 895                             !need_resched())
 896                                 goto hit_next;
 897                 }
 898                 goto search_again;
 899         }
 900         /*
 901          * | ---- desired range ---- |
 902          *     | state | or               | state |
 903          *
 904          * There's a hole, we need to insert something in it and
 905          * ignore the extent we found.
 906          */
 907         if (state->start > start) {
 908                 u64 this_end;
 909                 if (end < last_start)
 910                         this_end = end;
 911                 else
 912                         this_end = last_start - 1;
 913
 914                 prealloc = alloc_extent_state_atomic(prealloc);
 915                 BUG_ON(!prealloc);
 916
 917                 /*
 918                  * Avoid to free 'prealloc' if it can be merged with
 919                  * the later extent.
 920                  */
 921                 err = insert_state(tree, prealloc, start, this_end,
 922                                    &bits);
 923                 if (err)
 924                         extent_io_tree_panic(tree, err);
 925
 926                 cache_state(prealloc, cached_state);
 927                 prealloc = NULL;
 928                 start = this_end + 1;
 929                 goto search_again;
 930         }
 931         /*
 932          * | ---- desired range ---- |
 933          *                        | state |
 934          * We need to split the extent, and set the bit
 935          * on the first half
 936          */
 937         if (state->start <= end && state->end > end) {
 938                 if (state->state & exclusive_bits) {
 939                         *failed_start = start;
 940                         err = -EEXIST;
 941                         goto out;
 942                 }
 943
 944                 prealloc = alloc_extent_state_atomic(prealloc);
 945                 BUG_ON(!prealloc);
 946                 err = split_state(tree, state, prealloc, end + 1);
 947                 if (err)
 948                         extent_io_tree_panic(tree, err);
 949
 950                 set_state_bits(tree, prealloc, &bits);
 951                 cache_state(prealloc, cached_state);
 952                 merge_state(tree, prealloc);
 953                 prealloc = NULL;
 954                 goto out;
 955         }
 956
 957         goto search_again;
 958
 959 out:
 960         spin_unlock(&tree->lock);
 961         if (prealloc)
 962                 free_extent_state(prealloc);
 963
 964         return err;
 965
 966 search_again:
 967         if (start > end)
 968                 goto out;
 969         spin_unlock(&tree->lock);
 970         if (mask & __GFP_WAIT)
 971                 cond_resched();
 972         goto again;
 973 }
 974
 975 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 976                    unsigned long bits, u64 * failed_start,
 977                    struct extent_state **cached_state, gfp_t mask)
 978 {
 979         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
 980                                 cached_state, mask);
 981 }
 982
 983
 984 /**
 985  * convert_extent_bit - convert all bits in a given range from one bit to
 986  *                      another
 987  * @tree:       the io tree to search
 988  * @start:      the start offset in bytes
 989  * @end:        the end offset in bytes (inclusive)
 990  * @bits:       the bits to set in this range
 991  * @clear_bits: the bits to clear in this range
 992  * @cached_state:       state that we're going to cache
 993  * @mask:       the allocation mask
 994  *
 995  * This will go through and set bits for the given range.  If any states exist
 996  * already in this range they are set with the given bit and cleared of the
 997  * clear_bits.  This is only meant to be used by things that are mergeable, ie
 998  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
 999  * boundary bits like LOCK.
1000  */
1001 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1002                        unsigned long bits, unsigned long clear_bits,
1003                        struct extent_state **cached_state, gfp_t mask)
1004 {
1005         struct extent_state *state;
1006         struct extent_state *prealloc = NULL;
1007         struct rb_node *node;
1008         int err = 0;
1009         u64 last_start;
1010         u64 last_end;
1011
1012         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
1013
1014 again:
1015         if (!prealloc && (mask & __GFP_WAIT)) {
1016                 prealloc = alloc_extent_state(mask);
1017                 if (!prealloc)
1018                         return -ENOMEM;
1019         }
1020
1021         spin_lock(&tree->lock);
1022         if (cached_state && *cached_state) {
1023                 state = *cached_state;
1024                 if (state->start <= start && state->end > start &&
1025                     state->tree) {
1026                         node = &state->rb_node;
1027                         goto hit_next;
1028                 }
1029         }
1030
1031         /*
1032          * this search will find all the extents that end after
1033          * our range starts.
1034          */
1035         node = tree_search(tree, start);
1036         if (!node) {
1037                 prealloc = alloc_extent_state_atomic(prealloc);
1038                 if (!prealloc) {
1039                         err = -ENOMEM;
1040                         goto out;
1041                 }
1042                 err = insert_state(tree, prealloc, start, end, &bits);
1043                 prealloc = NULL;
1044                 if (err)
1045                         extent_io_tree_panic(tree, err);
1046                 goto out;
1047         }
1048         state = rb_entry(node, struct extent_state, rb_node);
1049 hit_next:
1050         last_start = state->start;
1051         last_end = state->end;
1052
1053         /*
1054          * | ---- desired range ---- |
1055          * | state |
1056          *
1057          * Just lock what we found and keep going
1058          */
1059         if (state->start == start && state->end <= end) {
1060                 set_state_bits(tree, state, &bits);
1061                 cache_state(state, cached_state);
1062                 state = clear_state_bit(tree, state, &clear_bits, 0);
1063                 if (last_end == (u64)-1)
1064                         goto out;
1065                 start = last_end + 1;
1066                 if (start < end && state && state->start == start &&
1067                     !need_resched())
1068                         goto hit_next;
1069                 goto search_again;
1070         }
1071
1072         /*
1073          *     | ---- desired range ---- |
1074          * | state |
1075          *   or
1076          * | ------------- state -------------- |
1077          *
1078          * We need to split the extent we found, and may flip bits on
1079          * second half.
1080          *
1081          * If the extent we found extends past our
1082          * range, we just split and search again.  It'll get split
1083          * again the next time though.
1084          *
1085          * If the extent we found is inside our range, we set the
1086          * desired bit on it.
1087          */
1088         if (state->start < start) {
1089                 prealloc = alloc_extent_state_atomic(prealloc);
1090                 if (!prealloc) {
1091                         err = -ENOMEM;
1092                         goto out;
1093                 }
1094                 err = split_state(tree, state, prealloc, start);
1095                 if (err)
1096                         extent_io_tree_panic(tree, err);
1097                 prealloc = NULL;
1098                 if (err)
1099                         goto out;
1100                 if (state->end <= end) {
1101                         set_state_bits(tree, state, &bits);
1102                         cache_state(state, cached_state);
1103                         state = clear_state_bit(tree, state, &clear_bits, 0);
1104                         if (last_end == (u64)-1)
1105                                 goto out;
1106                         start = last_end + 1;
1107                         if (start < end && state && state->start == start &&
1108                             !need_resched())
1109                                 goto hit_next;
1110                 }
1111                 goto search_again;
1112         }
1113         /*
1114          * | ---- desired range ---- |
1115          *     | state | or               | state |
1116          *
1117          * There's a hole, we need to insert something in it and
1118          * ignore the extent we found.
1119          */
1120         if (state->start > start) {
1121                 u64 this_end;
1122                 if (end < last_start)
1123                         this_end = end;
1124                 else
1125                         this_end = last_start - 1;
1126
1127                 prealloc = alloc_extent_state_atomic(prealloc);
1128                 if (!prealloc) {
1129                         err = -ENOMEM;
1130                         goto out;
1131                 }
1132
1133                 /*
1134                  * Avoid to free 'prealloc' if it can be merged with
1135                  * the later extent.
1136                  */
1137                 err = insert_state(tree, prealloc, start, this_end,
1138                                    &bits);
1139                 if (err)
1140                         extent_io_tree_panic(tree, err);
1141                 cache_state(prealloc, cached_state);
1142                 prealloc = NULL;
1143                 start = this_end + 1;
1144                 goto search_again;
1145         }
1146         /*
1147          * | ---- desired range ---- |
1148          *                        | state |
1149          * We need to split the extent, and set the bit
1150          * on the first half
1151          */
1152         if (state->start <= end && state->end > end) {
1153                 prealloc = alloc_extent_state_atomic(prealloc);
1154                 if (!prealloc) {
1155                         err = -ENOMEM;
1156                         goto out;
1157                 }
1158
1159                 err = split_state(tree, state, prealloc, end + 1);
1160                 if (err)
1161                         extent_io_tree_panic(tree, err);
1162
1163                 set_state_bits(tree, prealloc, &bits);
1164                 cache_state(prealloc, cached_state);
1165                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1166                 prealloc = NULL;
1167                 goto out;
1168         }
1169
1170         goto search_again;
1171
1172 out:
1173         spin_unlock(&tree->lock);
1174         if (prealloc)
1175                 free_extent_state(prealloc);
1176
1177         return err;
1178
1179 search_again:
1180         if (start > end)
1181                 goto out;
1182         spin_unlock(&tree->lock);
1183         if (mask & __GFP_WAIT)
1184                 cond_resched();
1185         goto again;
1186 }
1187
1188 /* wrappers around set/clear extent bit */
1189 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1190                      gfp_t mask)
1191 {
1192         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1193                               NULL, mask);
1194 }
1195
1196 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1197                     unsigned long bits, gfp_t mask)
1198 {
1199         return set_extent_bit(tree, start, end, bits, NULL,
1200                               NULL, mask);
1201 }
1202
1203 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1204                       unsigned long bits, gfp_t mask)
1205 {
1206         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1207 }
1208
1209 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1210                         struct extent_state **cached_state, gfp_t mask)
1211 {
1212         return set_extent_bit(tree, start, end,
1213                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1214                               NULL, cached_state, mask);
1215 }
1216
1217 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1218                       struct extent_state **cached_state, gfp_t mask)
1219 {
1220         return set_extent_bit(tree, start, end,
1221                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1222                               NULL, cached_state, mask);
1223 }
1224
1225 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1226                        gfp_t mask)
1227 {
1228         return clear_extent_bit(tree, start, end,
1229                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1230                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1231 }
1232
1233 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1234                      gfp_t mask)
1235 {
1236         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1237                               NULL, mask);
1238 }
1239
1240 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1241                         struct extent_state **cached_state, gfp_t mask)
1242 {
1243         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1244                               cached_state, mask);
1245 }
1246
1247 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1248                           struct extent_state **cached_state, gfp_t mask)
1249 {
1250         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1251                                 cached_state, mask);
1252 }
1253
1254 /*
1255  * either insert or lock state struct between start and end use mask to tell
1256  * us if waiting is desired.
1257  */
1258 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1259                      unsigned long bits, struct extent_state **cached_state)
1260 {
1261         int err;
1262         u64 failed_start;
1263         while (1) {
1264                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1265                                        EXTENT_LOCKED, &failed_start,
1266                                        cached_state, GFP_NOFS);
1267                 if (err == -EEXIST) {
1268                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1269                         start = failed_start;
1270                 } else
1271                         break;
1272                 WARN_ON(start > end);
1273         }
1274         return err;
1275 }
1276
1277 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1278 {
1279         return lock_extent_bits(tree, start, end, 0, NULL);
1280 }
1281
1282 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1283 {
1284         int err;
1285         u64 failed_start;
1286
1287         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1288                                &failed_start, NULL, GFP_NOFS);
1289         if (err == -EEXIST) {
1290                 if (failed_start > start)
1291                         clear_extent_bit(tree, start, failed_start - 1,
1292                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1293                 return 0;
1294         }
1295         return 1;
1296 }
1297
1298 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1299                          struct extent_state **cached, gfp_t mask)
1300 {
1301         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1302                                 mask);
1303 }
1304
1305 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1306 {
1307         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1308                                 GFP_NOFS);
1309 }
1310
1311 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1312 {
1313         unsigned long index = start >> PAGE_CACHE_SHIFT;
1314         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1315         struct page *page;
1316
1317         while (index <= end_index) {
1318                 page = find_get_page(inode->i_mapping, index);
1319                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1320                 clear_page_dirty_for_io(page);
1321                 page_cache_release(page);
1322                 index++;
1323         }
1324         return 0;
1325 }
1326
1327 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1328 {
1329         unsigned long index = start >> PAGE_CACHE_SHIFT;
1330         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1331         struct page *page;
1332
1333         while (index <= end_index) {
1334                 page = find_get_page(inode->i_mapping, index);
1335                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1336                 account_page_redirty(page);
1337                 __set_page_dirty_nobuffers(page);
1338                 page_cache_release(page);
1339                 index++;
1340         }
1341         return 0;
1342 }
1343
1344 /*
1345  * helper function to set both pages and extents in the tree writeback
1346  */
1347 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1348 {
1349         unsigned long index = start >> PAGE_CACHE_SHIFT;
1350         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1351         struct page *page;
1352
1353         while (index <= end_index) {
1354                 page = find_get_page(tree->mapping, index);
1355                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1356                 set_page_writeback(page);
1357                 page_cache_release(page);
1358                 index++;
1359         }
1360         return 0;
1361 }
1362
1363 /* find the first state struct with 'bits' set after 'start', and
1364  * return it.  tree->lock must be held.  NULL will returned if
1365  * nothing was found after 'start'
1366  */
1367 static struct extent_state *
1368 find_first_extent_bit_state(struct extent_io_tree *tree,
1369                             u64 start, unsigned long bits)
1370 {
1371         struct rb_node *node;
1372         struct extent_state *state;
1373
1374         /*
1375          * this search will find all the extents that end after
1376          * our range starts.
1377          */
1378         node = tree_search(tree, start);
1379         if (!node)
1380                 goto out;
1381
1382         while (1) {
1383                 state = rb_entry(node, struct extent_state, rb_node);
1384                 if (state->end >= start && (state->state & bits))
1385                         return state;
1386
1387                 node = rb_next(node);
1388                 if (!node)
1389                         break;
1390         }
1391 out:
1392         return NULL;
1393 }
1394
1395 /*
1396  * find the first offset in the io tree with 'bits' set. zero is
1397  * returned if we find something, and *start_ret and *end_ret are
1398  * set to reflect the state struct that was found.
1399  *
1400  * If nothing was found, 1 is returned. If found something, return 0.
1401  */
1402 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1403                           u64 *start_ret, u64 *end_ret, unsigned long bits,
1404                           struct extent_state **cached_state)
1405 {
1406         struct extent_state *state;
1407         struct rb_node *n;
1408         int ret = 1;
1409
1410         spin_lock(&tree->lock);
1411         if (cached_state && *cached_state) {
1412                 state = *cached_state;
1413                 if (state->end == start - 1 && state->tree) {
1414                         n = rb_next(&state->rb_node);
1415                         while (n) {
1416                                 state = rb_entry(n, struct extent_state,
1417                                                  rb_node);
1418                                 if (state->state & bits)
1419                                         goto got_it;
1420                                 n = rb_next(n);
1421                         }
1422                         free_extent_state(*cached_state);
1423                         *cached_state = NULL;
1424                         goto out;
1425                 }
1426                 free_extent_state(*cached_state);
1427                 *cached_state = NULL;
1428         }
1429
1430         state = find_first_extent_bit_state(tree, start, bits);
1431 got_it:
1432         if (state) {
1433                 cache_state(state, cached_state);
1434                 *start_ret = state->start;
1435                 *end_ret = state->end;
1436                 ret = 0;
1437         }
1438 out:
1439         spin_unlock(&tree->lock);
1440         return ret;
1441 }
1442
1443 /*
1444  * find a contiguous range of bytes in the file marked as delalloc, not
1445  * more than 'max_bytes'.  start and end are used to return the range,
1446  *
1447  * 1 is returned if we find something, 0 if nothing was in the tree
1448  */
1449 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1450                                         u64 *start, u64 *end, u64 max_bytes,
1451                                         struct extent_state **cached_state)
1452 {
1453         struct rb_node *node;
1454         struct extent_state *state;
1455         u64 cur_start = *start;
1456         u64 found = 0;
1457         u64 total_bytes = 0;
1458
1459         spin_lock(&tree->lock);
1460
1461         /*
1462          * this search will find all the extents that end after
1463          * our range starts.
1464          */
1465         node = tree_search(tree, cur_start);
1466         if (!node) {
1467                 if (!found)
1468                         *end = (u64)-1;
1469                 goto out;
1470         }
1471
1472         while (1) {
1473                 state = rb_entry(node, struct extent_state, rb_node);
1474                 if (found && (state->start != cur_start ||
1475                               (state->state & EXTENT_BOUNDARY))) {
1476                         goto out;
1477                 }
1478                 if (!(state->state & EXTENT_DELALLOC)) {
1479                         if (!found)
1480                                 *end = state->end;
1481                         goto out;
1482                 }
1483                 if (!found) {
1484                         *start = state->start;
1485                         *cached_state = state;
1486                         atomic_inc(&state->refs);
1487                 }
1488                 found++;
1489                 *end = state->end;
1490                 cur_start = state->end + 1;
1491                 node = rb_next(node);
1492                 total_bytes += state->end - state->start + 1;
1493                 if (total_bytes >= max_bytes)
1494                         break;
1495                 if (!node)
1496                         break;
1497         }
1498 out:
1499         spin_unlock(&tree->lock);
1500         return found;
1501 }
1502
1503 static noinline void __unlock_for_delalloc(struct inode *inode,
1504                                            struct page *locked_page,
1505                                            u64 start, u64 end)
1506 {
1507         int ret;
1508         struct page *pages[16];
1509         unsigned long index = start >> PAGE_CACHE_SHIFT;
1510         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1511         unsigned long nr_pages = end_index - index + 1;
1512         int i;
1513
1514         if (index == locked_page->index && end_index == index)
1515                 return;
1516
1517         while (nr_pages > 0) {
1518                 ret = find_get_pages_contig(inode->i_mapping, index,
1519                                      min_t(unsigned long, nr_pages,
1520                                      ARRAY_SIZE(pages)), pages);
1521                 for (i = 0; i < ret; i++) {
1522                         if (pages[i] != locked_page)
1523                                 unlock_page(pages[i]);
1524                         page_cache_release(pages[i]);
1525                 }
1526                 nr_pages -= ret;
1527                 index += ret;
1528                 cond_resched();
1529         }
1530 }
1531
1532 static noinline int lock_delalloc_pages(struct inode *inode,
1533                                         struct page *locked_page,
1534                                         u64 delalloc_start,
1535                                         u64 delalloc_end)
1536 {
1537         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1538         unsigned long start_index = index;
1539         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1540         unsigned long pages_locked = 0;
1541         struct page *pages[16];
1542         unsigned long nrpages;
1543         int ret;
1544         int i;
1545
1546         /* the caller is responsible for locking the start index */
1547         if (index == locked_page->index && index == end_index)
1548                 return 0;
1549
1550         /* skip the page at the start index */
1551         nrpages = end_index - index + 1;
1552         while (nrpages > 0) {
1553                 ret = find_get_pages_contig(inode->i_mapping, index,
1554                                      min_t(unsigned long,
1555                                      nrpages, ARRAY_SIZE(pages)), pages);
1556                 if (ret == 0) {
1557                         ret = -EAGAIN;
1558                         goto done;
1559                 }
1560                 /* now we have an array of pages, lock them all */
1561                 for (i = 0; i < ret; i++) {
1562                         /*
1563                          * the caller is taking responsibility for
1564                          * locked_page
1565                          */
1566                         if (pages[i] != locked_page) {
1567                                 lock_page(pages[i]);
1568                                 if (!PageDirty(pages[i]) ||
1569                                     pages[i]->mapping != inode->i_mapping) {
1570                                         ret = -EAGAIN;
1571                                         unlock_page(pages[i]);
1572                                         page_cache_release(pages[i]);
1573                                         goto done;
1574                                 }
1575                         }
1576                         page_cache_release(pages[i]);
1577                         pages_locked++;
1578                 }
1579                 nrpages -= ret;
1580                 index += ret;
1581                 cond_resched();
1582         }
1583         ret = 0;
1584 done:
1585         if (ret && pages_locked) {
1586                 __unlock_for_delalloc(inode, locked_page,
1587                               delalloc_start,
1588                               ((u64)(start_index + pages_locked - 1)) <<
1589                               PAGE_CACHE_SHIFT);
1590         }
1591         return ret;
1592 }
1593
1594 /*
1595  * find a contiguous range of bytes in the file marked as delalloc, not
1596  * more than 'max_bytes'.  start and end are used to return the range,
1597  *
1598  * 1 is returned if we find something, 0 if nothing was in the tree
1599  */
1600 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1601                                     struct extent_io_tree *tree,
1602                                     struct page *locked_page, u64 *start,
1603                                     u64 *end, u64 max_bytes)
1604 {
1605         u64 delalloc_start;
1606         u64 delalloc_end;
1607         u64 found;
1608         struct extent_state *cached_state = NULL;
1609         int ret;
1610         int loops = 0;
1611
1612 again:
1613         /* step one, find a bunch of delalloc bytes starting at start */
1614         delalloc_start = *start;
1615         delalloc_end = 0;
1616         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1617                                     max_bytes, &cached_state);
1618         if (!found || delalloc_end <= *start) {
1619                 *start = delalloc_start;
1620                 *end = delalloc_end;
1621                 free_extent_state(cached_state);
1622                 return 0;
1623         }
1624
1625         /*
1626          * start comes from the offset of locked_page.  We have to lock
1627          * pages in order, so we can't process delalloc bytes before
1628          * locked_page
1629          */
1630         if (delalloc_start < *start)
1631                 delalloc_start = *start;
1632
1633         /*
1634          * make sure to limit the number of pages we try to lock down
1635          */
1636         if (delalloc_end + 1 - delalloc_start > max_bytes)
1637                 delalloc_end = delalloc_start + max_bytes - 1;
1638
1639         /* step two, lock all the pages after the page that has start */
1640         ret = lock_delalloc_pages(inode, locked_page,
1641                                   delalloc_start, delalloc_end);
1642         if (ret == -EAGAIN) {
1643                 /* some of the pages are gone, lets avoid looping by
1644                  * shortening the size of the delalloc range we're searching
1645                  */
1646                 free_extent_state(cached_state);
1647                 if (!loops) {
1648                         max_bytes = PAGE_CACHE_SIZE;
1649                         loops = 1;
1650                         goto again;
1651                 } else {
1652                         found = 0;
1653                         goto out_failed;
1654                 }
1655         }
1656         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1657
1658         /* step three, lock the state bits for the whole range */
1659         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1660
1661         /* then test to make sure it is all still delalloc */
1662         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1663                              EXTENT_DELALLOC, 1, cached_state);
1664         if (!ret) {
1665                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1666                                      &cached_state, GFP_NOFS);
1667                 __unlock_for_delalloc(inode, locked_page,
1668                               delalloc_start, delalloc_end);
1669                 cond_resched();
1670                 goto again;
1671         }
1672         free_extent_state(cached_state);
1673         *start = delalloc_start;
1674         *end = delalloc_end;
1675 out_failed:
1676         return found;
1677 }
1678
1679 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1680                                  struct page *locked_page,
1681                                  unsigned long clear_bits,
1682                                  unsigned long page_ops)
1683 {
1684         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1685         int ret;
1686         struct page *pages[16];
1687         unsigned long index = start >> PAGE_CACHE_SHIFT;
1688         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1689         unsigned long nr_pages = end_index - index + 1;
1690         int i;
1691
1692         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1693         if (page_ops == 0)
1694                 return 0;
1695
1696         while (nr_pages > 0) {
1697                 ret = find_get_pages_contig(inode->i_mapping, index,
1698                                      min_t(unsigned long,
1699                                      nr_pages, ARRAY_SIZE(pages)), pages);
1700                 for (i = 0; i < ret; i++) {
1701
1702                         if (page_ops & PAGE_SET_PRIVATE2)
1703                                 SetPagePrivate2(pages[i]);
1704
1705                         if (pages[i] == locked_page) {
1706                                 page_cache_release(pages[i]);
1707                                 continue;
1708                         }
1709                         if (page_ops & PAGE_CLEAR_DIRTY)
1710                                 clear_page_dirty_for_io(pages[i]);
1711                         if (page_ops & PAGE_SET_WRITEBACK)
1712                                 set_page_writeback(pages[i]);
1713                         if (page_ops & PAGE_END_WRITEBACK)
1714                                 end_page_writeback(pages[i]);
1715                         if (page_ops & PAGE_UNLOCK)
1716                                 unlock_page(pages[i]);
1717                         page_cache_release(pages[i]);
1718                 }
1719                 nr_pages -= ret;
1720                 index += ret;
1721                 cond_resched();
1722         }
1723         return 0;
1724 }
1725
1726 /*
1727  * count the number of bytes in the tree that have a given bit(s)
1728  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1729  * cached.  The total number found is returned.
1730  */
1731 u64 count_range_bits(struct extent_io_tree *tree,
1732                      u64 *start, u64 search_end, u64 max_bytes,
1733                      unsigned long bits, int contig)
1734 {
1735         struct rb_node *node;
1736         struct extent_state *state;
1737         u64 cur_start = *start;
1738         u64 total_bytes = 0;
1739         u64 last = 0;
1740         int found = 0;
1741
1742         if (WARN_ON(search_end <= cur_start))
1743                 return 0;
1744
1745         spin_lock(&tree->lock);
1746         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1747                 total_bytes = tree->dirty_bytes;
1748                 goto out;
1749         }
1750         /*
1751          * this search will find all the extents that end after
1752          * our range starts.
1753          */
1754         node = tree_search(tree, cur_start);
1755         if (!node)
1756                 goto out;
1757
1758         while (1) {
1759                 state = rb_entry(node, struct extent_state, rb_node);
1760                 if (state->start > search_end)
1761                         break;
1762                 if (contig && found && state->start > last + 1)
1763                         break;
1764                 if (state->end >= cur_start && (state->state & bits) == bits) {
1765                         total_bytes += min(search_end, state->end) + 1 -
1766                                        max(cur_start, state->start);
1767                         if (total_bytes >= max_bytes)
1768                                 break;
1769                         if (!found) {
1770                                 *start = max(cur_start, state->start);
1771                                 found = 1;
1772                         }
1773                         last = state->end;
1774                 } else if (contig && found) {
1775                         break;
1776                 }
1777                 node = rb_next(node);
1778                 if (!node)
1779                         break;
1780         }
1781 out:
1782         spin_unlock(&tree->lock);
1783         return total_bytes;
1784 }
1785
1786 /*
1787  * set the private field for a given byte offset in the tree.  If there isn't
1788  * an extent_state there already, this does nothing.
1789  */
1790 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1791 {
1792         struct rb_node *node;
1793         struct extent_state *state;
1794         int ret = 0;
1795
1796         spin_lock(&tree->lock);
1797         /*
1798          * this search will find all the extents that end after
1799          * our range starts.
1800          */
1801         node = tree_search(tree, start);
1802         if (!node) {
1803                 ret = -ENOENT;
1804                 goto out;
1805         }
1806         state = rb_entry(node, struct extent_state, rb_node);
1807         if (state->start != start) {
1808                 ret = -ENOENT;
1809                 goto out;
1810         }
1811         state->private = private;
1812 out:
1813         spin_unlock(&tree->lock);
1814         return ret;
1815 }
1816
1817 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1818 {
1819         struct rb_node *node;
1820         struct extent_state *state;
1821         int ret = 0;
1822
1823         spin_lock(&tree->lock);
1824         /*
1825          * this search will find all the extents that end after
1826          * our range starts.
1827          */
1828         node = tree_search(tree, start);
1829         if (!node) {
1830                 ret = -ENOENT;
1831                 goto out;
1832         }
1833         state = rb_entry(node, struct extent_state, rb_node);
1834         if (state->start != start) {
1835                 ret = -ENOENT;
1836                 goto out;
1837         }
1838         *private = state->private;
1839 out:
1840         spin_unlock(&tree->lock);
1841         return ret;
1842 }
1843
1844 /*
1845  * searches a range in the state tree for a given mask.
1846  * If 'filled' == 1, this returns 1 only if every extent in the tree
1847  * has the bits set.  Otherwise, 1 is returned if any bit in the
1848  * range is found set.
1849  */
1850 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1851                    unsigned long bits, int filled, struct extent_state *cached)
1852 {
1853         struct extent_state *state = NULL;
1854         struct rb_node *node;
1855         int bitset = 0;
1856
1857         spin_lock(&tree->lock);
1858         if (cached && cached->tree && cached->start <= start &&
1859             cached->end > start)
1860                 node = &cached->rb_node;
1861         else
1862                 node = tree_search(tree, start);
1863         while (node && start <= end) {
1864                 state = rb_entry(node, struct extent_state, rb_node);
1865
1866                 if (filled && state->start > start) {
1867                         bitset = 0;
1868                         break;
1869                 }
1870
1871                 if (state->start > end)
1872                         break;
1873
1874                 if (state->state & bits) {
1875                         bitset = 1;
1876                         if (!filled)
1877                                 break;
1878                 } else if (filled) {
1879                         bitset = 0;
1880                         break;
1881                 }
1882
1883                 if (state->end == (u64)-1)
1884                         break;
1885
1886                 start = state->end + 1;
1887                 if (start > end)
1888                         break;
1889                 node = rb_next(node);
1890                 if (!node) {
1891                         if (filled)
1892                                 bitset = 0;
1893                         break;
1894                 }
1895         }
1896         spin_unlock(&tree->lock);
1897         return bitset;
1898 }
1899
1900 /*
1901  * helper function to set a given page up to date if all the
1902  * extents in the tree for that page are up to date
1903  */
1904 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1905 {
1906         u64 start = page_offset(page);
1907         u64 end = start + PAGE_CACHE_SIZE - 1;
1908         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1909                 SetPageUptodate(page);
1910 }
1911
1912 /*
1913  * When IO fails, either with EIO or csum verification fails, we
1914  * try other mirrors that might have a good copy of the data.  This
1915  * io_failure_record is used to record state as we go through all the
1916  * mirrors.  If another mirror has good data, the page is set up to date
1917  * and things continue.  If a good mirror can't be found, the original
1918  * bio end_io callback is called to indicate things have failed.
1919  */
1920 struct io_failure_record {
1921         struct page *page;
1922         u64 start;
1923         u64 len;
1924         u64 logical;
1925         unsigned long bio_flags;
1926         int this_mirror;
1927         int failed_mirror;
1928         int in_validation;
1929 };
1930
1931 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1932                                 int did_repair)
1933 {
1934         int ret;
1935         int err = 0;
1936         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1937
1938         set_state_private(failure_tree, rec->start, 0);
1939         ret = clear_extent_bits(failure_tree, rec->start,
1940                                 rec->start + rec->len - 1,
1941                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1942         if (ret)
1943                 err = ret;
1944
1945         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1946                                 rec->start + rec->len - 1,
1947                                 EXTENT_DAMAGED, GFP_NOFS);
1948         if (ret && !err)
1949                 err = ret;
1950
1951         kfree(rec);
1952         return err;
1953 }
1954
1955 static void repair_io_failure_callback(struct bio *bio, int err)
1956 {
1957         complete(bio->bi_private);
1958 }
1959
1960 /*
1961  * this bypasses the standard btrfs submit functions deliberately, as
1962  * the standard behavior is to write all copies in a raid setup. here we only
1963  * want to write the one bad copy. so we do the mapping for ourselves and issue
1964  * submit_bio directly.
1965  * to avoid any synchronization issues, wait for the data after writing, which
1966  * actually prevents the read that triggered the error from finishing.
1967  * currently, there can be no more than two copies of every data bit. thus,
1968  * exactly one rewrite is required.
1969  */
1970 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 start,
1971                         u64 length, u64 logical, struct page *page,
1972                         int mirror_num)
1973 {
1974         struct bio *bio;
1975         struct btrfs_device *dev;
1976         DECLARE_COMPLETION_ONSTACK(compl);
1977         u64 map_length = 0;
1978         u64 sector;
1979         struct btrfs_bio *bbio = NULL;
1980         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
1981         int ret;
1982
1983         BUG_ON(!mirror_num);
1984
1985         /* we can't repair anything in raid56 yet */
1986         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
1987                 return 0;
1988
1989         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1990         if (!bio)
1991                 return -EIO;
1992         bio->bi_private = &compl;
1993         bio->bi_end_io = repair_io_failure_callback;
1994         bio->bi_size = 0;
1995         map_length = length;
1996
1997         ret = btrfs_map_block(fs_info, WRITE, logical,
1998                               &map_length, &bbio, mirror_num);
1999         if (ret) {
2000                 bio_put(bio);
2001                 return -EIO;
2002         }
2003         BUG_ON(mirror_num != bbio->mirror_num);
2004         sector = bbio->stripes[mirror_num-1].physical >> 9;
2005         bio->bi_sector = sector;
2006         dev = bbio->stripes[mirror_num-1].dev;
2007         kfree(bbio);
2008         if (!dev || !dev->bdev || !dev->writeable) {
2009                 bio_put(bio);
2010                 return -EIO;
2011         }
2012         bio->bi_bdev = dev->bdev;
2013         bio_add_page(bio, page, length, start - page_offset(page));
2014         btrfsic_submit_bio(WRITE_SYNC, bio);
2015         wait_for_completion(&compl);
2016
2017         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2018                 /* try to remap that extent elsewhere? */
2019                 bio_put(bio);
2020                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2021                 return -EIO;
2022         }
2023
2024         printk_ratelimited_in_rcu(KERN_INFO "btrfs read error corrected: ino %lu off %llu "
2025                       "(dev %s sector %llu)\n", page->mapping->host->i_ino,
2026                       start, rcu_str_deref(dev->name), sector);
2027
2028         bio_put(bio);
2029         return 0;
2030 }
2031
2032 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2033                          int mirror_num)
2034 {
2035         u64 start = eb->start;
2036         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2037         int ret = 0;
2038
2039         for (i = 0; i < num_pages; i++) {
2040                 struct page *p = extent_buffer_page(eb, i);
2041                 ret = repair_io_failure(root->fs_info, start, PAGE_CACHE_SIZE,
2042                                         start, p, mirror_num);
2043                 if (ret)
2044                         break;
2045                 start += PAGE_CACHE_SIZE;
2046         }
2047
2048         return ret;
2049 }
2050
2051 /*
2052  * each time an IO finishes, we do a fast check in the IO failure tree
2053  * to see if we need to process or clean up an io_failure_record
2054  */
2055 static int clean_io_failure(u64 start, struct page *page)
2056 {
2057         u64 private;
2058         u64 private_failure;
2059         struct io_failure_record *failrec;
2060         struct btrfs_fs_info *fs_info;
2061         struct extent_state *state;
2062         int num_copies;
2063         int did_repair = 0;
2064         int ret;
2065         struct inode *inode = page->mapping->host;
2066
2067         private = 0;
2068         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2069                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2070         if (!ret)
2071                 return 0;
2072
2073         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2074                                 &private_failure);
2075         if (ret)
2076                 return 0;
2077
2078         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2079         BUG_ON(!failrec->this_mirror);
2080
2081         if (failrec->in_validation) {
2082                 /* there was no real error, just free the record */
2083                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2084                          failrec->start);
2085                 did_repair = 1;
2086                 goto out;
2087         }
2088
2089         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2090         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2091                                             failrec->start,
2092                                             EXTENT_LOCKED);
2093         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2094
2095         if (state && state->start <= failrec->start &&
2096             state->end >= failrec->start + failrec->len - 1) {
2097                 fs_info = BTRFS_I(inode)->root->fs_info;
2098                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2099                                               failrec->len);
2100                 if (num_copies > 1)  {
2101                         ret = repair_io_failure(fs_info, start, failrec->len,
2102                                                 failrec->logical, page,
2103                                                 failrec->failed_mirror);
2104                         did_repair = !ret;
2105                 }
2106                 ret = 0;
2107         }
2108
2109 out:
2110         if (!ret)
2111                 ret = free_io_failure(inode, failrec, did_repair);
2112
2113         return ret;
2114 }
2115
2116 /*
2117  * this is a generic handler for readpage errors (default
2118  * readpage_io_failed_hook). if other copies exist, read those and write back
2119  * good data to the failed position. does not investigate in remapping the
2120  * failed extent elsewhere, hoping the device will be smart enough to do this as
2121  * needed
2122  */
2123
2124 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2125                               struct page *page, u64 start, u64 end,
2126                               int failed_mirror)
2127 {
2128         struct io_failure_record *failrec = NULL;
2129         u64 private;
2130         struct extent_map *em;
2131         struct inode *inode = page->mapping->host;
2132         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2133         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2134         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2135         struct bio *bio;
2136         struct btrfs_io_bio *btrfs_failed_bio;
2137         struct btrfs_io_bio *btrfs_bio;
2138         int num_copies;
2139         int ret;
2140         int read_mode;
2141         u64 logical;
2142
2143         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2144
2145         ret = get_state_private(failure_tree, start, &private);
2146         if (ret) {
2147                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2148                 if (!failrec)
2149                         return -ENOMEM;
2150                 failrec->start = start;
2151                 failrec->len = end - start + 1;
2152                 failrec->this_mirror = 0;
2153                 failrec->bio_flags = 0;
2154                 failrec->in_validation = 0;
2155
2156                 read_lock(&em_tree->lock);
2157                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2158                 if (!em) {
2159                         read_unlock(&em_tree->lock);
2160                         kfree(failrec);
2161                         return -EIO;
2162                 }
2163
2164                 if (em->start > start || em->start + em->len < start) {
2165                         free_extent_map(em);
2166                         em = NULL;
2167                 }
2168                 read_unlock(&em_tree->lock);
2169
2170                 if (!em) {
2171                         kfree(failrec);
2172                         return -EIO;
2173                 }
2174                 logical = start - em->start;
2175                 logical = em->block_start + logical;
2176                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2177                         logical = em->block_start;
2178                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2179                         extent_set_compress_type(&failrec->bio_flags,
2180                                                  em->compress_type);
2181                 }
2182                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2183                          "len=%llu\n", logical, start, failrec->len);
2184                 failrec->logical = logical;
2185                 free_extent_map(em);
2186
2187                 /* set the bits in the private failure tree */
2188                 ret = set_extent_bits(failure_tree, start, end,
2189                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2190                 if (ret >= 0)
2191                         ret = set_state_private(failure_tree, start,
2192                                                 (u64)(unsigned long)failrec);
2193                 /* set the bits in the inode's tree */
2194                 if (ret >= 0)
2195                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2196                                                 GFP_NOFS);
2197                 if (ret < 0) {
2198                         kfree(failrec);
2199                         return ret;
2200                 }
2201         } else {
2202                 failrec = (struct io_failure_record *)(unsigned long)private;
2203                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2204                          "start=%llu, len=%llu, validation=%d\n",
2205                          failrec->logical, failrec->start, failrec->len,
2206                          failrec->in_validation);
2207                 /*
2208                  * when data can be on disk more than twice, add to failrec here
2209                  * (e.g. with a list for failed_mirror) to make
2210                  * clean_io_failure() clean all those errors at once.
2211                  */
2212         }
2213         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2214                                       failrec->logical, failrec->len);
2215         if (num_copies == 1) {
2216                 /*
2217                  * we only have a single copy of the data, so don't bother with
2218                  * all the retry and error correction code that follows. no
2219                  * matter what the error is, it is very likely to persist.
2220                  */
2221                 pr_debug("bio_readpage_error: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2222                          num_copies, failrec->this_mirror, failed_mirror);
2223                 free_io_failure(inode, failrec, 0);
2224                 return -EIO;
2225         }
2226
2227         /*
2228          * there are two premises:
2229          *      a) deliver good data to the caller
2230          *      b) correct the bad sectors on disk
2231          */
2232         if (failed_bio->bi_vcnt > 1) {
2233                 /*
2234                  * to fulfill b), we need to know the exact failing sectors, as
2235                  * we don't want to rewrite any more than the failed ones. thus,
2236                  * we need separate read requests for the failed bio
2237                  *
2238                  * if the following BUG_ON triggers, our validation request got
2239                  * merged. we need separate requests for our algorithm to work.
2240                  */
2241                 BUG_ON(failrec->in_validation);
2242                 failrec->in_validation = 1;
2243                 failrec->this_mirror = failed_mirror;
2244                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2245         } else {
2246                 /*
2247                  * we're ready to fulfill a) and b) alongside. get a good copy
2248                  * of the failed sector and if we succeed, we have setup
2249                  * everything for repair_io_failure to do the rest for us.
2250                  */
2251                 if (failrec->in_validation) {
2252                         BUG_ON(failrec->this_mirror != failed_mirror);
2253                         failrec->in_validation = 0;
2254                         failrec->this_mirror = 0;
2255                 }
2256                 failrec->failed_mirror = failed_mirror;
2257                 failrec->this_mirror++;
2258                 if (failrec->this_mirror == failed_mirror)
2259                         failrec->this_mirror++;
2260                 read_mode = READ_SYNC;
2261         }
2262
2263         if (failrec->this_mirror > num_copies) {
2264                 pr_debug("bio_readpage_error: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2265                          num_copies, failrec->this_mirror, failed_mirror);
2266                 free_io_failure(inode, failrec, 0);
2267                 return -EIO;
2268         }
2269
2270         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2271         if (!bio) {
2272                 free_io_failure(inode, failrec, 0);
2273                 return -EIO;
2274         }
2275         bio->bi_end_io = failed_bio->bi_end_io;
2276         bio->bi_sector = failrec->logical >> 9;
2277         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2278         bio->bi_size = 0;
2279
2280         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2281         if (btrfs_failed_bio->csum) {
2282                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2283                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2284
2285                 btrfs_bio = btrfs_io_bio(bio);
2286                 btrfs_bio->csum = btrfs_bio->csum_inline;
2287                 phy_offset >>= inode->i_sb->s_blocksize_bits;
2288                 phy_offset *= csum_size;
2289                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + phy_offset,
2290                        csum_size);
2291         }
2292
2293         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2294
2295         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2296                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2297                  failrec->this_mirror, num_copies, failrec->in_validation);
2298
2299         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2300                                          failrec->this_mirror,
2301                                          failrec->bio_flags, 0);
2302         return ret;
2303 }
2304
2305 /* lots and lots of room for performance fixes in the end_bio funcs */
2306
2307 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2308 {
2309         int uptodate = (err == 0);
2310         struct extent_io_tree *tree;
2311         int ret;
2312
2313         tree = &BTRFS_I(page->mapping->host)->io_tree;
2314
2315         if (tree->ops && tree->ops->writepage_end_io_hook) {
2316                 ret = tree->ops->writepage_end_io_hook(page, start,
2317                                                end, NULL, uptodate);
2318                 if (ret)
2319                         uptodate = 0;
2320         }
2321
2322         if (!uptodate) {
2323                 ClearPageUptodate(page);
2324                 SetPageError(page);
2325         }
2326         return 0;
2327 }
2328
2329 /*
2330  * after a writepage IO is done, we need to:
2331  * clear the uptodate bits on error
2332  * clear the writeback bits in the extent tree for this IO
2333  * end_page_writeback if the page has no more pending IO
2334  *
2335  * Scheduling is not allowed, so the extent state tree is expected
2336  * to have one and only one object corresponding to this IO.
2337  */
2338 static void end_bio_extent_writepage(struct bio *bio, int err)
2339 {
2340         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2341         struct extent_io_tree *tree;
2342         u64 start;
2343         u64 end;
2344
2345         do {
2346                 struct page *page = bvec->bv_page;
2347                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2348
2349                 /* We always issue full-page reads, but if some block
2350                  * in a page fails to read, blk_update_request() will
2351                  * advance bv_offset and adjust bv_len to compensate.
2352                  * Print a warning for nonzero offsets, and an error
2353                  * if they don't add up to a full page.  */
2354                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE)
2355                         printk("%s page write in btrfs with offset %u and length %u\n",
2356                                bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE
2357                                ? KERN_ERR "partial" : KERN_INFO "incomplete",
2358                                bvec->bv_offset, bvec->bv_len);
2359
2360                 start = page_offset(page);
2361                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2362
2363                 if (--bvec >= bio->bi_io_vec)
2364                         prefetchw(&bvec->bv_page->flags);
2365
2366                 if (end_extent_writepage(page, err, start, end))
2367                         continue;
2368
2369                 end_page_writeback(page);
2370         } while (bvec >= bio->bi_io_vec);
2371
2372         bio_put(bio);
2373 }
2374
2375 static void
2376 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2377                               int uptodate)
2378 {
2379         struct extent_state *cached = NULL;
2380         u64 end = start + len - 1;
2381
2382         if (uptodate && tree->track_uptodate)
2383                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2384         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2385 }
2386
2387 /*
2388  * after a readpage IO is done, we need to:
2389  * clear the uptodate bits on error
2390  * set the uptodate bits if things worked
2391  * set the page up to date if all extents in the tree are uptodate
2392  * clear the lock bit in the extent tree
2393  * unlock the page if there are no other extents locked for it
2394  *
2395  * Scheduling is not allowed, so the extent state tree is expected
2396  * to have one and only one object corresponding to this IO.
2397  */
2398 static void end_bio_extent_readpage(struct bio *bio, int err)
2399 {
2400         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2401         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2402         struct bio_vec *bvec = bio->bi_io_vec;
2403         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2404         struct extent_io_tree *tree;
2405         u64 offset = 0;
2406         u64 start;
2407         u64 end;
2408         u64 len;
2409         u64 extent_start = 0;
2410         u64 extent_len = 0;
2411         int mirror;
2412         int ret;
2413
2414         if (err)
2415                 uptodate = 0;
2416
2417         do {
2418                 struct page *page = bvec->bv_page;
2419                 struct inode *inode = page->mapping->host;
2420
2421                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2422                          "mirror=%lu\n", (u64)bio->bi_sector, err,
2423                          io_bio->mirror_num);
2424                 tree = &BTRFS_I(inode)->io_tree;
2425
2426                 /* We always issue full-page reads, but if some block
2427                  * in a page fails to read, blk_update_request() will
2428                  * advance bv_offset and adjust bv_len to compensate.
2429                  * Print a warning for nonzero offsets, and an error
2430                  * if they don't add up to a full page.  */
2431                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE)
2432                         printk("%s page read in btrfs with offset %u and length %u\n",
2433                                bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE
2434                                ? KERN_ERR "partial" : KERN_INFO "incomplete",
2435                                bvec->bv_offset, bvec->bv_len);
2436
2437                 start = page_offset(page);
2438                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2439                 len = bvec->bv_len;
2440
2441                 if (++bvec <= bvec_end)
2442                         prefetchw(&bvec->bv_page->flags);
2443
2444                 mirror = io_bio->mirror_num;
2445                 if (likely(uptodate && tree->ops &&
2446                            tree->ops->readpage_end_io_hook)) {
2447                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2448                                                               page, start, end,
2449                                                               mirror);
2450                         if (ret)
2451                                 uptodate = 0;
2452                         else
2453                                 clean_io_failure(start, page);
2454                 }
2455
2456                 if (likely(uptodate))
2457                         goto readpage_ok;
2458
2459                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2460                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2461                         if (!ret && !err &&
2462                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2463                                 uptodate = 1;
2464                 } else {
2465                         /*
2466                          * The generic bio_readpage_error handles errors the
2467                          * following way: If possible, new read requests are
2468                          * created and submitted and will end up in
2469                          * end_bio_extent_readpage as well (if we're lucky, not
2470                          * in the !uptodate case). In that case it returns 0 and
2471                          * we just go on with the next page in our bio. If it
2472                          * can't handle the error it will return -EIO and we
2473                          * remain responsible for that page.
2474                          */
2475                         ret = bio_readpage_error(bio, offset, page, start, end,
2476                                                  mirror);
2477                         if (ret == 0) {
2478                                 uptodate =
2479                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2480                                 if (err)
2481                                         uptodate = 0;
2482                                 continue;
2483                         }
2484                 }
2485 readpage_ok:
2486                 if (likely(uptodate)) {
2487                         loff_t i_size = i_size_read(inode);
2488                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2489                         unsigned offset;
2490
2491                         /* Zero out the end if this page straddles i_size */
2492                         offset = i_size & (PAGE_CACHE_SIZE-1);
2493                         if (page->index == end_index && offset)
2494                                 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2495                         SetPageUptodate(page);
2496                 } else {
2497                         ClearPageUptodate(page);
2498                         SetPageError(page);
2499                 }
2500                 unlock_page(page);
2501                 offset += len;
2502
2503                 if (unlikely(!uptodate)) {
2504                         if (extent_len) {
2505                                 endio_readpage_release_extent(tree,
2506                                                               extent_start,
2507                                                               extent_len, 1);
2508                                 extent_start = 0;
2509                                 extent_len = 0;
2510                         }
2511                         endio_readpage_release_extent(tree, start,
2512                                                       end - start + 1, 0);
2513                 } else if (!extent_len) {
2514                         extent_start = start;
2515                         extent_len = end + 1 - start;
2516                 } else if (extent_start + extent_len == start) {
2517                         extent_len += end + 1 - start;
2518                 } else {
2519                         endio_readpage_release_extent(tree, extent_start,
2520                                                       extent_len, uptodate);
2521                         extent_start = start;
2522                         extent_len = end + 1 - start;
2523                 }
2524         } while (bvec <= bvec_end);
2525
2526         if (extent_len)
2527                 endio_readpage_release_extent(tree, extent_start, extent_len,
2528                                               uptodate);
2529         if (io_bio->end_io)
2530                 io_bio->end_io(io_bio, err);
2531         bio_put(bio);
2532 }
2533
2534 /*
2535  * this allocates from the btrfs_bioset.  We're returning a bio right now
2536  * but you can call btrfs_io_bio for the appropriate container_of magic
2537  */
2538 struct bio *
2539 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2540                 gfp_t gfp_flags)
2541 {
2542         struct btrfs_io_bio *btrfs_bio;
2543         struct bio *bio;
2544
2545         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2546
2547         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2548                 while (!bio && (nr_vecs /= 2)) {
2549                         bio = bio_alloc_bioset(gfp_flags,
2550                                                nr_vecs, btrfs_bioset);
2551                 }
2552         }
2553
2554         if (bio) {
2555                 bio->bi_size = 0;
2556                 bio->bi_bdev = bdev;
2557                 bio->bi_sector = first_sector;
2558                 btrfs_bio = btrfs_io_bio(bio);
2559                 btrfs_bio->csum = NULL;
2560                 btrfs_bio->csum_allocated = NULL;
2561                 btrfs_bio->end_io = NULL;
2562         }
2563         return bio;
2564 }
2565
2566 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2567 {
2568         return bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2569 }
2570
2571
2572 /* this also allocates from the btrfs_bioset */
2573 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2574 {
2575         struct btrfs_io_bio *btrfs_bio;
2576         struct bio *bio;
2577
2578         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2579         if (bio) {
2580                 btrfs_bio = btrfs_io_bio(bio);
2581                 btrfs_bio->csum = NULL;
2582                 btrfs_bio->csum_allocated = NULL;
2583                 btrfs_bio->end_io = NULL;
2584         }
2585         return bio;
2586 }
2587
2588
2589 static int __must_check submit_one_bio(int rw, struct bio *bio,
2590                                        int mirror_num, unsigned long bio_flags)
2591 {
2592         int ret = 0;
2593         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2594         struct page *page = bvec->bv_page;
2595         struct extent_io_tree *tree = bio->bi_private;
2596         u64 start;
2597
2598         start = page_offset(page) + bvec->bv_offset;
2599
2600         bio->bi_private = NULL;
2601
2602         bio_get(bio);
2603
2604         if (tree->ops && tree->ops->submit_bio_hook)
2605                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2606                                            mirror_num, bio_flags, start);
2607         else
2608                 btrfsic_submit_bio(rw, bio);
2609
2610         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2611                 ret = -EOPNOTSUPP;
2612         bio_put(bio);
2613         return ret;
2614 }
2615
2616 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2617                      unsigned long offset, size_t size, struct bio *bio,
2618                      unsigned long bio_flags)
2619 {
2620         int ret = 0;
2621         if (tree->ops && tree->ops->merge_bio_hook)
2622                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2623                                                 bio_flags);
2624         BUG_ON(ret < 0);
2625         return ret;
2626
2627 }
2628
2629 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2630                               struct page *page, sector_t sector,
2631                               size_t size, unsigned long offset,
2632                               struct block_device *bdev,
2633                               struct bio **bio_ret,
2634                               unsigned long max_pages,
2635                               bio_end_io_t end_io_func,
2636                               int mirror_num,
2637                               unsigned long prev_bio_flags,
2638                               unsigned long bio_flags)
2639 {
2640         int ret = 0;
2641         struct bio *bio;
2642         int nr;
2643         int contig = 0;
2644         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2645         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2646         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2647
2648         if (bio_ret && *bio_ret) {
2649                 bio = *bio_ret;
2650                 if (old_compressed)
2651                         contig = bio->bi_sector == sector;
2652                 else
2653                         contig = bio_end_sector(bio) == sector;
2654
2655                 if (prev_bio_flags != bio_flags || !contig ||
2656                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2657                     bio_add_page(bio, page, page_size, offset) < page_size) {
2658                         ret = submit_one_bio(rw, bio, mirror_num,
2659                                              prev_bio_flags);
2660                         if (ret < 0)
2661                                 return ret;
2662                         bio = NULL;
2663                 } else {
2664                         return 0;
2665                 }
2666         }
2667         if (this_compressed)
2668                 nr = BIO_MAX_PAGES;
2669         else
2670                 nr = bio_get_nr_vecs(bdev);
2671
2672         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2673         if (!bio)
2674                 return -ENOMEM;
2675
2676         bio_add_page(bio, page, page_size, offset);
2677         bio->bi_end_io = end_io_func;
2678         bio->bi_private = tree;
2679
2680         if (bio_ret)
2681                 *bio_ret = bio;
2682         else
2683                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2684
2685         return ret;
2686 }
2687
2688 static void attach_extent_buffer_page(struct extent_buffer *eb,
2689                                       struct page *page)
2690 {
2691         if (!PagePrivate(page)) {
2692                 SetPagePrivate(page);
2693                 page_cache_get(page);
2694                 set_page_private(page, (unsigned long)eb);
2695         } else {
2696                 WARN_ON(page->private != (unsigned long)eb);
2697         }
2698 }
2699
2700 void set_page_extent_mapped(struct page *page)
2701 {
2702         if (!PagePrivate(page)) {
2703                 SetPagePrivate(page);
2704                 page_cache_get(page);
2705                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2706         }
2707 }
2708
2709 static struct extent_map *
2710 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2711                  u64 start, u64 len, get_extent_t *get_extent,
2712                  struct extent_map **em_cached)
2713 {
2714         struct extent_map *em;
2715
2716         if (em_cached && *em_cached) {
2717                 em = *em_cached;
2718                 if (em->in_tree && start >= em->start &&
2719                     start < extent_map_end(em)) {
2720                         atomic_inc(&em->refs);
2721                         return em;
2722                 }
2723
2724                 free_extent_map(em);
2725                 *em_cached = NULL;
2726         }
2727
2728         em = get_extent(inode, page, pg_offset, start, len, 0);
2729         if (em_cached && !IS_ERR_OR_NULL(em)) {
2730                 BUG_ON(*em_cached);
2731                 atomic_inc(&em->refs);
2732                 *em_cached = em;
2733         }
2734         return em;
2735 }
2736 /*
2737  * basic readpage implementation.  Locked extent state structs are inserted
2738  * into the tree that are removed when the IO is done (by the end_io
2739  * handlers)
2740  * XXX JDM: This needs looking at to ensure proper page locking
2741  */
2742 static int __do_readpage(struct extent_io_tree *tree,
2743                          struct page *page,
2744                          get_extent_t *get_extent,
2745                          struct extent_map **em_cached,
2746                          struct bio **bio, int mirror_num,
2747                          unsigned long *bio_flags, int rw)
2748 {
2749         struct inode *inode = page->mapping->host;
2750         u64 start = page_offset(page);
2751         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2752         u64 end;
2753         u64 cur = start;
2754         u64 extent_offset;
2755         u64 last_byte = i_size_read(inode);
2756         u64 block_start;
2757         u64 cur_end;
2758         sector_t sector;
2759         struct extent_map *em;
2760         struct block_device *bdev;
2761         int ret;
2762         int nr = 0;
2763         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2764         size_t pg_offset = 0;
2765         size_t iosize;
2766         size_t disk_io_size;
2767         size_t blocksize = inode->i_sb->s_blocksize;
2768         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2769
2770         set_page_extent_mapped(page);
2771
2772         end = page_end;
2773         if (!PageUptodate(page)) {
2774                 if (cleancache_get_page(page) == 0) {
2775                         BUG_ON(blocksize != PAGE_SIZE);
2776                         unlock_extent(tree, start, end);
2777                         goto out;
2778                 }
2779         }
2780
2781         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2782                 char *userpage;
2783                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2784
2785                 if (zero_offset) {
2786                         iosize = PAGE_CACHE_SIZE - zero_offset;
2787                         userpage = kmap_atomic(page);
2788                         memset(userpage + zero_offset, 0, iosize);
2789                         flush_dcache_page(page);
2790                         kunmap_atomic(userpage);
2791                 }
2792         }
2793         while (cur <= end) {
2794                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2795
2796                 if (cur >= last_byte) {
2797                         char *userpage;
2798                         struct extent_state *cached = NULL;
2799
2800                         iosize = PAGE_CACHE_SIZE - pg_offset;
2801                         userpage = kmap_atomic(page);
2802                         memset(userpage + pg_offset, 0, iosize);
2803                         flush_dcache_page(page);
2804                         kunmap_atomic(userpage);
2805                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2806                                             &cached, GFP_NOFS);
2807                         if (!parent_locked)
2808                                 unlock_extent_cached(tree, cur,
2809                                                      cur + iosize - 1,
2810                                                      &cached, GFP_NOFS);
2811                         break;
2812                 }
2813                 em = __get_extent_map(inode, page, pg_offset, cur,
2814                                       end - cur + 1, get_extent, em_cached);
2815                 if (IS_ERR_OR_NULL(em)) {
2816                         SetPageError(page);
2817                         if (!parent_locked)
2818                                 unlock_extent(tree, cur, end);
2819                         break;
2820                 }
2821                 extent_offset = cur - em->start;
2822                 BUG_ON(extent_map_end(em) <= cur);
2823                 BUG_ON(end < cur);
2824
2825                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2826                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2827                         extent_set_compress_type(&this_bio_flag,
2828                                                  em->compress_type);
2829                 }
2830
2831                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2832                 cur_end = min(extent_map_end(em) - 1, end);
2833                 iosize = ALIGN(iosize, blocksize);
2834                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2835                         disk_io_size = em->block_len;
2836                         sector = em->block_start >> 9;
2837                 } else {
2838                         sector = (em->block_start + extent_offset) >> 9;
2839                         disk_io_size = iosize;
2840                 }
2841                 bdev = em->bdev;
2842                 block_start = em->block_start;
2843                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2844                         block_start = EXTENT_MAP_HOLE;
2845                 free_extent_map(em);
2846                 em = NULL;
2847
2848                 /* we've found a hole, just zero and go on */
2849                 if (block_start == EXTENT_MAP_HOLE) {
2850                         char *userpage;
2851                         struct extent_state *cached = NULL;
2852
2853                         userpage = kmap_atomic(page);
2854                         memset(userpage + pg_offset, 0, iosize);
2855                         flush_dcache_page(page);
2856                         kunmap_atomic(userpage);
2857
2858                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2859                                             &cached, GFP_NOFS);
2860                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2861                                              &cached, GFP_NOFS);
2862                         cur = cur + iosize;
2863                         pg_offset += iosize;
2864                         continue;
2865                 }
2866                 /* the get_extent function already copied into the page */
2867                 if (test_range_bit(tree, cur, cur_end,
2868                                    EXTENT_UPTODATE, 1, NULL)) {
2869                         check_page_uptodate(tree, page);
2870                         if (!parent_locked)
2871                                 unlock_extent(tree, cur, cur + iosize - 1);
2872                         cur = cur + iosize;
2873                         pg_offset += iosize;
2874                         continue;
2875                 }
2876                 /* we have an inline extent but it didn't get marked up
2877                  * to date.  Error out
2878                  */
2879                 if (block_start == EXTENT_MAP_INLINE) {
2880                         SetPageError(page);
2881                         if (!parent_locked)
2882                                 unlock_extent(tree, cur, cur + iosize - 1);
2883                         cur = cur + iosize;
2884                         pg_offset += iosize;
2885                         continue;
2886                 }
2887
2888                 pnr -= page->index;
2889                 ret = submit_extent_page(rw, tree, page,
2890                                          sector, disk_io_size, pg_offset,
2891                                          bdev, bio, pnr,
2892                                          end_bio_extent_readpage, mirror_num,
2893                                          *bio_flags,
2894                                          this_bio_flag);
2895                 if (!ret) {
2896                         nr++;
2897                         *bio_flags = this_bio_flag;
2898                 } else {
2899                         SetPageError(page);
2900                         if (!parent_locked)
2901                                 unlock_extent(tree, cur, cur + iosize - 1);
2902                 }
2903                 cur = cur + iosize;
2904                 pg_offset += iosize;
2905         }
2906 out:
2907         if (!nr) {
2908                 if (!PageError(page))
2909                         SetPageUptodate(page);
2910                 unlock_page(page);
2911         }
2912         return 0;
2913 }
2914
2915 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
2916                                              struct page *pages[], int nr_pages,
2917                                              u64 start, u64 end,
2918                                              get_extent_t *get_extent,
2919                                              struct extent_map **em_cached,
2920                                              struct bio **bio, int mirror_num,
2921                                              unsigned long *bio_flags, int rw)
2922 {
2923         struct inode *inode;
2924         struct btrfs_ordered_extent *ordered;
2925         int index;
2926
2927         inode = pages[0]->mapping->host;
2928         while (1) {
2929                 lock_extent(tree, start, end);
2930                 ordered = btrfs_lookup_ordered_range(inode, start,
2931                                                      end - start + 1);
2932                 if (!ordered)
2933                         break;
2934                 unlock_extent(tree, start, end);
2935                 btrfs_start_ordered_extent(inode, ordered, 1);
2936                 btrfs_put_ordered_extent(ordered);
2937         }
2938
2939         for (index = 0; index < nr_pages; index++) {
2940                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
2941                               mirror_num, bio_flags, rw);
2942                 page_cache_release(pages[index]);
2943         }
2944 }
2945
2946 static void __extent_readpages(struct extent_io_tree *tree,
2947                                struct page *pages[],
2948                                int nr_pages, get_extent_t *get_extent,
2949                                struct extent_map **em_cached,
2950                                struct bio **bio, int mirror_num,
2951                                unsigned long *bio_flags, int rw)
2952 {
2953         u64 start = 0;
2954         u64 end = 0;
2955         u64 page_start;
2956         int index;
2957         int first_index = 0;
2958
2959         for (index = 0; index < nr_pages; index++) {
2960                 page_start = page_offset(pages[index]);
2961                 if (!end) {
2962                         start = page_start;
2963                         end = start + PAGE_CACHE_SIZE - 1;
2964                         first_index = index;
2965                 } else if (end + 1 == page_start) {
2966                         end += PAGE_CACHE_SIZE;
2967                 } else {
2968                         __do_contiguous_readpages(tree, &pages[first_index],
2969                                                   index - first_index, start,
2970                                                   end, get_extent, em_cached,
2971                                                   bio, mirror_num, bio_flags,
2972                                                   rw);
2973                         start = page_start;
2974                         end = start + PAGE_CACHE_SIZE - 1;
2975                         first_index = index;
2976                 }
2977         }
2978
2979         if (end)
2980                 __do_contiguous_readpages(tree, &pages[first_index],
2981                                           index - first_index, start,
2982                                           end, get_extent, em_cached, bio,
2983                                           mirror_num, bio_flags, rw);
2984 }
2985
2986 static int __extent_read_full_page(struct extent_io_tree *tree,
2987                                    struct page *page,
2988                                    get_extent_t *get_extent,
2989                                    struct bio **bio, int mirror_num,
2990                                    unsigned long *bio_flags, int rw)
2991 {
2992         struct inode *inode = page->mapping->host;
2993         struct btrfs_ordered_extent *ordered;
2994         u64 start = page_offset(page);
2995         u64 end = start + PAGE_CACHE_SIZE - 1;
2996         int ret;
2997
2998         while (1) {
2999                 lock_extent(tree, start, end);
3000                 ordered = btrfs_lookup_ordered_extent(inode, start);
3001                 if (!ordered)
3002                         break;
3003                 unlock_extent(tree, start, end);
3004                 btrfs_start_ordered_extent(inode, ordered, 1);
3005                 btrfs_put_ordered_extent(ordered);
3006         }
3007
3008         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3009                             bio_flags, rw);
3010         return ret;
3011 }
3012
3013 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3014                             get_extent_t *get_extent, int mirror_num)
3015 {
3016         struct bio *bio = NULL;
3017         unsigned long bio_flags = 0;
3018         int ret;
3019
3020         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3021                                       &bio_flags, READ);
3022         if (bio)
3023                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3024         return ret;
3025 }
3026
3027 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3028                                  get_extent_t *get_extent, int mirror_num)
3029 {
3030         struct bio *bio = NULL;
3031         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3032         int ret;
3033
3034         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3035                                       &bio_flags, READ);
3036         if (bio)
3037                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3038         return ret;
3039 }
3040
3041 static noinline void update_nr_written(struct page *page,
3042                                       struct writeback_control *wbc,
3043                                       unsigned long nr_written)
3044 {
3045         wbc->nr_to_write -= nr_written;
3046         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3047             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3048                 page->mapping->writeback_index = page->index + nr_written;
3049 }
3050
3051 /*
3052  * the writepage semantics are similar to regular writepage.  extent
3053  * records are inserted to lock ranges in the tree, and as dirty areas
3054  * are found, they are marked writeback.  Then the lock bits are removed
3055  * and the end_io handler clears the writeback ranges
3056  */
3057 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3058                               void *data)
3059 {
3060         struct inode *inode = page->mapping->host;
3061         struct extent_page_data *epd = data;
3062         struct extent_io_tree *tree = epd->tree;
3063         u64 start = page_offset(page);
3064         u64 delalloc_start;
3065         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3066         u64 end;
3067         u64 cur = start;
3068         u64 extent_offset;
3069         u64 last_byte = i_size_read(inode);
3070         u64 block_start;
3071         u64 iosize;
3072         sector_t sector;
3073         struct extent_state *cached_state = NULL;
3074         struct extent_map *em;
3075         struct block_device *bdev;
3076         int ret;
3077         int nr = 0;
3078         size_t pg_offset = 0;
3079         size_t blocksize;
3080         loff_t i_size = i_size_read(inode);
3081         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3082         u64 nr_delalloc;
3083         u64 delalloc_end;
3084         int page_started;
3085         int compressed;
3086         int write_flags;
3087         unsigned long nr_written = 0;
3088         bool fill_delalloc = true;
3089
3090         if (wbc->sync_mode == WB_SYNC_ALL)
3091                 write_flags = WRITE_SYNC;
3092         else
3093                 write_flags = WRITE;
3094
3095         trace___extent_writepage(page, inode, wbc);
3096
3097         WARN_ON(!PageLocked(page));
3098
3099         ClearPageError(page);
3100
3101         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3102         if (page->index > end_index ||
3103            (page->index == end_index && !pg_offset)) {
3104                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3105                 unlock_page(page);
3106                 return 0;
3107         }
3108
3109         if (page->index == end_index) {
3110                 char *userpage;
3111
3112                 userpage = kmap_atomic(page);
3113                 memset(userpage + pg_offset, 0,
3114                        PAGE_CACHE_SIZE - pg_offset);
3115                 kunmap_atomic(userpage);
3116                 flush_dcache_page(page);
3117         }
3118         pg_offset = 0;
3119
3120         set_page_extent_mapped(page);
3121
3122         if (!tree->ops || !tree->ops->fill_delalloc)
3123                 fill_delalloc = false;
3124
3125         delalloc_start = start;
3126         delalloc_end = 0;
3127         page_started = 0;
3128         if (!epd->extent_locked && fill_delalloc) {
3129                 u64 delalloc_to_write = 0;
3130                 /*
3131                  * make sure the wbc mapping index is at least updated
3132                  * to this page.
3133                  */
3134                 update_nr_written(page, wbc, 0);
3135
3136                 while (delalloc_end < page_end) {
3137                         nr_delalloc = find_lock_delalloc_range(inode, tree,
3138                                                        page,
3139                                                        &delalloc_start,
3140                                                        &delalloc_end,
3141                                                        128 * 1024 * 1024);
3142                         if (nr_delalloc == 0) {
3143                                 delalloc_start = delalloc_end + 1;
3144                                 continue;
3145                         }
3146                         ret = tree->ops->fill_delalloc(inode, page,
3147                                                        delalloc_start,
3148                                                        delalloc_end,
3149                                                        &page_started,
3150                                                        &nr_written);
3151                         /* File system has been set read-only */
3152                         if (ret) {
3153                                 SetPageError(page);
3154                                 goto done;
3155                         }
3156                         /*
3157                          * delalloc_end is already one less than the total
3158                          * length, so we don't subtract one from
3159                          * PAGE_CACHE_SIZE
3160                          */
3161                         delalloc_to_write += (delalloc_end - delalloc_start +
3162                                               PAGE_CACHE_SIZE) >>
3163                                               PAGE_CACHE_SHIFT;
3164                         delalloc_start = delalloc_end + 1;
3165                 }
3166                 if (wbc->nr_to_write < delalloc_to_write) {
3167                         int thresh = 8192;
3168
3169                         if (delalloc_to_write < thresh * 2)
3170                                 thresh = delalloc_to_write;
3171                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
3172                                                  thresh);
3173                 }
3174
3175                 /* did the fill delalloc function already unlock and start
3176                  * the IO?
3177                  */
3178                 if (page_started) {
3179                         ret = 0;
3180                         /*
3181                          * we've unlocked the page, so we can't update
3182                          * the mapping's writeback index, just update
3183                          * nr_to_write.
3184                          */
3185                         wbc->nr_to_write -= nr_written;
3186                         goto done_unlocked;
3187                 }
3188         }
3189         if (tree->ops && tree->ops->writepage_start_hook) {
3190                 ret = tree->ops->writepage_start_hook(page, start,
3191                                                       page_end);
3192                 if (ret) {
3193                         /* Fixup worker will requeue */
3194                         if (ret == -EBUSY)
3195                                 wbc->pages_skipped++;
3196                         else
3197                                 redirty_page_for_writepage(wbc, page);
3198                         update_nr_written(page, wbc, nr_written);
3199                         unlock_page(page);
3200                         ret = 0;
3201                         goto done_unlocked;
3202                 }
3203         }
3204
3205         /*
3206          * we don't want to touch the inode after unlocking the page,
3207          * so we update the mapping writeback index now
3208          */
3209         update_nr_written(page, wbc, nr_written + 1);
3210
3211         end = page_end;
3212         if (last_byte <= start) {
3213                 if (tree->ops && tree->ops->writepage_end_io_hook)
3214                         tree->ops->writepage_end_io_hook(page, start,
3215                                                          page_end, NULL, 1);
3216                 goto done;
3217         }
3218
3219         blocksize = inode->i_sb->s_blocksize;
3220
3221         while (cur <= end) {
3222                 if (cur >= last_byte) {
3223                         if (tree->ops && tree->ops->writepage_end_io_hook)
3224                                 tree->ops->writepage_end_io_hook(page, cur,
3225                                                          page_end, NULL, 1);
3226                         break;
3227                 }
3228                 em = epd->get_extent(inode, page, pg_offset, cur,
3229                                      end - cur + 1, 1);
3230                 if (IS_ERR_OR_NULL(em)) {
3231                         SetPageError(page);
3232                         break;
3233                 }
3234
3235                 extent_offset = cur - em->start;
3236                 BUG_ON(extent_map_end(em) <= cur);
3237                 BUG_ON(end < cur);
3238                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3239                 iosize = ALIGN(iosize, blocksize);
3240                 sector = (em->block_start + extent_offset) >> 9;
3241                 bdev = em->bdev;
3242                 block_start = em->block_start;
3243                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3244                 free_extent_map(em);
3245                 em = NULL;
3246
3247                 /*
3248                  * compressed and inline extents are written through other
3249                  * paths in the FS
3250                  */
3251                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3252                     block_start == EXTENT_MAP_INLINE) {
3253                         /*
3254                          * end_io notification does not happen here for
3255                          * compressed extents
3256                          */
3257                         if (!compressed && tree->ops &&
3258                             tree->ops->writepage_end_io_hook)
3259                                 tree->ops->writepage_end_io_hook(page, cur,
3260                                                          cur + iosize - 1,
3261                                                          NULL, 1);
3262                         else if (compressed) {
3263                                 /* we don't want to end_page_writeback on
3264                                  * a compressed extent.  this happens
3265                                  * elsewhere
3266                                  */
3267                                 nr++;
3268                         }
3269
3270                         cur += iosize;
3271                         pg_offset += iosize;
3272                         continue;
3273                 }
3274                 /* leave this out until we have a page_mkwrite call */
3275                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
3276                                    EXTENT_DIRTY, 0, NULL)) {
3277                         cur = cur + iosize;
3278                         pg_offset += iosize;
3279                         continue;
3280                 }
3281
3282                 if (tree->ops && tree->ops->writepage_io_hook) {
3283                         ret = tree->ops->writepage_io_hook(page, cur,
3284                                                 cur + iosize - 1);
3285                 } else {
3286                         ret = 0;
3287                 }
3288                 if (ret) {
3289                         SetPageError(page);
3290                 } else {
3291                         unsigned long max_nr = end_index + 1;
3292
3293                         set_range_writeback(tree, cur, cur + iosize - 1);
3294                         if (!PageWriteback(page)) {
3295                                 printk(KERN_ERR "btrfs warning page %lu not "
3296                                        "writeback, cur %llu end %llu\n",
3297                                        page->index, cur, end);
3298                         }
3299
3300                         ret = submit_extent_page(write_flags, tree, page,
3301                                                  sector, iosize, pg_offset,
3302                                                  bdev, &epd->bio, max_nr,
3303                                                  end_bio_extent_writepage,
3304                                                  0, 0, 0);
3305                         if (ret)
3306                                 SetPageError(page);
3307                 }
3308                 cur = cur + iosize;
3309                 pg_offset += iosize;
3310                 nr++;
3311         }
3312 done:
3313         if (nr == 0) {
3314                 /* make sure the mapping tag for page dirty gets cleared */
3315                 set_page_writeback(page);
3316                 end_page_writeback(page);
3317         }
3318         unlock_page(page);
3319
3320 done_unlocked:
3321
3322         /* drop our reference on any cached states */
3323         free_extent_state(cached_state);
3324         return 0;
3325 }
3326
3327 static int eb_wait(void *word)
3328 {
3329         io_schedule();
3330         return 0;
3331 }
3332
3333 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3334 {
3335         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3336                     TASK_UNINTERRUPTIBLE);
3337 }
3338
3339 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3340                                      struct btrfs_fs_info *fs_info,
3341                                      struct extent_page_data *epd)
3342 {
3343         unsigned long i, num_pages;
3344         int flush = 0;
3345         int ret = 0;
3346
3347         if (!btrfs_try_tree_write_lock(eb)) {
3348                 flush = 1;
3349                 flush_write_bio(epd);
3350                 btrfs_tree_lock(eb);
3351         }
3352
3353         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3354                 btrfs_tree_unlock(eb);
3355                 if (!epd->sync_io)
3356                         return 0;
3357                 if (!flush) {
3358                         flush_write_bio(epd);
3359                         flush = 1;
3360                 }
3361                 while (1) {
3362                         wait_on_extent_buffer_writeback(eb);
3363                         btrfs_tree_lock(eb);
3364                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3365                                 break;
3366                         btrfs_tree_unlock(eb);
3367                 }
3368         }
3369
3370         /*
3371          * We need to do this to prevent races in people who check if the eb is
3372          * under IO since we can end up having no IO bits set for a short period
3373          * of time.
3374          */
3375         spin_lock(&eb->refs_lock);
3376         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3377                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3378                 spin_unlock(&eb->refs_lock);
3379                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3380                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3381                                      -eb->len,
3382                                      fs_info->dirty_metadata_batch);
3383                 ret = 1;
3384         } else {
3385                 spin_unlock(&eb->refs_lock);
3386         }
3387
3388         btrfs_tree_unlock(eb);
3389
3390         if (!ret)
3391                 return ret;
3392
3393         num_pages = num_extent_pages(eb->start, eb->len);
3394         for (i = 0; i < num_pages; i++) {
3395                 struct page *p = extent_buffer_page(eb, i);
3396
3397                 if (!trylock_page(p)) {
3398                         if (!flush) {
3399                                 flush_write_bio(epd);
3400                                 flush = 1;
3401                         }
3402                         lock_page(p);
3403                 }
3404         }
3405
3406         return ret;
3407 }
3408
3409 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3410 {
3411         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3412         smp_mb__after_clear_bit();
3413         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3414 }
3415
3416 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3417 {
3418         int uptodate = err == 0;
3419         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
3420         struct extent_buffer *eb;
3421         int done;
3422
3423         do {
3424                 struct page *page = bvec->bv_page;
3425
3426                 bvec--;
3427                 eb = (struct extent_buffer *)page->private;
3428                 BUG_ON(!eb);
3429                 done = atomic_dec_and_test(&eb->io_pages);
3430
3431                 if (!uptodate || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3432                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3433                         ClearPageUptodate(page);
3434                         SetPageError(page);
3435                 }
3436
3437                 end_page_writeback(page);
3438
3439                 if (!done)
3440                         continue;
3441
3442                 end_extent_buffer_writeback(eb);
3443         } while (bvec >= bio->bi_io_vec);
3444
3445         bio_put(bio);
3446
3447 }
3448
3449 static int write_one_eb(struct extent_buffer *eb,
3450                         struct btrfs_fs_info *fs_info,
3451                         struct writeback_control *wbc,
3452                         struct extent_page_data *epd)
3453 {
3454         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3455         u64 offset = eb->start;
3456         unsigned long i, num_pages;
3457         unsigned long bio_flags = 0;
3458         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3459         int ret = 0;
3460
3461         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3462         num_pages = num_extent_pages(eb->start, eb->len);
3463         atomic_set(&eb->io_pages, num_pages);
3464         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3465                 bio_flags = EXTENT_BIO_TREE_LOG;
3466
3467         for (i = 0; i < num_pages; i++) {
3468                 struct page *p = extent_buffer_page(eb, i);
3469
3470                 clear_page_dirty_for_io(p);
3471                 set_page_writeback(p);
3472                 ret = submit_extent_page(rw, eb->tree, p, offset >> 9,
3473                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3474                                          -1, end_bio_extent_buffer_writepage,
3475                                          0, epd->bio_flags, bio_flags);
3476                 epd->bio_flags = bio_flags;
3477                 if (ret) {
3478                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3479                         SetPageError(p);
3480                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3481                                 end_extent_buffer_writeback(eb);
3482                         ret = -EIO;
3483                         break;
3484                 }
3485                 offset += PAGE_CACHE_SIZE;
3486                 update_nr_written(p, wbc, 1);
3487                 unlock_page(p);
3488         }
3489
3490         if (unlikely(ret)) {
3491                 for (; i < num_pages; i++) {
3492                         struct page *p = extent_buffer_page(eb, i);
3493                         unlock_page(p);
3494                 }
3495         }
3496
3497         return ret;
3498 }
3499
3500 int btree_write_cache_pages(struct address_space *mapping,
3501                                    struct writeback_control *wbc)
3502 {
3503         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3504         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3505         struct extent_buffer *eb, *prev_eb = NULL;
3506         struct extent_page_data epd = {
3507                 .bio = NULL,
3508                 .tree = tree,
3509                 .extent_locked = 0,
3510                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3511                 .bio_flags = 0,
3512         };
3513         int ret = 0;
3514         int done = 0;
3515         int nr_to_write_done = 0;
3516         struct pagevec pvec;
3517         int nr_pages;
3518         pgoff_t index;
3519         pgoff_t end;            /* Inclusive */
3520         int scanned = 0;
3521         int tag;
3522
3523         pagevec_init(&pvec, 0);
3524         if (wbc->range_cyclic) {
3525                 index = mapping->writeback_index; /* Start from prev offset */
3526                 end = -1;
3527         } else {
3528                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3529                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3530                 scanned = 1;
3531         }
3532         if (wbc->sync_mode == WB_SYNC_ALL)
3533                 tag = PAGECACHE_TAG_TOWRITE;
3534         else
3535                 tag = PAGECACHE_TAG_DIRTY;
3536 retry:
3537         if (wbc->sync_mode == WB_SYNC_ALL)
3538                 tag_pages_for_writeback(mapping, index, end);
3539         while (!done && !nr_to_write_done && (index <= end) &&
3540                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3541                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3542                 unsigned i;
3543
3544                 scanned = 1;
3545                 for (i = 0; i < nr_pages; i++) {
3546                         struct page *page = pvec.pages[i];
3547
3548                         if (!PagePrivate(page))
3549                                 continue;
3550
3551                         if (!wbc->range_cyclic && page->index > end) {
3552                                 done = 1;
3553                                 break;
3554                         }
3555
3556                         spin_lock(&mapping->private_lock);
3557                         if (!PagePrivate(page)) {
3558                                 spin_unlock(&mapping->private_lock);
3559                                 continue;
3560                         }
3561
3562                         eb = (struct extent_buffer *)page->private;
3563
3564                         /*
3565                          * Shouldn't happen and normally this would be a BUG_ON
3566                          * but no sense in crashing the users box for something
3567                          * we can survive anyway.
3568                          */
3569                         if (WARN_ON(!eb)) {
3570                                 spin_unlock(&mapping->private_lock);
3571                                 continue;
3572                         }
3573
3574                         if (eb == prev_eb) {
3575                                 spin_unlock(&mapping->private_lock);
3576                                 continue;
3577                         }
3578
3579                         ret = atomic_inc_not_zero(&eb->refs);
3580                         spin_unlock(&mapping->private_lock);
3581                         if (!ret)
3582                                 continue;
3583
3584                         prev_eb = eb;
3585                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3586                         if (!ret) {
3587                                 free_extent_buffer(eb);
3588                                 continue;
3589                         }
3590
3591                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3592                         if (ret) {
3593                                 done = 1;
3594                                 free_extent_buffer(eb);
3595                                 break;
3596                         }
3597                         free_extent_buffer(eb);
3598
3599                         /*
3600                          * the filesystem may choose to bump up nr_to_write.
3601                          * We have to make sure to honor the new nr_to_write
3602                          * at any time
3603                          */
3604                         nr_to_write_done = wbc->nr_to_write <= 0;
3605                 }
3606                 pagevec_release(&pvec);
3607                 cond_resched();
3608         }
3609         if (!scanned && !done) {
3610                 /*
3611                  * We hit the last page and there is more work to be done: wrap
3612                  * back to the start of the file
3613                  */
3614                 scanned = 1;
3615                 index = 0;
3616                 goto retry;
3617         }
3618         flush_write_bio(&epd);
3619         return ret;
3620 }
3621
3622 /**
3623  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3624  * @mapping: address space structure to write
3625  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3626  * @writepage: function called for each page
3627  * @data: data passed to writepage function
3628  *
3629  * If a page is already under I/O, write_cache_pages() skips it, even
3630  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3631  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3632  * and msync() need to guarantee that all the data which was dirty at the time
3633  * the call was made get new I/O started against them.  If wbc->sync_mode is
3634  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3635  * existing IO to complete.
3636  */
3637 static int extent_write_cache_pages(struct extent_io_tree *tree,
3638                              struct address_space *mapping,
3639                              struct writeback_control *wbc,
3640                              writepage_t writepage, void *data,
3641                              void (*flush_fn)(void *))
3642 {
3643         struct inode *inode = mapping->host;
3644         int ret = 0;
3645         int done = 0;
3646         int nr_to_write_done = 0;
3647         struct pagevec pvec;
3648         int nr_pages;
3649         pgoff_t index;
3650         pgoff_t end;            /* Inclusive */
3651         int scanned = 0;
3652         int tag;
3653
3654         /*
3655          * We have to hold onto the inode so that ordered extents can do their
3656          * work when the IO finishes.  The alternative to this is failing to add
3657          * an ordered extent if the igrab() fails there and that is a huge pain
3658          * to deal with, so instead just hold onto the inode throughout the
3659          * writepages operation.  If it fails here we are freeing up the inode
3660          * anyway and we'd rather not waste our time writing out stuff that is
3661          * going to be truncated anyway.
3662          */
3663         if (!igrab(inode))
3664                 return 0;
3665
3666         pagevec_init(&pvec, 0);
3667         if (wbc->range_cyclic) {
3668                 index = mapping->writeback_index; /* Start from prev offset */
3669                 end = -1;
3670         } else {
3671                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3672                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3673                 scanned = 1;
3674         }
3675         if (wbc->sync_mode == WB_SYNC_ALL)
3676                 tag = PAGECACHE_TAG_TOWRITE;
3677         else
3678                 tag = PAGECACHE_TAG_DIRTY;
3679 retry:
3680         if (wbc->sync_mode == WB_SYNC_ALL)
3681                 tag_pages_for_writeback(mapping, index, end);
3682         while (!done && !nr_to_write_done && (index <= end) &&
3683                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3684                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3685                 unsigned i;
3686
3687                 scanned = 1;
3688                 for (i = 0; i < nr_pages; i++) {
3689                         struct page *page = pvec.pages[i];
3690
3691                         /*
3692                          * At this point we hold neither mapping->tree_lock nor
3693                          * lock on the page itself: the page may be truncated or
3694                          * invalidated (changing page->mapping to NULL), or even
3695                          * swizzled back from swapper_space to tmpfs file
3696                          * mapping
3697                          */
3698                         if (!trylock_page(page)) {
3699                                 flush_fn(data);
3700                                 lock_page(page);
3701                         }
3702
3703                         if (unlikely(page->mapping != mapping)) {
3704                                 unlock_page(page);
3705                                 continue;
3706                         }
3707
3708                         if (!wbc->range_cyclic && page->index > end) {
3709                                 done = 1;
3710                                 unlock_page(page);
3711                                 continue;
3712                         }
3713
3714                         if (wbc->sync_mode != WB_SYNC_NONE) {
3715                                 if (PageWriteback(page))
3716                                         flush_fn(data);
3717                                 wait_on_page_writeback(page);
3718                         }
3719
3720                         if (PageWriteback(page) ||
3721                             !clear_page_dirty_for_io(page)) {
3722                                 unlock_page(page);
3723                                 continue;
3724                         }
3725
3726                         ret = (*writepage)(page, wbc, data);
3727
3728                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3729                                 unlock_page(page);
3730                                 ret = 0;
3731                         }
3732                         if (ret)
3733                                 done = 1;
3734
3735                         /*
3736                          * the filesystem may choose to bump up nr_to_write.
3737                          * We have to make sure to honor the new nr_to_write
3738                          * at any time
3739                          */
3740                         nr_to_write_done = wbc->nr_to_write <= 0;
3741                 }
3742                 pagevec_release(&pvec);
3743                 cond_resched();
3744         }
3745         if (!scanned && !done) {
3746                 /*
3747                  * We hit the last page and there is more work to be done: wrap
3748                  * back to the start of the file
3749                  */
3750                 scanned = 1;
3751                 index = 0;
3752                 goto retry;
3753         }
3754         btrfs_add_delayed_iput(inode);
3755         return ret;
3756 }
3757
3758 static void flush_epd_write_bio(struct extent_page_data *epd)
3759 {
3760         if (epd->bio) {
3761                 int rw = WRITE;
3762                 int ret;
3763
3764                 if (epd->sync_io)
3765                         rw = WRITE_SYNC;
3766
3767                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3768                 BUG_ON(ret < 0); /* -ENOMEM */
3769                 epd->bio = NULL;
3770         }
3771 }
3772
3773 static noinline void flush_write_bio(void *data)
3774 {
3775         struct extent_page_data *epd = data;
3776         flush_epd_write_bio(epd);
3777 }
3778
3779 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3780                           get_extent_t *get_extent,
3781                           struct writeback_control *wbc)
3782 {
3783         int ret;
3784         struct extent_page_data epd = {
3785                 .bio = NULL,
3786                 .tree = tree,
3787                 .get_extent = get_extent,
3788                 .extent_locked = 0,
3789                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3790                 .bio_flags = 0,
3791         };
3792
3793         ret = __extent_writepage(page, wbc, &epd);
3794
3795         flush_epd_write_bio(&epd);
3796         return ret;
3797 }
3798
3799 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3800                               u64 start, u64 end, get_extent_t *get_extent,
3801                               int mode)
3802 {
3803         int ret = 0;
3804         struct address_space *mapping = inode->i_mapping;
3805         struct page *page;
3806         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3807                 PAGE_CACHE_SHIFT;
3808
3809         struct extent_page_data epd = {
3810                 .bio = NULL,
3811                 .tree = tree,
3812                 .get_extent = get_extent,
3813                 .extent_locked = 1,
3814                 .sync_io = mode == WB_SYNC_ALL,
3815                 .bio_flags = 0,
3816         };
3817         struct writeback_control wbc_writepages = {
3818                 .sync_mode      = mode,
3819                 .nr_to_write    = nr_pages * 2,
3820                 .range_start    = start,
3821                 .range_end      = end + 1,
3822         };
3823
3824         while (start <= end) {
3825                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3826                 if (clear_page_dirty_for_io(page))
3827                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3828                 else {
3829                         if (tree->ops && tree->ops->writepage_end_io_hook)
3830                                 tree->ops->writepage_end_io_hook(page, start,
3831                                                  start + PAGE_CACHE_SIZE - 1,
3832                                                  NULL, 1);
3833                         unlock_page(page);
3834                 }
3835                 page_cache_release(page);
3836                 start += PAGE_CACHE_SIZE;
3837         }
3838
3839         flush_epd_write_bio(&epd);
3840         return ret;
3841 }
3842
3843 int extent_writepages(struct extent_io_tree *tree,
3844                       struct address_space *mapping,
3845                       get_extent_t *get_extent,
3846                       struct writeback_control *wbc)
3847 {
3848         int ret = 0;
3849         struct extent_page_data epd = {
3850                 .bio = NULL,
3851                 .tree = tree,
3852                 .get_extent = get_extent,
3853                 .extent_locked = 0,
3854                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3855                 .bio_flags = 0,
3856         };
3857
3858         ret = extent_write_cache_pages(tree, mapping, wbc,
3859                                        __extent_writepage, &epd,
3860                                        flush_write_bio);
3861         flush_epd_write_bio(&epd);
3862         return ret;
3863 }
3864
3865 int extent_readpages(struct extent_io_tree *tree,
3866                      struct address_space *mapping,
3867                      struct list_head *pages, unsigned nr_pages,
3868                      get_extent_t get_extent)
3869 {
3870         struct bio *bio = NULL;
3871         unsigned page_idx;
3872         unsigned long bio_flags = 0;
3873         struct page *pagepool[16];
3874         struct page *page;
3875         struct extent_map *em_cached = NULL;
3876         int nr = 0;
3877
3878         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3879                 page = list_entry(pages->prev, struct page, lru);
3880
3881                 prefetchw(&page->flags);
3882                 list_del(&page->lru);
3883                 if (add_to_page_cache_lru(page, mapping,
3884                                         page->index, GFP_NOFS)) {
3885                         page_cache_release(page);
3886                         continue;
3887                 }
3888
3889                 pagepool[nr++] = page;
3890                 if (nr < ARRAY_SIZE(pagepool))
3891                         continue;
3892                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3893                                    &bio, 0, &bio_flags, READ);
3894                 nr = 0;
3895         }
3896         if (nr)
3897                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3898                                    &bio, 0, &bio_flags, READ);
3899
3900         if (em_cached)
3901                 free_extent_map(em_cached);
3902
3903         BUG_ON(!list_empty(pages));
3904         if (bio)
3905                 return submit_one_bio(READ, bio, 0, bio_flags);
3906         return 0;
3907 }
3908
3909 /*
3910  * basic invalidatepage code, this waits on any locked or writeback
3911  * ranges corresponding to the page, and then deletes any extent state
3912  * records from the tree
3913  */
3914 int extent_invalidatepage(struct extent_io_tree *tree,
3915                           struct page *page, unsigned long offset)
3916 {
3917         struct extent_state *cached_state = NULL;
3918         u64 start = page_offset(page);
3919         u64 end = start + PAGE_CACHE_SIZE - 1;
3920         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3921
3922         start += ALIGN(offset, blocksize);
3923         if (start > end)
3924                 return 0;
3925
3926         lock_extent_bits(tree, start, end, 0, &cached_state);
3927         wait_on_page_writeback(page);
3928         clear_extent_bit(tree, start, end,
3929                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3930                          EXTENT_DO_ACCOUNTING,
3931                          1, 1, &cached_state, GFP_NOFS);
3932         return 0;
3933 }
3934
3935 /*
3936  * a helper for releasepage, this tests for areas of the page that
3937  * are locked or under IO and drops the related state bits if it is safe
3938  * to drop the page.
3939  */
3940 static int try_release_extent_state(struct extent_map_tree *map,
3941                                     struct extent_io_tree *tree,
3942                                     struct page *page, gfp_t mask)
3943 {
3944         u64 start = page_offset(page);
3945         u64 end = start + PAGE_CACHE_SIZE - 1;
3946         int ret = 1;
3947
3948         if (test_range_bit(tree, start, end,
3949                            EXTENT_IOBITS, 0, NULL))
3950                 ret = 0;
3951         else {
3952                 if ((mask & GFP_NOFS) == GFP_NOFS)
3953                         mask = GFP_NOFS;
3954                 /*
3955                  * at this point we can safely clear everything except the
3956                  * locked bit and the nodatasum bit
3957                  */
3958                 ret = clear_extent_bit(tree, start, end,
3959                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3960                                  0, 0, NULL, mask);
3961
3962                 /* if clear_extent_bit failed for enomem reasons,
3963                  * we can't allow the release to continue.
3964                  */
3965                 if (ret < 0)
3966                         ret = 0;
3967                 else
3968                         ret = 1;
3969         }
3970         return ret;
3971 }
3972
3973 /*
3974  * a helper for releasepage.  As long as there are no locked extents
3975  * in the range corresponding to the page, both state records and extent
3976  * map records are removed
3977  */
3978 int try_release_extent_mapping(struct extent_map_tree *map,
3979                                struct extent_io_tree *tree, struct page *page,
3980                                gfp_t mask)
3981 {
3982         struct extent_map *em;
3983         u64 start = page_offset(page);
3984         u64 end = start + PAGE_CACHE_SIZE - 1;
3985
3986         if ((mask & __GFP_WAIT) &&
3987             page->mapping->host->i_size > 16 * 1024 * 1024) {
3988                 u64 len;
3989                 while (start <= end) {
3990                         len = end - start + 1;
3991                         write_lock(&map->lock);
3992                         em = lookup_extent_mapping(map, start, len);
3993                         if (!em) {
3994                                 write_unlock(&map->lock);
3995                                 break;
3996                         }
3997                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3998                             em->start != start) {
3999                                 write_unlock(&map->lock);
4000                                 free_extent_map(em);
4001                                 break;
4002                         }
4003                         if (!test_range_bit(tree, em->start,
4004                                             extent_map_end(em) - 1,
4005                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4006                                             0, NULL)) {
4007                                 remove_extent_mapping(map, em);
4008                                 /* once for the rb tree */
4009                                 free_extent_map(em);
4010                         }
4011                         start = extent_map_end(em);
4012                         write_unlock(&map->lock);
4013
4014                         /* once for us */
4015                         free_extent_map(em);
4016                 }
4017         }
4018         return try_release_extent_state(map, tree, page, mask);
4019 }
4020
4021 /*
4022  * helper function for fiemap, which doesn't want to see any holes.
4023  * This maps until we find something past 'last'
4024  */
4025 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4026                                                 u64 offset,
4027                                                 u64 last,
4028                                                 get_extent_t *get_extent)
4029 {
4030         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4031         struct extent_map *em;
4032         u64 len;
4033
4034         if (offset >= last)
4035                 return NULL;
4036
4037         while (1) {
4038                 len = last - offset;
4039                 if (len == 0)
4040                         break;
4041                 len = ALIGN(len, sectorsize);
4042                 em = get_extent(inode, NULL, 0, offset, len, 0);
4043                 if (IS_ERR_OR_NULL(em))
4044                         return em;
4045
4046                 /* if this isn't a hole return it */
4047                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4048                     em->block_start != EXTENT_MAP_HOLE) {
4049                         return em;
4050                 }
4051
4052                 /* this is a hole, advance to the next extent */
4053                 offset = extent_map_end(em);
4054                 free_extent_map(em);
4055                 if (offset >= last)
4056                         break;
4057         }
4058         return NULL;
4059 }
4060
4061 static noinline int count_ext_ref(u64 inum, u64 offset, u64 root_id, void *ctx)
4062 {
4063         unsigned long cnt = *((unsigned long *)ctx);
4064
4065         cnt++;
4066         *((unsigned long *)ctx) = cnt;
4067
4068         /* Now we're sure that the extent is shared. */
4069         if (cnt > 1)
4070                 return 1;
4071         return 0;
4072 }
4073
4074 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4075                 __u64 start, __u64 len, get_extent_t *get_extent)
4076 {
4077         int ret = 0;
4078         u64 off = start;
4079         u64 max = start + len;
4080         u32 flags = 0;
4081         u32 found_type;
4082         u64 last;
4083         u64 last_for_get_extent = 0;
4084         u64 disko = 0;
4085         u64 isize = i_size_read(inode);
4086         struct btrfs_key found_key;
4087         struct extent_map *em = NULL;
4088         struct extent_state *cached_state = NULL;
4089         struct btrfs_path *path;
4090         struct btrfs_file_extent_item *item;
4091         int end = 0;
4092         u64 em_start = 0;
4093         u64 em_len = 0;
4094         u64 em_end = 0;
4095         unsigned long emflags;
4096
4097         if (len == 0)
4098                 return -EINVAL;
4099
4100         path = btrfs_alloc_path();
4101         if (!path)
4102                 return -ENOMEM;
4103         path->leave_spinning = 1;
4104
4105         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
4106         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
4107
4108         /*
4109          * lookup the last file extent.  We're not using i_size here
4110          * because there might be preallocation past i_size
4111          */
4112         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
4113                                        path, btrfs_ino(inode), -1, 0);
4114         if (ret < 0) {
4115                 btrfs_free_path(path);
4116                 return ret;
4117         }
4118         WARN_ON(!ret);
4119         path->slots[0]--;
4120         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
4121                               struct btrfs_file_extent_item);
4122         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4123         found_type = btrfs_key_type(&found_key);
4124
4125         /* No extents, but there might be delalloc bits */
4126         if (found_key.objectid != btrfs_ino(inode) ||
4127             found_type != BTRFS_EXTENT_DATA_KEY) {
4128                 /* have to trust i_size as the end */
4129                 last = (u64)-1;
4130                 last_for_get_extent = isize;
4131         } else {
4132                 /*
4133                  * remember the start of the last extent.  There are a
4134                  * bunch of different factors that go into the length of the
4135                  * extent, so its much less complex to remember where it started
4136                  */
4137                 last = found_key.offset;
4138                 last_for_get_extent = last + 1;
4139         }
4140         btrfs_release_path(path);
4141
4142         /*
4143          * we might have some extents allocated but more delalloc past those
4144          * extents.  so, we trust isize unless the start of the last extent is
4145          * beyond isize
4146          */
4147         if (last < isize) {
4148                 last = (u64)-1;
4149                 last_for_get_extent = isize;
4150         }
4151
4152         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4153                          &cached_state);
4154
4155         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4156                                    get_extent);
4157         if (!em)
4158                 goto out;
4159         if (IS_ERR(em)) {
4160                 ret = PTR_ERR(em);
4161                 goto out;
4162         }
4163
4164         while (!end) {
4165                 u64 offset_in_extent = 0;
4166
4167                 /* break if the extent we found is outside the range */
4168                 if (em->start >= max || extent_map_end(em) < off)
4169                         break;
4170
4171                 /*
4172                  * get_extent may return an extent that starts before our
4173                  * requested range.  We have to make sure the ranges
4174                  * we return to fiemap always move forward and don't
4175                  * overlap, so adjust the offsets here
4176                  */
4177                 em_start = max(em->start, off);
4178
4179                 /*
4180                  * record the offset from the start of the extent
4181                  * for adjusting the disk offset below.  Only do this if the
4182                  * extent isn't compressed since our in ram offset may be past
4183                  * what we have actually allocated on disk.
4184                  */
4185                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4186                         offset_in_extent = em_start - em->start;
4187                 em_end = extent_map_end(em);
4188                 em_len = em_end - em_start;
4189                 emflags = em->flags;
4190                 disko = 0;
4191                 flags = 0;
4192
4193                 /*
4194                  * bump off for our next call to get_extent
4195                  */
4196                 off = extent_map_end(em);
4197                 if (off >= max)
4198                         end = 1;
4199
4200                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4201                         end = 1;
4202                         flags |= FIEMAP_EXTENT_LAST;
4203                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4204                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4205                                   FIEMAP_EXTENT_NOT_ALIGNED);
4206                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4207                         flags |= (FIEMAP_EXTENT_DELALLOC |
4208                                   FIEMAP_EXTENT_UNKNOWN);
4209                 } else {
4210                         unsigned long ref_cnt = 0;
4211
4212                         disko = em->block_start + offset_in_extent;
4213
4214                         /*
4215                          * As btrfs supports shared space, this information
4216                          * can be exported to userspace tools via
4217                          * flag FIEMAP_EXTENT_SHARED.
4218                          */
4219                         ret = iterate_inodes_from_logical(
4220                                         em->block_start,
4221                                         BTRFS_I(inode)->root->fs_info,
4222                                         path, count_ext_ref, &ref_cnt);
4223                         if (ret < 0 && ret != -ENOENT)
4224                                 goto out_free;
4225
4226                         if (ref_cnt > 1)
4227                                 flags |= FIEMAP_EXTENT_SHARED;
4228                 }
4229                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4230                         flags |= FIEMAP_EXTENT_ENCODED;
4231
4232                 free_extent_map(em);
4233                 em = NULL;
4234                 if ((em_start >= last) || em_len == (u64)-1 ||
4235                    (last == (u64)-1 && isize <= em_end)) {
4236                         flags |= FIEMAP_EXTENT_LAST;
4237                         end = 1;
4238                 }
4239
4240                 /* now scan forward to see if this is really the last extent. */
4241                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4242                                            get_extent);
4243                 if (IS_ERR(em)) {
4244                         ret = PTR_ERR(em);
4245                         goto out;
4246                 }
4247                 if (!em) {
4248                         flags |= FIEMAP_EXTENT_LAST;
4249                         end = 1;
4250                 }
4251                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4252                                               em_len, flags);
4253                 if (ret)
4254                         goto out_free;
4255         }
4256 out_free:
4257         free_extent_map(em);
4258 out:
4259         btrfs_free_path(path);
4260         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4261                              &cached_state, GFP_NOFS);
4262         return ret;
4263 }
4264
4265 static void __free_extent_buffer(struct extent_buffer *eb)
4266 {
4267         btrfs_leak_debug_del(&eb->leak_list);
4268         kmem_cache_free(extent_buffer_cache, eb);
4269 }
4270
4271 static int extent_buffer_under_io(struct extent_buffer *eb)
4272 {
4273         return (atomic_read(&eb->io_pages) ||
4274                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4275                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4276 }
4277
4278 /*
4279  * Helper for releasing extent buffer page.
4280  */
4281 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4282                                                 unsigned long start_idx)
4283 {
4284         unsigned long index;
4285         unsigned long num_pages;
4286         struct page *page;
4287         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4288
4289         BUG_ON(extent_buffer_under_io(eb));
4290
4291         num_pages = num_extent_pages(eb->start, eb->len);
4292         index = start_idx + num_pages;
4293         if (start_idx >= index)
4294                 return;
4295
4296         do {
4297                 index--;
4298                 page = extent_buffer_page(eb, index);
4299                 if (page && mapped) {
4300                         spin_lock(&page->mapping->private_lock);
4301                         /*
4302                          * We do this since we'll remove the pages after we've
4303                          * removed the eb from the radix tree, so we could race
4304                          * and have this page now attached to the new eb.  So
4305                          * only clear page_private if it's still connected to
4306                          * this eb.
4307                          */
4308                         if (PagePrivate(page) &&
4309                             page->private == (unsigned long)eb) {
4310                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4311                                 BUG_ON(PageDirty(page));
4312                                 BUG_ON(PageWriteback(page));
4313                                 /*
4314                                  * We need to make sure we haven't be attached
4315                                  * to a new eb.
4316                                  */
4317                                 ClearPagePrivate(page);
4318                                 set_page_private(page, 0);
4319                                 /* One for the page private */
4320                                 page_cache_release(page);
4321                         }
4322                         spin_unlock(&page->mapping->private_lock);
4323
4324                 }
4325                 if (page) {
4326                         /* One for when we alloced the page */
4327                         page_cache_release(page);
4328                 }
4329         } while (index != start_idx);
4330 }
4331
4332 /*
4333  * Helper for releasing the extent buffer.
4334  */
4335 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4336 {
4337         btrfs_release_extent_buffer_page(eb, 0);
4338         __free_extent_buffer(eb);
4339 }
4340
4341 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
4342                                                    u64 start,
4343                                                    unsigned long len,
4344                                                    gfp_t mask)
4345 {
4346         struct extent_buffer *eb = NULL;
4347
4348         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4349         if (eb == NULL)
4350                 return NULL;
4351         eb->start = start;
4352         eb->len = len;
4353         eb->tree = tree;
4354         eb->bflags = 0;
4355         rwlock_init(&eb->lock);
4356         atomic_set(&eb->write_locks, 0);
4357         atomic_set(&eb->read_locks, 0);
4358         atomic_set(&eb->blocking_readers, 0);
4359         atomic_set(&eb->blocking_writers, 0);
4360         atomic_set(&eb->spinning_readers, 0);
4361         atomic_set(&eb->spinning_writers, 0);
4362         eb->lock_nested = 0;
4363         init_waitqueue_head(&eb->write_lock_wq);
4364         init_waitqueue_head(&eb->read_lock_wq);
4365
4366         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4367
4368         spin_lock_init(&eb->refs_lock);
4369         atomic_set(&eb->refs, 1);
4370         atomic_set(&eb->io_pages, 0);
4371
4372         /*
4373          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4374          */
4375         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4376                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4377         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4378
4379         return eb;
4380 }
4381
4382 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4383 {
4384         unsigned long i;
4385         struct page *p;
4386         struct extent_buffer *new;
4387         unsigned long num_pages = num_extent_pages(src->start, src->len);
4388
4389         new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_NOFS);
4390         if (new == NULL)
4391                 return NULL;
4392
4393         for (i = 0; i < num_pages; i++) {
4394                 p = alloc_page(GFP_NOFS);
4395                 if (!p) {
4396                         btrfs_release_extent_buffer(new);
4397                         return NULL;
4398                 }
4399                 attach_extent_buffer_page(new, p);
4400                 WARN_ON(PageDirty(p));
4401                 SetPageUptodate(p);
4402                 new->pages[i] = p;
4403         }
4404
4405         copy_extent_buffer(new, src, 0, 0, src->len);
4406         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4407         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4408
4409         return new;
4410 }
4411
4412 struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4413 {
4414         struct extent_buffer *eb;
4415         unsigned long num_pages = num_extent_pages(0, len);
4416         unsigned long i;
4417
4418         eb = __alloc_extent_buffer(NULL, start, len, GFP_NOFS);
4419         if (!eb)
4420                 return NULL;
4421
4422         for (i = 0; i < num_pages; i++) {
4423                 eb->pages[i] = alloc_page(GFP_NOFS);
4424                 if (!eb->pages[i])
4425                         goto err;
4426         }
4427         set_extent_buffer_uptodate(eb);
4428         btrfs_set_header_nritems(eb, 0);
4429         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4430
4431         return eb;
4432 err:
4433         for (; i > 0; i--)
4434                 __free_page(eb->pages[i - 1]);
4435         __free_extent_buffer(eb);
4436         return NULL;
4437 }
4438
4439 static void check_buffer_tree_ref(struct extent_buffer *eb)
4440 {
4441         int refs;
4442         /* the ref bit is tricky.  We have to make sure it is set
4443          * if we have the buffer dirty.   Otherwise the
4444          * code to free a buffer can end up dropping a dirty
4445          * page
4446          *
4447          * Once the ref bit is set, it won't go away while the
4448          * buffer is dirty or in writeback, and it also won't
4449          * go away while we have the reference count on the
4450          * eb bumped.
4451          *
4452          * We can't just set the ref bit without bumping the
4453          * ref on the eb because free_extent_buffer might
4454          * see the ref bit and try to clear it.  If this happens
4455          * free_extent_buffer might end up dropping our original
4456          * ref by mistake and freeing the page before we are able
4457          * to add one more ref.
4458          *
4459          * So bump the ref count first, then set the bit.  If someone
4460          * beat us to it, drop the ref we added.
4461          */
4462         refs = atomic_read(&eb->refs);
4463         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4464                 return;
4465
4466         spin_lock(&eb->refs_lock);
4467         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4468                 atomic_inc(&eb->refs);
4469         spin_unlock(&eb->refs_lock);
4470 }
4471
4472 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4473 {
4474         unsigned long num_pages, i;
4475
4476         check_buffer_tree_ref(eb);
4477
4478         num_pages = num_extent_pages(eb->start, eb->len);
4479         for (i = 0; i < num_pages; i++) {
4480                 struct page *p = extent_buffer_page(eb, i);
4481                 mark_page_accessed(p);
4482         }
4483 }
4484
4485 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
4486                                                         u64 start)
4487 {
4488         struct extent_buffer *eb;
4489
4490         rcu_read_lock();
4491         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4492         if (eb && atomic_inc_not_zero(&eb->refs)) {
4493                 rcu_read_unlock();
4494                 mark_extent_buffer_accessed(eb);
4495                 return eb;
4496         }
4497         rcu_read_unlock();
4498
4499         return NULL;
4500 }
4501
4502 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
4503                                           u64 start, unsigned long len)
4504 {
4505         unsigned long num_pages = num_extent_pages(start, len);
4506         unsigned long i;
4507         unsigned long index = start >> PAGE_CACHE_SHIFT;
4508         struct extent_buffer *eb;
4509         struct extent_buffer *exists = NULL;
4510         struct page *p;
4511         struct address_space *mapping = tree->mapping;
4512         int uptodate = 1;
4513         int ret;
4514
4515
4516         eb = find_extent_buffer(tree, start);
4517         if (eb)
4518                 return eb;
4519
4520         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
4521         if (!eb)
4522                 return NULL;
4523
4524         for (i = 0; i < num_pages; i++, index++) {
4525                 p = find_or_create_page(mapping, index, GFP_NOFS);
4526                 if (!p)
4527                         goto free_eb;
4528
4529                 spin_lock(&mapping->private_lock);
4530                 if (PagePrivate(p)) {
4531                         /*
4532                          * We could have already allocated an eb for this page
4533                          * and attached one so lets see if we can get a ref on
4534                          * the existing eb, and if we can we know it's good and
4535                          * we can just return that one, else we know we can just
4536                          * overwrite page->private.
4537                          */
4538                         exists = (struct extent_buffer *)p->private;
4539                         if (atomic_inc_not_zero(&exists->refs)) {
4540                                 spin_unlock(&mapping->private_lock);
4541                                 unlock_page(p);
4542                                 page_cache_release(p);
4543                                 mark_extent_buffer_accessed(exists);
4544                                 goto free_eb;
4545                         }
4546
4547                         /*
4548                          * Do this so attach doesn't complain and we need to
4549                          * drop the ref the old guy had.
4550                          */
4551                         ClearPagePrivate(p);
4552                         WARN_ON(PageDirty(p));
4553                         page_cache_release(p);
4554                 }
4555                 attach_extent_buffer_page(eb, p);
4556                 spin_unlock(&mapping->private_lock);
4557                 WARN_ON(PageDirty(p));
4558                 mark_page_accessed(p);
4559                 eb->pages[i] = p;
4560                 if (!PageUptodate(p))
4561                         uptodate = 0;
4562
4563                 /*
4564                  * see below about how we avoid a nasty race with release page
4565                  * and why we unlock later
4566                  */
4567         }
4568         if (uptodate)
4569                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4570 again:
4571         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4572         if (ret)
4573                 goto free_eb;
4574
4575         spin_lock(&tree->buffer_lock);
4576         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
4577         spin_unlock(&tree->buffer_lock);
4578         radix_tree_preload_end();
4579         if (ret == -EEXIST) {
4580                 exists = find_extent_buffer(tree, start);
4581                 if (exists)
4582                         goto free_eb;
4583                 else
4584                         goto again;
4585         }
4586         /* add one reference for the tree */
4587         check_buffer_tree_ref(eb);
4588
4589         /*
4590          * there is a race where release page may have
4591          * tried to find this extent buffer in the radix
4592          * but failed.  It will tell the VM it is safe to
4593          * reclaim the, and it will clear the page private bit.
4594          * We must make sure to set the page private bit properly
4595          * after the extent buffer is in the radix tree so
4596          * it doesn't get lost
4597          */
4598         SetPageChecked(eb->pages[0]);
4599         for (i = 1; i < num_pages; i++) {
4600                 p = extent_buffer_page(eb, i);
4601                 ClearPageChecked(p);
4602                 unlock_page(p);
4603         }
4604         unlock_page(eb->pages[0]);
4605         return eb;
4606
4607 free_eb:
4608         for (i = 0; i < num_pages; i++) {
4609                 if (eb->pages[i])
4610                         unlock_page(eb->pages[i]);
4611         }
4612
4613         WARN_ON(!atomic_dec_and_test(&eb->refs));
4614         btrfs_release_extent_buffer(eb);
4615         return exists;
4616 }
4617
4618 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4619 {
4620         struct extent_buffer *eb =
4621                         container_of(head, struct extent_buffer, rcu_head);
4622
4623         __free_extent_buffer(eb);
4624 }
4625
4626 /* Expects to have eb->eb_lock already held */
4627 static int release_extent_buffer(struct extent_buffer *eb)
4628 {
4629         WARN_ON(atomic_read(&eb->refs) == 0);
4630         if (atomic_dec_and_test(&eb->refs)) {
4631                 if (test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags)) {
4632                         spin_unlock(&eb->refs_lock);
4633                 } else {
4634                         struct extent_io_tree *tree = eb->tree;
4635
4636                         spin_unlock(&eb->refs_lock);
4637
4638                         spin_lock(&tree->buffer_lock);
4639                         radix_tree_delete(&tree->buffer,
4640                                           eb->start >> PAGE_CACHE_SHIFT);
4641                         spin_unlock(&tree->buffer_lock);
4642                 }
4643
4644                 /* Should be safe to release our pages at this point */
4645                 btrfs_release_extent_buffer_page(eb, 0);
4646                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4647                 return 1;
4648         }
4649         spin_unlock(&eb->refs_lock);
4650
4651         return 0;
4652 }
4653
4654 void free_extent_buffer(struct extent_buffer *eb)
4655 {
4656         int refs;
4657         int old;
4658         if (!eb)
4659                 return;
4660
4661         while (1) {
4662                 refs = atomic_read(&eb->refs);
4663                 if (refs <= 3)
4664                         break;
4665                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
4666                 if (old == refs)
4667                         return;
4668         }
4669
4670         spin_lock(&eb->refs_lock);
4671         if (atomic_read(&eb->refs) == 2 &&
4672             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4673                 atomic_dec(&eb->refs);
4674
4675         if (atomic_read(&eb->refs) == 2 &&
4676             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4677             !extent_buffer_under_io(eb) &&
4678             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4679                 atomic_dec(&eb->refs);
4680
4681         /*
4682          * I know this is terrible, but it's temporary until we stop tracking
4683          * the uptodate bits and such for the extent buffers.
4684          */
4685         release_extent_buffer(eb);
4686 }
4687
4688 void free_extent_buffer_stale(struct extent_buffer *eb)
4689 {
4690         if (!eb)
4691                 return;
4692
4693         spin_lock(&eb->refs_lock);
4694         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4695
4696         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4697             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4698                 atomic_dec(&eb->refs);
4699         release_extent_buffer(eb);
4700 }
4701
4702 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4703 {
4704         unsigned long i;
4705         unsigned long num_pages;
4706         struct page *page;
4707
4708         num_pages = num_extent_pages(eb->start, eb->len);
4709
4710         for (i = 0; i < num_pages; i++) {
4711                 page = extent_buffer_page(eb, i);
4712                 if (!PageDirty(page))
4713                         continue;
4714
4715                 lock_page(page);
4716                 WARN_ON(!PagePrivate(page));
4717
4718                 clear_page_dirty_for_io(page);
4719                 spin_lock_irq(&page->mapping->tree_lock);
4720                 if (!PageDirty(page)) {
4721                         radix_tree_tag_clear(&page->mapping->page_tree,
4722                                                 page_index(page),
4723                                                 PAGECACHE_TAG_DIRTY);
4724                 }
4725                 spin_unlock_irq(&page->mapping->tree_lock);
4726                 ClearPageError(page);
4727                 unlock_page(page);
4728         }
4729         WARN_ON(atomic_read(&eb->refs) == 0);
4730 }
4731
4732 int set_extent_buffer_dirty(struct extent_buffer *eb)
4733 {
4734         unsigned long i;
4735         unsigned long num_pages;
4736         int was_dirty = 0;
4737
4738         check_buffer_tree_ref(eb);
4739
4740         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4741
4742         num_pages = num_extent_pages(eb->start, eb->len);
4743         WARN_ON(atomic_read(&eb->refs) == 0);
4744         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4745
4746         for (i = 0; i < num_pages; i++)
4747                 set_page_dirty(extent_buffer_page(eb, i));
4748         return was_dirty;
4749 }
4750
4751 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4752 {
4753         unsigned long i;
4754         struct page *page;
4755         unsigned long num_pages;
4756
4757         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4758         num_pages = num_extent_pages(eb->start, eb->len);
4759         for (i = 0; i < num_pages; i++) {
4760                 page = extent_buffer_page(eb, i);
4761                 if (page)
4762                         ClearPageUptodate(page);
4763         }
4764         return 0;
4765 }
4766
4767 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4768 {
4769         unsigned long i;
4770         struct page *page;
4771         unsigned long num_pages;
4772
4773         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4774         num_pages = num_extent_pages(eb->start, eb->len);
4775         for (i = 0; i < num_pages; i++) {
4776                 page = extent_buffer_page(eb, i);
4777                 SetPageUptodate(page);
4778         }
4779         return 0;
4780 }
4781
4782 int extent_buffer_uptodate(struct extent_buffer *eb)
4783 {
4784         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4785 }
4786
4787 int read_extent_buffer_pages(struct extent_io_tree *tree,
4788                              struct extent_buffer *eb, u64 start, int wait,
4789                              get_extent_t *get_extent, int mirror_num)
4790 {
4791         unsigned long i;
4792         unsigned long start_i;
4793         struct page *page;
4794         int err;
4795         int ret = 0;
4796         int locked_pages = 0;
4797         int all_uptodate = 1;
4798         unsigned long num_pages;
4799         unsigned long num_reads = 0;
4800         struct bio *bio = NULL;
4801         unsigned long bio_flags = 0;
4802
4803         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4804                 return 0;
4805
4806         if (start) {
4807                 WARN_ON(start < eb->start);
4808                 start_i = (start >> PAGE_CACHE_SHIFT) -
4809                         (eb->start >> PAGE_CACHE_SHIFT);
4810         } else {
4811                 start_i = 0;
4812         }
4813
4814         num_pages = num_extent_pages(eb->start, eb->len);
4815         for (i = start_i; i < num_pages; i++) {
4816                 page = extent_buffer_page(eb, i);
4817                 if (wait == WAIT_NONE) {
4818                         if (!trylock_page(page))
4819                                 goto unlock_exit;
4820                 } else {
4821                         lock_page(page);
4822                 }
4823                 locked_pages++;
4824                 if (!PageUptodate(page)) {
4825                         num_reads++;
4826                         all_uptodate = 0;
4827                 }
4828         }
4829         if (all_uptodate) {
4830                 if (start_i == 0)
4831                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4832                 goto unlock_exit;
4833         }
4834
4835         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4836         eb->read_mirror = 0;
4837         atomic_set(&eb->io_pages, num_reads);
4838         for (i = start_i; i < num_pages; i++) {
4839                 page = extent_buffer_page(eb, i);
4840                 if (!PageUptodate(page)) {
4841                         ClearPageError(page);
4842                         err = __extent_read_full_page(tree, page,
4843                                                       get_extent, &bio,
4844                                                       mirror_num, &bio_flags,
4845                                                       READ | REQ_META);
4846                         if (err)
4847                                 ret = err;
4848                 } else {
4849                         unlock_page(page);
4850                 }
4851         }
4852
4853         if (bio) {
4854                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
4855                                      bio_flags);
4856                 if (err)
4857                         return err;
4858         }
4859
4860         if (ret || wait != WAIT_COMPLETE)
4861                 return ret;
4862
4863         for (i = start_i; i < num_pages; i++) {
4864                 page = extent_buffer_page(eb, i);
4865                 wait_on_page_locked(page);
4866                 if (!PageUptodate(page))
4867                         ret = -EIO;
4868         }
4869
4870         return ret;
4871
4872 unlock_exit:
4873         i = start_i;
4874         while (locked_pages > 0) {
4875                 page = extent_buffer_page(eb, i);
4876                 i++;
4877                 unlock_page(page);
4878                 locked_pages--;
4879         }
4880         return ret;
4881 }
4882
4883 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4884                         unsigned long start,
4885                         unsigned long len)
4886 {
4887         size_t cur;
4888         size_t offset;
4889         struct page *page;
4890         char *kaddr;
4891         char *dst = (char *)dstv;
4892         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4893         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4894
4895         WARN_ON(start > eb->len);
4896         WARN_ON(start + len > eb->start + eb->len);
4897
4898         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
4899
4900         while (len > 0) {
4901                 page = extent_buffer_page(eb, i);
4902
4903                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4904                 kaddr = page_address(page);
4905                 memcpy(dst, kaddr + offset, cur);
4906
4907                 dst += cur;
4908                 len -= cur;
4909                 offset = 0;
4910                 i++;
4911         }
4912 }
4913
4914 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4915                                unsigned long min_len, char **map,
4916                                unsigned long *map_start,
4917                                unsigned long *map_len)
4918 {
4919         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4920         char *kaddr;
4921         struct page *p;
4922         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4923         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4924         unsigned long end_i = (start_offset + start + min_len - 1) >>
4925                 PAGE_CACHE_SHIFT;
4926
4927         if (i != end_i)
4928                 return -EINVAL;
4929
4930         if (i == 0) {
4931                 offset = start_offset;
4932                 *map_start = 0;
4933         } else {
4934                 offset = 0;
4935                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4936         }
4937
4938         if (start + min_len > eb->len) {
4939                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4940                        "wanted %lu %lu\n",
4941                        eb->start, eb->len, start, min_len);
4942                 return -EINVAL;
4943         }
4944
4945         p = extent_buffer_page(eb, i);
4946         kaddr = page_address(p);
4947         *map = kaddr + offset;
4948         *map_len = PAGE_CACHE_SIZE - offset;
4949         return 0;
4950 }
4951
4952 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4953                           unsigned long start,
4954                           unsigned long len)
4955 {
4956         size_t cur;
4957         size_t offset;
4958         struct page *page;
4959         char *kaddr;
4960         char *ptr = (char *)ptrv;
4961         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4962         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4963         int ret = 0;
4964
4965         WARN_ON(start > eb->len);
4966         WARN_ON(start + len > eb->start + eb->len);
4967
4968         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
4969
4970         while (len > 0) {
4971                 page = extent_buffer_page(eb, i);
4972
4973                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4974
4975                 kaddr = page_address(page);
4976                 ret = memcmp(ptr, kaddr + offset, cur);
4977                 if (ret)
4978                         break;
4979
4980                 ptr += cur;
4981                 len -= cur;
4982                 offset = 0;
4983                 i++;
4984         }
4985         return ret;
4986 }
4987
4988 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4989                          unsigned long start, unsigned long len)
4990 {
4991         size_t cur;
4992         size_t offset;
4993         struct page *page;
4994         char *kaddr;
4995         char *src = (char *)srcv;
4996         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4997         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4998
4999         WARN_ON(start > eb->len);
5000         WARN_ON(start + len > eb->start + eb->len);
5001
5002         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5003
5004         while (len > 0) {
5005                 page = extent_buffer_page(eb, i);
5006                 WARN_ON(!PageUptodate(page));
5007
5008                 cur = min(len, PAGE_CACHE_SIZE - offset);
5009                 kaddr = page_address(page);
5010                 memcpy(kaddr + offset, src, cur);
5011
5012                 src += cur;
5013                 len -= cur;
5014                 offset = 0;
5015                 i++;
5016         }
5017 }
5018
5019 void memset_extent_buffer(struct extent_buffer *eb, char c,
5020                           unsigned long start, unsigned long len)
5021 {
5022         size_t cur;
5023         size_t offset;
5024         struct page *page;
5025         char *kaddr;
5026         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5027         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5028
5029         WARN_ON(start > eb->len);
5030         WARN_ON(start + len > eb->start + eb->len);
5031
5032         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5033
5034         while (len > 0) {
5035                 page = extent_buffer_page(eb, i);
5036                 WARN_ON(!PageUptodate(page));
5037
5038                 cur = min(len, PAGE_CACHE_SIZE - offset);
5039                 kaddr = page_address(page);
5040                 memset(kaddr + offset, c, cur);
5041
5042                 len -= cur;
5043                 offset = 0;
5044                 i++;
5045         }
5046 }
5047
5048 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5049                         unsigned long dst_offset, unsigned long src_offset,
5050                         unsigned long len)
5051 {
5052         u64 dst_len = dst->len;
5053         size_t cur;
5054         size_t offset;
5055         struct page *page;
5056         char *kaddr;
5057         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5058         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5059
5060         WARN_ON(src->len != dst_len);
5061
5062         offset = (start_offset + dst_offset) &
5063                 (PAGE_CACHE_SIZE - 1);
5064
5065         while (len > 0) {
5066                 page = extent_buffer_page(dst, i);
5067                 WARN_ON(!PageUptodate(page));
5068
5069                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5070
5071                 kaddr = page_address(page);
5072                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5073
5074                 src_offset += cur;
5075                 len -= cur;
5076                 offset = 0;
5077                 i++;
5078         }
5079 }
5080
5081 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5082 {
5083         unsigned long distance = (src > dst) ? src - dst : dst - src;
5084         return distance < len;
5085 }
5086
5087 static void copy_pages(struct page *dst_page, struct page *src_page,
5088                        unsigned long dst_off, unsigned long src_off,
5089                        unsigned long len)
5090 {
5091         char *dst_kaddr = page_address(dst_page);
5092         char *src_kaddr;
5093         int must_memmove = 0;
5094
5095         if (dst_page != src_page) {
5096                 src_kaddr = page_address(src_page);
5097         } else {
5098                 src_kaddr = dst_kaddr;
5099                 if (areas_overlap(src_off, dst_off, len))
5100                         must_memmove = 1;
5101         }
5102
5103         if (must_memmove)
5104                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5105         else
5106                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5107 }
5108
5109 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5110                            unsigned long src_offset, unsigned long len)
5111 {
5112         size_t cur;
5113         size_t dst_off_in_page;
5114         size_t src_off_in_page;
5115         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5116         unsigned long dst_i;
5117         unsigned long src_i;
5118
5119         if (src_offset + len > dst->len) {
5120                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
5121                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5122                 BUG_ON(1);
5123         }
5124         if (dst_offset + len > dst->len) {
5125                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
5126                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5127                 BUG_ON(1);
5128         }
5129
5130         while (len > 0) {
5131                 dst_off_in_page = (start_offset + dst_offset) &
5132                         (PAGE_CACHE_SIZE - 1);
5133                 src_off_in_page = (start_offset + src_offset) &
5134                         (PAGE_CACHE_SIZE - 1);
5135
5136                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5137                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5138
5139                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5140                                                src_off_in_page));
5141                 cur = min_t(unsigned long, cur,
5142                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5143
5144                 copy_pages(extent_buffer_page(dst, dst_i),
5145                            extent_buffer_page(dst, src_i),
5146                            dst_off_in_page, src_off_in_page, cur);
5147
5148                 src_offset += cur;
5149                 dst_offset += cur;
5150                 len -= cur;
5151         }
5152 }
5153
5154 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5155                            unsigned long src_offset, unsigned long len)
5156 {
5157         size_t cur;
5158         size_t dst_off_in_page;
5159         size_t src_off_in_page;
5160         unsigned long dst_end = dst_offset + len - 1;
5161         unsigned long src_end = src_offset + len - 1;
5162         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5163         unsigned long dst_i;
5164         unsigned long src_i;
5165
5166         if (src_offset + len > dst->len) {
5167                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
5168                        "len %lu len %lu\n", src_offset, len, dst->len);
5169                 BUG_ON(1);
5170         }
5171         if (dst_offset + len > dst->len) {
5172                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
5173                        "len %lu len %lu\n", dst_offset, len, dst->len);
5174                 BUG_ON(1);
5175         }
5176         if (dst_offset < src_offset) {
5177                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5178                 return;
5179         }
5180         while (len > 0) {
5181                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5182                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5183
5184                 dst_off_in_page = (start_offset + dst_end) &
5185                         (PAGE_CACHE_SIZE - 1);
5186                 src_off_in_page = (start_offset + src_end) &
5187                         (PAGE_CACHE_SIZE - 1);
5188
5189                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5190                 cur = min(cur, dst_off_in_page + 1);
5191                 copy_pages(extent_buffer_page(dst, dst_i),
5192                            extent_buffer_page(dst, src_i),
5193                            dst_off_in_page - cur + 1,
5194                            src_off_in_page - cur + 1, cur);
5195
5196                 dst_end -= cur;
5197                 src_end -= cur;
5198                 len -= cur;
5199         }
5200 }
5201
5202 int try_release_extent_buffer(struct page *page)
5203 {
5204         struct extent_buffer *eb;
5205
5206         /*
5207          * We need to make sure noboody is attaching this page to an eb right
5208          * now.
5209          */
5210         spin_lock(&page->mapping->private_lock);
5211         if (!PagePrivate(page)) {
5212                 spin_unlock(&page->mapping->private_lock);
5213                 return 1;
5214         }
5215
5216         eb = (struct extent_buffer *)page->private;
5217         BUG_ON(!eb);
5218
5219         /*
5220          * This is a little awful but should be ok, we need to make sure that
5221          * the eb doesn't disappear out from under us while we're looking at
5222          * this page.
5223          */
5224         spin_lock(&eb->refs_lock);
5225         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5226                 spin_unlock(&eb->refs_lock);
5227                 spin_unlock(&page->mapping->private_lock);
5228                 return 0;
5229         }
5230         spin_unlock(&page->mapping->private_lock);
5231
5232         /*
5233          * If tree ref isn't set then we know the ref on this eb is a real ref,
5234          * so just return, this page will likely be freed soon anyway.
5235          */
5236         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5237                 spin_unlock(&eb->refs_lock);
5238                 return 0;
5239         }
5240
5241         return release_extent_buffer(eb);
5242 }