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 "compat.h"
  17 #include "ctree.h"
  18 #include "btrfs_inode.h"
  19 #include "volumes.h"
  20 #include "check-integrity.h"
  21 #include "locking.h"
  22 #include "rcu-string.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 noinline u64 find_lock_delalloc_range(struct inode *inode,
1601                                              struct extent_io_tree *tree,
1602                                              struct page *locked_page,
1603                                              u64 *start, u64 *end,
1604                                              u64 max_bytes)
1605 {
1606         u64 delalloc_start;
1607         u64 delalloc_end;
1608         u64 found;
1609         struct extent_state *cached_state = NULL;
1610         int ret;
1611         int loops = 0;
1612
1613 again:
1614         /* step one, find a bunch of delalloc bytes starting at start */
1615         delalloc_start = *start;
1616         delalloc_end = 0;
1617         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1618                                     max_bytes, &cached_state);
1619         if (!found || delalloc_end <= *start) {
1620                 *start = delalloc_start;
1621                 *end = delalloc_end;
1622                 free_extent_state(cached_state);
1623                 return 0;
1624         }
1625
1626         /*
1627          * start comes from the offset of locked_page.  We have to lock
1628          * pages in order, so we can't process delalloc bytes before
1629          * locked_page
1630          */
1631         if (delalloc_start < *start)
1632                 delalloc_start = *start;
1633
1634         /*
1635          * make sure to limit the number of pages we try to lock down
1636          */
1637         if (delalloc_end + 1 - delalloc_start > max_bytes)
1638                 delalloc_end = delalloc_start + max_bytes - 1;
1639
1640         /* step two, lock all the pages after the page that has start */
1641         ret = lock_delalloc_pages(inode, locked_page,
1642                                   delalloc_start, delalloc_end);
1643         if (ret == -EAGAIN) {
1644                 /* some of the pages are gone, lets avoid looping by
1645                  * shortening the size of the delalloc range we're searching
1646                  */
1647                 free_extent_state(cached_state);
1648                 if (!loops) {
1649                         max_bytes = PAGE_CACHE_SIZE;
1650                         loops = 1;
1651                         goto again;
1652                 } else {
1653                         found = 0;
1654                         goto out_failed;
1655                 }
1656         }
1657         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1658
1659         /* step three, lock the state bits for the whole range */
1660         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1661
1662         /* then test to make sure it is all still delalloc */
1663         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1664                              EXTENT_DELALLOC, 1, cached_state);
1665         if (!ret) {
1666                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1667                                      &cached_state, GFP_NOFS);
1668                 __unlock_for_delalloc(inode, locked_page,
1669                               delalloc_start, delalloc_end);
1670                 cond_resched();
1671                 goto again;
1672         }
1673         free_extent_state(cached_state);
1674         *start = delalloc_start;
1675         *end = delalloc_end;
1676 out_failed:
1677         return found;
1678 }
1679
1680 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1681                                  struct page *locked_page,
1682                                  unsigned long clear_bits,
1683                                  unsigned long page_ops)
1684 {
1685         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1686         int ret;
1687         struct page *pages[16];
1688         unsigned long index = start >> PAGE_CACHE_SHIFT;
1689         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1690         unsigned long nr_pages = end_index - index + 1;
1691         int i;
1692
1693         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1694         if (page_ops == 0)
1695                 return 0;
1696
1697         while (nr_pages > 0) {
1698                 ret = find_get_pages_contig(inode->i_mapping, index,
1699                                      min_t(unsigned long,
1700                                      nr_pages, ARRAY_SIZE(pages)), pages);
1701                 for (i = 0; i < ret; i++) {
1702
1703                         if (page_ops & PAGE_SET_PRIVATE2)
1704                                 SetPagePrivate2(pages[i]);
1705
1706                         if (pages[i] == locked_page) {
1707                                 page_cache_release(pages[i]);
1708                                 continue;
1709                         }
1710                         if (page_ops & PAGE_CLEAR_DIRTY)
1711                                 clear_page_dirty_for_io(pages[i]);
1712                         if (page_ops & PAGE_SET_WRITEBACK)
1713                                 set_page_writeback(pages[i]);
1714                         if (page_ops & PAGE_END_WRITEBACK)
1715                                 end_page_writeback(pages[i]);
1716                         if (page_ops & PAGE_UNLOCK)
1717                                 unlock_page(pages[i]);
1718                         page_cache_release(pages[i]);
1719                 }
1720                 nr_pages -= ret;
1721                 index += ret;
1722                 cond_resched();
1723         }
1724         return 0;
1725 }
1726
1727 /*
1728  * count the number of bytes in the tree that have a given bit(s)
1729  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1730  * cached.  The total number found is returned.
1731  */
1732 u64 count_range_bits(struct extent_io_tree *tree,
1733                      u64 *start, u64 search_end, u64 max_bytes,
1734                      unsigned long bits, int contig)
1735 {
1736         struct rb_node *node;
1737         struct extent_state *state;
1738         u64 cur_start = *start;
1739         u64 total_bytes = 0;
1740         u64 last = 0;
1741         int found = 0;
1742
1743         if (search_end <= cur_start) {
1744                 WARN_ON(1);
1745                 return 0;
1746         }
1747
1748         spin_lock(&tree->lock);
1749         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1750                 total_bytes = tree->dirty_bytes;
1751                 goto out;
1752         }
1753         /*
1754          * this search will find all the extents that end after
1755          * our range starts.
1756          */
1757         node = tree_search(tree, cur_start);
1758         if (!node)
1759                 goto out;
1760
1761         while (1) {
1762                 state = rb_entry(node, struct extent_state, rb_node);
1763                 if (state->start > search_end)
1764                         break;
1765                 if (contig && found && state->start > last + 1)
1766                         break;
1767                 if (state->end >= cur_start && (state->state & bits) == bits) {
1768                         total_bytes += min(search_end, state->end) + 1 -
1769                                        max(cur_start, state->start);
1770                         if (total_bytes >= max_bytes)
1771                                 break;
1772                         if (!found) {
1773                                 *start = max(cur_start, state->start);
1774                                 found = 1;
1775                         }
1776                         last = state->end;
1777                 } else if (contig && found) {
1778                         break;
1779                 }
1780                 node = rb_next(node);
1781                 if (!node)
1782                         break;
1783         }
1784 out:
1785         spin_unlock(&tree->lock);
1786         return total_bytes;
1787 }
1788
1789 /*
1790  * set the private field for a given byte offset in the tree.  If there isn't
1791  * an extent_state there already, this does nothing.
1792  */
1793 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1794 {
1795         struct rb_node *node;
1796         struct extent_state *state;
1797         int ret = 0;
1798
1799         spin_lock(&tree->lock);
1800         /*
1801          * this search will find all the extents that end after
1802          * our range starts.
1803          */
1804         node = tree_search(tree, start);
1805         if (!node) {
1806                 ret = -ENOENT;
1807                 goto out;
1808         }
1809         state = rb_entry(node, struct extent_state, rb_node);
1810         if (state->start != start) {
1811                 ret = -ENOENT;
1812                 goto out;
1813         }
1814         state->private = private;
1815 out:
1816         spin_unlock(&tree->lock);
1817         return ret;
1818 }
1819
1820 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1821 {
1822         struct rb_node *node;
1823         struct extent_state *state;
1824         int ret = 0;
1825
1826         spin_lock(&tree->lock);
1827         /*
1828          * this search will find all the extents that end after
1829          * our range starts.
1830          */
1831         node = tree_search(tree, start);
1832         if (!node) {
1833                 ret = -ENOENT;
1834                 goto out;
1835         }
1836         state = rb_entry(node, struct extent_state, rb_node);
1837         if (state->start != start) {
1838                 ret = -ENOENT;
1839                 goto out;
1840         }
1841         *private = state->private;
1842 out:
1843         spin_unlock(&tree->lock);
1844         return ret;
1845 }
1846
1847 /*
1848  * searches a range in the state tree for a given mask.
1849  * If 'filled' == 1, this returns 1 only if every extent in the tree
1850  * has the bits set.  Otherwise, 1 is returned if any bit in the
1851  * range is found set.
1852  */
1853 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1854                    unsigned long bits, int filled, struct extent_state *cached)
1855 {
1856         struct extent_state *state = NULL;
1857         struct rb_node *node;
1858         int bitset = 0;
1859
1860         spin_lock(&tree->lock);
1861         if (cached && cached->tree && cached->start <= start &&
1862             cached->end > start)
1863                 node = &cached->rb_node;
1864         else
1865                 node = tree_search(tree, start);
1866         while (node && start <= end) {
1867                 state = rb_entry(node, struct extent_state, rb_node);
1868
1869                 if (filled && state->start > start) {
1870                         bitset = 0;
1871                         break;
1872                 }
1873
1874                 if (state->start > end)
1875                         break;
1876
1877                 if (state->state & bits) {
1878                         bitset = 1;
1879                         if (!filled)
1880                                 break;
1881                 } else if (filled) {
1882                         bitset = 0;
1883                         break;
1884                 }
1885
1886                 if (state->end == (u64)-1)
1887                         break;
1888
1889                 start = state->end + 1;
1890                 if (start > end)
1891                         break;
1892                 node = rb_next(node);
1893                 if (!node) {
1894                         if (filled)
1895                                 bitset = 0;
1896                         break;
1897                 }
1898         }
1899         spin_unlock(&tree->lock);
1900         return bitset;
1901 }
1902
1903 /*
1904  * helper function to set a given page up to date if all the
1905  * extents in the tree for that page are up to date
1906  */
1907 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1908 {
1909         u64 start = page_offset(page);
1910         u64 end = start + PAGE_CACHE_SIZE - 1;
1911         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1912                 SetPageUptodate(page);
1913 }
1914
1915 /*
1916  * When IO fails, either with EIO or csum verification fails, we
1917  * try other mirrors that might have a good copy of the data.  This
1918  * io_failure_record is used to record state as we go through all the
1919  * mirrors.  If another mirror has good data, the page is set up to date
1920  * and things continue.  If a good mirror can't be found, the original
1921  * bio end_io callback is called to indicate things have failed.
1922  */
1923 struct io_failure_record {
1924         struct page *page;
1925         u64 start;
1926         u64 len;
1927         u64 logical;
1928         unsigned long bio_flags;
1929         int this_mirror;
1930         int failed_mirror;
1931         int in_validation;
1932 };
1933
1934 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1935                                 int did_repair)
1936 {
1937         int ret;
1938         int err = 0;
1939         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1940
1941         set_state_private(failure_tree, rec->start, 0);
1942         ret = clear_extent_bits(failure_tree, rec->start,
1943                                 rec->start + rec->len - 1,
1944                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1945         if (ret)
1946                 err = ret;
1947
1948         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1949                                 rec->start + rec->len - 1,
1950                                 EXTENT_DAMAGED, GFP_NOFS);
1951         if (ret && !err)
1952                 err = ret;
1953
1954         kfree(rec);
1955         return err;
1956 }
1957
1958 static void repair_io_failure_callback(struct bio *bio, int err)
1959 {
1960         complete(bio->bi_private);
1961 }
1962
1963 /*
1964  * this bypasses the standard btrfs submit functions deliberately, as
1965  * the standard behavior is to write all copies in a raid setup. here we only
1966  * want to write the one bad copy. so we do the mapping for ourselves and issue
1967  * submit_bio directly.
1968  * to avoid any synchronization issues, wait for the data after writing, which
1969  * actually prevents the read that triggered the error from finishing.
1970  * currently, there can be no more than two copies of every data bit. thus,
1971  * exactly one rewrite is required.
1972  */
1973 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 start,
1974                         u64 length, u64 logical, struct page *page,
1975                         int mirror_num)
1976 {
1977         struct bio *bio;
1978         struct btrfs_device *dev;
1979         DECLARE_COMPLETION_ONSTACK(compl);
1980         u64 map_length = 0;
1981         u64 sector;
1982         struct btrfs_bio *bbio = NULL;
1983         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
1984         int ret;
1985
1986         BUG_ON(!mirror_num);
1987
1988         /* we can't repair anything in raid56 yet */
1989         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
1990                 return 0;
1991
1992         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1993         if (!bio)
1994                 return -EIO;
1995         bio->bi_private = &compl;
1996         bio->bi_end_io = repair_io_failure_callback;
1997         bio->bi_size = 0;
1998         map_length = length;
1999
2000         ret = btrfs_map_block(fs_info, WRITE, logical,
2001                               &map_length, &bbio, mirror_num);
2002         if (ret) {
2003                 bio_put(bio);
2004                 return -EIO;
2005         }
2006         BUG_ON(mirror_num != bbio->mirror_num);
2007         sector = bbio->stripes[mirror_num-1].physical >> 9;
2008         bio->bi_sector = sector;
2009         dev = bbio->stripes[mirror_num-1].dev;
2010         kfree(bbio);
2011         if (!dev || !dev->bdev || !dev->writeable) {
2012                 bio_put(bio);
2013                 return -EIO;
2014         }
2015         bio->bi_bdev = dev->bdev;
2016         bio_add_page(bio, page, length, start - page_offset(page));
2017         btrfsic_submit_bio(WRITE_SYNC, bio);
2018         wait_for_completion(&compl);
2019
2020         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2021                 /* try to remap that extent elsewhere? */
2022                 bio_put(bio);
2023                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2024                 return -EIO;
2025         }
2026
2027         printk_ratelimited_in_rcu(KERN_INFO "btrfs read error corrected: ino %lu off %llu "
2028                       "(dev %s sector %llu)\n", page->mapping->host->i_ino,
2029                       start, rcu_str_deref(dev->name), sector);
2030
2031         bio_put(bio);
2032         return 0;
2033 }
2034
2035 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2036                          int mirror_num)
2037 {
2038         u64 start = eb->start;
2039         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2040         int ret = 0;
2041
2042         for (i = 0; i < num_pages; i++) {
2043                 struct page *p = extent_buffer_page(eb, i);
2044                 ret = repair_io_failure(root->fs_info, start, PAGE_CACHE_SIZE,
2045                                         start, p, mirror_num);
2046                 if (ret)
2047                         break;
2048                 start += PAGE_CACHE_SIZE;
2049         }
2050
2051         return ret;
2052 }
2053
2054 /*
2055  * each time an IO finishes, we do a fast check in the IO failure tree
2056  * to see if we need to process or clean up an io_failure_record
2057  */
2058 static int clean_io_failure(u64 start, struct page *page)
2059 {
2060         u64 private;
2061         u64 private_failure;
2062         struct io_failure_record *failrec;
2063         struct btrfs_fs_info *fs_info;
2064         struct extent_state *state;
2065         int num_copies;
2066         int did_repair = 0;
2067         int ret;
2068         struct inode *inode = page->mapping->host;
2069
2070         private = 0;
2071         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2072                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2073         if (!ret)
2074                 return 0;
2075
2076         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2077                                 &private_failure);
2078         if (ret)
2079                 return 0;
2080
2081         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2082         BUG_ON(!failrec->this_mirror);
2083
2084         if (failrec->in_validation) {
2085                 /* there was no real error, just free the record */
2086                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2087                          failrec->start);
2088                 did_repair = 1;
2089                 goto out;
2090         }
2091
2092         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2093         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2094                                             failrec->start,
2095                                             EXTENT_LOCKED);
2096         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2097
2098         if (state && state->start <= failrec->start &&
2099             state->end >= failrec->start + failrec->len - 1) {
2100                 fs_info = BTRFS_I(inode)->root->fs_info;
2101                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2102                                               failrec->len);
2103                 if (num_copies > 1)  {
2104                         ret = repair_io_failure(fs_info, start, failrec->len,
2105                                                 failrec->logical, page,
2106                                                 failrec->failed_mirror);
2107                         did_repair = !ret;
2108                 }
2109                 ret = 0;
2110         }
2111
2112 out:
2113         if (!ret)
2114                 ret = free_io_failure(inode, failrec, did_repair);
2115
2116         return ret;
2117 }
2118
2119 /*
2120  * this is a generic handler for readpage errors (default
2121  * readpage_io_failed_hook). if other copies exist, read those and write back
2122  * good data to the failed position. does not investigate in remapping the
2123  * failed extent elsewhere, hoping the device will be smart enough to do this as
2124  * needed
2125  */
2126
2127 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2128                               struct page *page, u64 start, u64 end,
2129                               int failed_mirror)
2130 {
2131         struct io_failure_record *failrec = NULL;
2132         u64 private;
2133         struct extent_map *em;
2134         struct inode *inode = page->mapping->host;
2135         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2136         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2137         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2138         struct bio *bio;
2139         struct btrfs_io_bio *btrfs_failed_bio;
2140         struct btrfs_io_bio *btrfs_bio;
2141         int num_copies;
2142         int ret;
2143         int read_mode;
2144         u64 logical;
2145
2146         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2147
2148         ret = get_state_private(failure_tree, start, &private);
2149         if (ret) {
2150                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2151                 if (!failrec)
2152                         return -ENOMEM;
2153                 failrec->start = start;
2154                 failrec->len = end - start + 1;
2155                 failrec->this_mirror = 0;
2156                 failrec->bio_flags = 0;
2157                 failrec->in_validation = 0;
2158
2159                 read_lock(&em_tree->lock);
2160                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2161                 if (!em) {
2162                         read_unlock(&em_tree->lock);
2163                         kfree(failrec);
2164                         return -EIO;
2165                 }
2166
2167                 if (em->start > start || em->start + em->len < start) {
2168                         free_extent_map(em);
2169                         em = NULL;
2170                 }
2171                 read_unlock(&em_tree->lock);
2172
2173                 if (!em) {
2174                         kfree(failrec);
2175                         return -EIO;
2176                 }
2177                 logical = start - em->start;
2178                 logical = em->block_start + logical;
2179                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2180                         logical = em->block_start;
2181                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2182                         extent_set_compress_type(&failrec->bio_flags,
2183                                                  em->compress_type);
2184                 }
2185                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2186                          "len=%llu\n", logical, start, failrec->len);
2187                 failrec->logical = logical;
2188                 free_extent_map(em);
2189
2190                 /* set the bits in the private failure tree */
2191                 ret = set_extent_bits(failure_tree, start, end,
2192                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2193                 if (ret >= 0)
2194                         ret = set_state_private(failure_tree, start,
2195                                                 (u64)(unsigned long)failrec);
2196                 /* set the bits in the inode's tree */
2197                 if (ret >= 0)
2198                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2199                                                 GFP_NOFS);
2200                 if (ret < 0) {
2201                         kfree(failrec);
2202                         return ret;
2203                 }
2204         } else {
2205                 failrec = (struct io_failure_record *)(unsigned long)private;
2206                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2207                          "start=%llu, len=%llu, validation=%d\n",
2208                          failrec->logical, failrec->start, failrec->len,
2209                          failrec->in_validation);
2210                 /*
2211                  * when data can be on disk more than twice, add to failrec here
2212                  * (e.g. with a list for failed_mirror) to make
2213                  * clean_io_failure() clean all those errors at once.
2214                  */
2215         }
2216         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2217                                       failrec->logical, failrec->len);
2218         if (num_copies == 1) {
2219                 /*
2220                  * we only have a single copy of the data, so don't bother with
2221                  * all the retry and error correction code that follows. no
2222                  * matter what the error is, it is very likely to persist.
2223                  */
2224                 pr_debug("bio_readpage_error: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2225                          num_copies, failrec->this_mirror, failed_mirror);
2226                 free_io_failure(inode, failrec, 0);
2227                 return -EIO;
2228         }
2229
2230         /*
2231          * there are two premises:
2232          *      a) deliver good data to the caller
2233          *      b) correct the bad sectors on disk
2234          */
2235         if (failed_bio->bi_vcnt > 1) {
2236                 /*
2237                  * to fulfill b), we need to know the exact failing sectors, as
2238                  * we don't want to rewrite any more than the failed ones. thus,
2239                  * we need separate read requests for the failed bio
2240                  *
2241                  * if the following BUG_ON triggers, our validation request got
2242                  * merged. we need separate requests for our algorithm to work.
2243                  */
2244                 BUG_ON(failrec->in_validation);
2245                 failrec->in_validation = 1;
2246                 failrec->this_mirror = failed_mirror;
2247                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2248         } else {
2249                 /*
2250                  * we're ready to fulfill a) and b) alongside. get a good copy
2251                  * of the failed sector and if we succeed, we have setup
2252                  * everything for repair_io_failure to do the rest for us.
2253                  */
2254                 if (failrec->in_validation) {
2255                         BUG_ON(failrec->this_mirror != failed_mirror);
2256                         failrec->in_validation = 0;
2257                         failrec->this_mirror = 0;
2258                 }
2259                 failrec->failed_mirror = failed_mirror;
2260                 failrec->this_mirror++;
2261                 if (failrec->this_mirror == failed_mirror)
2262                         failrec->this_mirror++;
2263                 read_mode = READ_SYNC;
2264         }
2265
2266         if (failrec->this_mirror > num_copies) {
2267                 pr_debug("bio_readpage_error: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2268                          num_copies, failrec->this_mirror, failed_mirror);
2269                 free_io_failure(inode, failrec, 0);
2270                 return -EIO;
2271         }
2272
2273         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2274         if (!bio) {
2275                 free_io_failure(inode, failrec, 0);
2276                 return -EIO;
2277         }
2278         bio->bi_end_io = failed_bio->bi_end_io;
2279         bio->bi_sector = failrec->logical >> 9;
2280         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2281         bio->bi_size = 0;
2282
2283         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2284         if (btrfs_failed_bio->csum) {
2285                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2286                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2287
2288                 btrfs_bio = btrfs_io_bio(bio);
2289                 btrfs_bio->csum = btrfs_bio->csum_inline;
2290                 phy_offset >>= inode->i_sb->s_blocksize_bits;
2291                 phy_offset *= csum_size;
2292                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + phy_offset,
2293                        csum_size);
2294         }
2295
2296         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2297
2298         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2299                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2300                  failrec->this_mirror, num_copies, failrec->in_validation);
2301
2302         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2303                                          failrec->this_mirror,
2304                                          failrec->bio_flags, 0);
2305         return ret;
2306 }
2307
2308 /* lots and lots of room for performance fixes in the end_bio funcs */
2309
2310 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2311 {
2312         int uptodate = (err == 0);
2313         struct extent_io_tree *tree;
2314         int ret;
2315
2316         tree = &BTRFS_I(page->mapping->host)->io_tree;
2317
2318         if (tree->ops && tree->ops->writepage_end_io_hook) {
2319                 ret = tree->ops->writepage_end_io_hook(page, start,
2320                                                end, NULL, uptodate);
2321                 if (ret)
2322                         uptodate = 0;
2323         }
2324
2325         if (!uptodate) {
2326                 ClearPageUptodate(page);
2327                 SetPageError(page);
2328         }
2329         return 0;
2330 }
2331
2332 /*
2333  * after a writepage IO is done, we need to:
2334  * clear the uptodate bits on error
2335  * clear the writeback bits in the extent tree for this IO
2336  * end_page_writeback if the page has no more pending IO
2337  *
2338  * Scheduling is not allowed, so the extent state tree is expected
2339  * to have one and only one object corresponding to this IO.
2340  */
2341 static void end_bio_extent_writepage(struct bio *bio, int err)
2342 {
2343         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2344         struct extent_io_tree *tree;
2345         u64 start;
2346         u64 end;
2347
2348         do {
2349                 struct page *page = bvec->bv_page;
2350                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2351
2352                 /* We always issue full-page reads, but if some block
2353                  * in a page fails to read, blk_update_request() will
2354                  * advance bv_offset and adjust bv_len to compensate.
2355                  * Print a warning for nonzero offsets, and an error
2356                  * if they don't add up to a full page.  */
2357                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE)
2358                         printk("%s page write in btrfs with offset %u and length %u\n",
2359                                bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE
2360                                ? KERN_ERR "partial" : KERN_INFO "incomplete",
2361                                bvec->bv_offset, bvec->bv_len);
2362
2363                 start = page_offset(page);
2364                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2365
2366                 if (--bvec >= bio->bi_io_vec)
2367                         prefetchw(&bvec->bv_page->flags);
2368
2369                 if (end_extent_writepage(page, err, start, end))
2370                         continue;
2371
2372                 end_page_writeback(page);
2373         } while (bvec >= bio->bi_io_vec);
2374
2375         bio_put(bio);
2376 }
2377
2378 static void
2379 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2380                               int uptodate)
2381 {
2382         struct extent_state *cached = NULL;
2383         u64 end = start + len - 1;
2384
2385         if (uptodate && tree->track_uptodate)
2386                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2387         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2388 }
2389
2390 /*
2391  * after a readpage IO is done, we need to:
2392  * clear the uptodate bits on error
2393  * set the uptodate bits if things worked
2394  * set the page up to date if all extents in the tree are uptodate
2395  * clear the lock bit in the extent tree
2396  * unlock the page if there are no other extents locked for it
2397  *
2398  * Scheduling is not allowed, so the extent state tree is expected
2399  * to have one and only one object corresponding to this IO.
2400  */
2401 static void end_bio_extent_readpage(struct bio *bio, int err)
2402 {
2403         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2404         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2405         struct bio_vec *bvec = bio->bi_io_vec;
2406         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2407         struct extent_io_tree *tree;
2408         u64 offset = 0;
2409         u64 start;
2410         u64 end;
2411         u64 len;
2412         u64 extent_start = 0;
2413         u64 extent_len = 0;
2414         int mirror;
2415         int ret;
2416
2417         if (err)
2418                 uptodate = 0;
2419
2420         do {
2421                 struct page *page = bvec->bv_page;
2422                 struct inode *inode = page->mapping->host;
2423
2424                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2425                          "mirror=%lu\n", (u64)bio->bi_sector, err,
2426                          io_bio->mirror_num);
2427                 tree = &BTRFS_I(inode)->io_tree;
2428
2429                 /* We always issue full-page reads, but if some block
2430                  * in a page fails to read, blk_update_request() will
2431                  * advance bv_offset and adjust bv_len to compensate.
2432                  * Print a warning for nonzero offsets, and an error
2433                  * if they don't add up to a full page.  */
2434                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE)
2435                         printk("%s page read in btrfs with offset %u and length %u\n",
2436                                bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE
2437                                ? KERN_ERR "partial" : KERN_INFO "incomplete",
2438                                bvec->bv_offset, bvec->bv_len);
2439
2440                 start = page_offset(page);
2441                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2442                 len = bvec->bv_len;
2443
2444                 if (++bvec <= bvec_end)
2445                         prefetchw(&bvec->bv_page->flags);
2446
2447                 mirror = io_bio->mirror_num;
2448                 if (likely(uptodate && tree->ops &&
2449                            tree->ops->readpage_end_io_hook)) {
2450                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2451                                                               page, start, end,
2452                                                               mirror);
2453                         if (ret)
2454                                 uptodate = 0;
2455                         else
2456                                 clean_io_failure(start, page);
2457                 }
2458
2459                 if (likely(uptodate))
2460                         goto readpage_ok;
2461
2462                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2463                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2464                         if (!ret && !err &&
2465                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2466                                 uptodate = 1;
2467                 } else {
2468                         /*
2469                          * The generic bio_readpage_error handles errors the
2470                          * following way: If possible, new read requests are
2471                          * created and submitted and will end up in
2472                          * end_bio_extent_readpage as well (if we're lucky, not
2473                          * in the !uptodate case). In that case it returns 0 and
2474                          * we just go on with the next page in our bio. If it
2475                          * can't handle the error it will return -EIO and we
2476                          * remain responsible for that page.
2477                          */
2478                         ret = bio_readpage_error(bio, offset, page, start, end,
2479                                                  mirror);
2480                         if (ret == 0) {
2481                                 uptodate =
2482                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2483                                 if (err)
2484                                         uptodate = 0;
2485                                 continue;
2486                         }
2487                 }
2488 readpage_ok:
2489                 if (likely(uptodate)) {
2490                         loff_t i_size = i_size_read(inode);
2491                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2492                         unsigned offset;
2493
2494                         /* Zero out the end if this page straddles i_size */
2495                         offset = i_size & (PAGE_CACHE_SIZE-1);
2496                         if (page->index == end_index && offset)
2497                                 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2498                         SetPageUptodate(page);
2499                 } else {
2500                         ClearPageUptodate(page);
2501                         SetPageError(page);
2502                 }
2503                 unlock_page(page);
2504                 offset += len;
2505
2506                 if (unlikely(!uptodate)) {
2507                         if (extent_len) {
2508                                 endio_readpage_release_extent(tree,
2509                                                               extent_start,
2510                                                               extent_len, 1);
2511                                 extent_start = 0;
2512                                 extent_len = 0;
2513                         }
2514                         endio_readpage_release_extent(tree, start,
2515                                                       end - start + 1, 0);
2516                 } else if (!extent_len) {
2517                         extent_start = start;
2518                         extent_len = end + 1 - start;
2519                 } else if (extent_start + extent_len == start) {
2520                         extent_len += end + 1 - start;
2521                 } else {
2522                         endio_readpage_release_extent(tree, extent_start,
2523                                                       extent_len, uptodate);
2524                         extent_start = start;
2525                         extent_len = end + 1 - start;
2526                 }
2527         } while (bvec <= bvec_end);
2528
2529         if (extent_len)
2530                 endio_readpage_release_extent(tree, extent_start, extent_len,
2531                                               uptodate);
2532         if (io_bio->end_io)
2533                 io_bio->end_io(io_bio, err);
2534         bio_put(bio);
2535 }
2536
2537 /*
2538  * this allocates from the btrfs_bioset.  We're returning a bio right now
2539  * but you can call btrfs_io_bio for the appropriate container_of magic
2540  */
2541 struct bio *
2542 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2543                 gfp_t gfp_flags)
2544 {
2545         struct btrfs_io_bio *btrfs_bio;
2546         struct bio *bio;
2547
2548         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2549
2550         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2551                 while (!bio && (nr_vecs /= 2)) {
2552                         bio = bio_alloc_bioset(gfp_flags,
2553                                                nr_vecs, btrfs_bioset);
2554                 }
2555         }
2556
2557         if (bio) {
2558                 bio->bi_size = 0;
2559                 bio->bi_bdev = bdev;
2560                 bio->bi_sector = first_sector;
2561                 btrfs_bio = btrfs_io_bio(bio);
2562                 btrfs_bio->csum = NULL;
2563                 btrfs_bio->csum_allocated = NULL;
2564                 btrfs_bio->end_io = NULL;
2565         }
2566         return bio;
2567 }
2568
2569 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2570 {
2571         return bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2572 }
2573
2574
2575 /* this also allocates from the btrfs_bioset */
2576 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2577 {
2578         struct btrfs_io_bio *btrfs_bio;
2579         struct bio *bio;
2580
2581         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2582         if (bio) {
2583                 btrfs_bio = btrfs_io_bio(bio);
2584                 btrfs_bio->csum = NULL;
2585                 btrfs_bio->csum_allocated = NULL;
2586                 btrfs_bio->end_io = NULL;
2587         }
2588         return bio;
2589 }
2590
2591
2592 static int __must_check submit_one_bio(int rw, struct bio *bio,
2593                                        int mirror_num, unsigned long bio_flags)
2594 {
2595         int ret = 0;
2596         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2597         struct page *page = bvec->bv_page;
2598         struct extent_io_tree *tree = bio->bi_private;
2599         u64 start;
2600
2601         start = page_offset(page) + bvec->bv_offset;
2602
2603         bio->bi_private = NULL;
2604
2605         bio_get(bio);
2606
2607         if (tree->ops && tree->ops->submit_bio_hook)
2608                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2609                                            mirror_num, bio_flags, start);
2610         else
2611                 btrfsic_submit_bio(rw, bio);
2612
2613         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2614                 ret = -EOPNOTSUPP;
2615         bio_put(bio);
2616         return ret;
2617 }
2618
2619 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2620                      unsigned long offset, size_t size, struct bio *bio,
2621                      unsigned long bio_flags)
2622 {
2623         int ret = 0;
2624         if (tree->ops && tree->ops->merge_bio_hook)
2625                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2626                                                 bio_flags);
2627         BUG_ON(ret < 0);
2628         return ret;
2629
2630 }
2631
2632 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2633                               struct page *page, sector_t sector,
2634                               size_t size, unsigned long offset,
2635                               struct block_device *bdev,
2636                               struct bio **bio_ret,
2637                               unsigned long max_pages,
2638                               bio_end_io_t end_io_func,
2639                               int mirror_num,
2640                               unsigned long prev_bio_flags,
2641                               unsigned long bio_flags)
2642 {
2643         int ret = 0;
2644         struct bio *bio;
2645         int nr;
2646         int contig = 0;
2647         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2648         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2649         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2650
2651         if (bio_ret && *bio_ret) {
2652                 bio = *bio_ret;
2653                 if (old_compressed)
2654                         contig = bio->bi_sector == sector;
2655                 else
2656                         contig = bio_end_sector(bio) == sector;
2657
2658                 if (prev_bio_flags != bio_flags || !contig ||
2659                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2660                     bio_add_page(bio, page, page_size, offset) < page_size) {
2661                         ret = submit_one_bio(rw, bio, mirror_num,
2662                                              prev_bio_flags);
2663                         if (ret < 0)
2664                                 return ret;
2665                         bio = NULL;
2666                 } else {
2667                         return 0;
2668                 }
2669         }
2670         if (this_compressed)
2671                 nr = BIO_MAX_PAGES;
2672         else
2673                 nr = bio_get_nr_vecs(bdev);
2674
2675         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2676         if (!bio)
2677                 return -ENOMEM;
2678
2679         bio_add_page(bio, page, page_size, offset);
2680         bio->bi_end_io = end_io_func;
2681         bio->bi_private = tree;
2682
2683         if (bio_ret)
2684                 *bio_ret = bio;
2685         else
2686                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2687
2688         return ret;
2689 }
2690
2691 static void attach_extent_buffer_page(struct extent_buffer *eb,
2692                                       struct page *page)
2693 {
2694         if (!PagePrivate(page)) {
2695                 SetPagePrivate(page);
2696                 page_cache_get(page);
2697                 set_page_private(page, (unsigned long)eb);
2698         } else {
2699                 WARN_ON(page->private != (unsigned long)eb);
2700         }
2701 }
2702
2703 void set_page_extent_mapped(struct page *page)
2704 {
2705         if (!PagePrivate(page)) {
2706                 SetPagePrivate(page);
2707                 page_cache_get(page);
2708                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2709         }
2710 }
2711
2712 static struct extent_map *
2713 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2714                  u64 start, u64 len, get_extent_t *get_extent,
2715                  struct extent_map **em_cached)
2716 {
2717         struct extent_map *em;
2718
2719         if (em_cached && *em_cached) {
2720                 em = *em_cached;
2721                 if (em->in_tree && start >= em->start &&
2722                     start < extent_map_end(em)) {
2723                         atomic_inc(&em->refs);
2724                         return em;
2725                 }
2726
2727                 free_extent_map(em);
2728                 *em_cached = NULL;
2729         }
2730
2731         em = get_extent(inode, page, pg_offset, start, len, 0);
2732         if (em_cached && !IS_ERR_OR_NULL(em)) {
2733                 BUG_ON(*em_cached);
2734                 atomic_inc(&em->refs);
2735                 *em_cached = em;
2736         }
2737         return em;
2738 }
2739 /*
2740  * basic readpage implementation.  Locked extent state structs are inserted
2741  * into the tree that are removed when the IO is done (by the end_io
2742  * handlers)
2743  * XXX JDM: This needs looking at to ensure proper page locking
2744  */
2745 static int __do_readpage(struct extent_io_tree *tree,
2746                          struct page *page,
2747                          get_extent_t *get_extent,
2748                          struct extent_map **em_cached,
2749                          struct bio **bio, int mirror_num,
2750                          unsigned long *bio_flags, int rw)
2751 {
2752         struct inode *inode = page->mapping->host;
2753         u64 start = page_offset(page);
2754         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2755         u64 end;
2756         u64 cur = start;
2757         u64 extent_offset;
2758         u64 last_byte = i_size_read(inode);
2759         u64 block_start;
2760         u64 cur_end;
2761         sector_t sector;
2762         struct extent_map *em;
2763         struct block_device *bdev;
2764         int ret;
2765         int nr = 0;
2766         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2767         size_t pg_offset = 0;
2768         size_t iosize;
2769         size_t disk_io_size;
2770         size_t blocksize = inode->i_sb->s_blocksize;
2771         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2772
2773         set_page_extent_mapped(page);
2774
2775         end = page_end;
2776         if (!PageUptodate(page)) {
2777                 if (cleancache_get_page(page) == 0) {
2778                         BUG_ON(blocksize != PAGE_SIZE);
2779                         unlock_extent(tree, start, end);
2780                         goto out;
2781                 }
2782         }
2783
2784         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2785                 char *userpage;
2786                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2787
2788                 if (zero_offset) {
2789                         iosize = PAGE_CACHE_SIZE - zero_offset;
2790                         userpage = kmap_atomic(page);
2791                         memset(userpage + zero_offset, 0, iosize);
2792                         flush_dcache_page(page);
2793                         kunmap_atomic(userpage);
2794                 }
2795         }
2796         while (cur <= end) {
2797                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2798
2799                 if (cur >= last_byte) {
2800                         char *userpage;
2801                         struct extent_state *cached = NULL;
2802
2803                         iosize = PAGE_CACHE_SIZE - pg_offset;
2804                         userpage = kmap_atomic(page);
2805                         memset(userpage + pg_offset, 0, iosize);
2806                         flush_dcache_page(page);
2807                         kunmap_atomic(userpage);
2808                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2809                                             &cached, GFP_NOFS);
2810                         if (!parent_locked)
2811                                 unlock_extent_cached(tree, cur,
2812                                                      cur + iosize - 1,
2813                                                      &cached, GFP_NOFS);
2814                         break;
2815                 }
2816                 em = __get_extent_map(inode, page, pg_offset, cur,
2817                                       end - cur + 1, get_extent, em_cached);
2818                 if (IS_ERR_OR_NULL(em)) {
2819                         SetPageError(page);
2820                         if (!parent_locked)
2821                                 unlock_extent(tree, cur, end);
2822                         break;
2823                 }
2824                 extent_offset = cur - em->start;
2825                 BUG_ON(extent_map_end(em) <= cur);
2826                 BUG_ON(end < cur);
2827
2828                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2829                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2830                         extent_set_compress_type(&this_bio_flag,
2831                                                  em->compress_type);
2832                 }
2833
2834                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2835                 cur_end = min(extent_map_end(em) - 1, end);
2836                 iosize = ALIGN(iosize, blocksize);
2837                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2838                         disk_io_size = em->block_len;
2839                         sector = em->block_start >> 9;
2840                 } else {
2841                         sector = (em->block_start + extent_offset) >> 9;
2842                         disk_io_size = iosize;
2843                 }
2844                 bdev = em->bdev;
2845                 block_start = em->block_start;
2846                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2847                         block_start = EXTENT_MAP_HOLE;
2848                 free_extent_map(em);
2849                 em = NULL;
2850
2851                 /* we've found a hole, just zero and go on */
2852                 if (block_start == EXTENT_MAP_HOLE) {
2853                         char *userpage;
2854                         struct extent_state *cached = NULL;
2855
2856                         userpage = kmap_atomic(page);
2857                         memset(userpage + pg_offset, 0, iosize);
2858                         flush_dcache_page(page);
2859                         kunmap_atomic(userpage);
2860
2861                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2862                                             &cached, GFP_NOFS);
2863                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2864                                              &cached, GFP_NOFS);
2865                         cur = cur + iosize;
2866                         pg_offset += iosize;
2867                         continue;
2868                 }
2869                 /* the get_extent function already copied into the page */
2870                 if (test_range_bit(tree, cur, cur_end,
2871                                    EXTENT_UPTODATE, 1, NULL)) {
2872                         check_page_uptodate(tree, page);
2873                         if (!parent_locked)
2874                                 unlock_extent(tree, cur, cur + iosize - 1);
2875                         cur = cur + iosize;
2876                         pg_offset += iosize;
2877                         continue;
2878                 }
2879                 /* we have an inline extent but it didn't get marked up
2880                  * to date.  Error out
2881                  */
2882                 if (block_start == EXTENT_MAP_INLINE) {
2883                         SetPageError(page);
2884                         if (!parent_locked)
2885                                 unlock_extent(tree, cur, cur + iosize - 1);
2886                         cur = cur + iosize;
2887                         pg_offset += iosize;
2888                         continue;
2889                 }
2890
2891                 pnr -= page->index;
2892                 ret = submit_extent_page(rw, tree, page,
2893                                          sector, disk_io_size, pg_offset,
2894                                          bdev, bio, pnr,
2895                                          end_bio_extent_readpage, mirror_num,
2896                                          *bio_flags,
2897                                          this_bio_flag);
2898                 if (!ret) {
2899                         nr++;
2900                         *bio_flags = this_bio_flag;
2901                 } else {
2902                         SetPageError(page);
2903                         if (!parent_locked)
2904                                 unlock_extent(tree, cur, cur + iosize - 1);
2905                 }
2906                 cur = cur + iosize;
2907                 pg_offset += iosize;
2908         }
2909 out:
2910         if (!nr) {
2911                 if (!PageError(page))
2912                         SetPageUptodate(page);
2913                 unlock_page(page);
2914         }
2915         return 0;
2916 }
2917
2918 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
2919                                              struct page *pages[], int nr_pages,
2920                                              u64 start, u64 end,
2921                                              get_extent_t *get_extent,
2922                                              struct extent_map **em_cached,
2923                                              struct bio **bio, int mirror_num,
2924                                              unsigned long *bio_flags, int rw)
2925 {
2926         struct inode *inode;
2927         struct btrfs_ordered_extent *ordered;
2928         int index;
2929
2930         inode = pages[0]->mapping->host;
2931         while (1) {
2932                 lock_extent(tree, start, end);
2933                 ordered = btrfs_lookup_ordered_range(inode, start,
2934                                                      end - start + 1);
2935                 if (!ordered)
2936                         break;
2937                 unlock_extent(tree, start, end);
2938                 btrfs_start_ordered_extent(inode, ordered, 1);
2939                 btrfs_put_ordered_extent(ordered);
2940         }
2941
2942         for (index = 0; index < nr_pages; index++) {
2943                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
2944                               mirror_num, bio_flags, rw);
2945                 page_cache_release(pages[index]);
2946         }
2947 }
2948
2949 static void __extent_readpages(struct extent_io_tree *tree,
2950                                struct page *pages[],
2951                                int nr_pages, get_extent_t *get_extent,
2952                                struct extent_map **em_cached,
2953                                struct bio **bio, int mirror_num,
2954                                unsigned long *bio_flags, int rw)
2955 {
2956         u64 start = 0;
2957         u64 end = 0;
2958         u64 page_start;
2959         int index;
2960         int first_index = 0;
2961
2962         for (index = 0; index < nr_pages; index++) {
2963                 page_start = page_offset(pages[index]);
2964                 if (!end) {
2965                         start = page_start;
2966                         end = start + PAGE_CACHE_SIZE - 1;
2967                         first_index = index;
2968                 } else if (end + 1 == page_start) {
2969                         end += PAGE_CACHE_SIZE;
2970                 } else {
2971                         __do_contiguous_readpages(tree, &pages[first_index],
2972                                                   index - first_index, start,
2973                                                   end, get_extent, em_cached,
2974                                                   bio, mirror_num, bio_flags,
2975                                                   rw);
2976                         start = page_start;
2977                         end = start + PAGE_CACHE_SIZE - 1;
2978                         first_index = index;
2979                 }
2980         }
2981
2982         if (end)
2983                 __do_contiguous_readpages(tree, &pages[first_index],
2984                                           index - first_index, start,
2985                                           end, get_extent, em_cached, bio,
2986                                           mirror_num, bio_flags, rw);
2987 }
2988
2989 static int __extent_read_full_page(struct extent_io_tree *tree,
2990                                    struct page *page,
2991                                    get_extent_t *get_extent,
2992                                    struct bio **bio, int mirror_num,
2993                                    unsigned long *bio_flags, int rw)
2994 {
2995         struct inode *inode = page->mapping->host;
2996         struct btrfs_ordered_extent *ordered;
2997         u64 start = page_offset(page);
2998         u64 end = start + PAGE_CACHE_SIZE - 1;
2999         int ret;
3000
3001         while (1) {
3002                 lock_extent(tree, start, end);
3003                 ordered = btrfs_lookup_ordered_extent(inode, start);
3004                 if (!ordered)
3005                         break;
3006                 unlock_extent(tree, start, end);
3007                 btrfs_start_ordered_extent(inode, ordered, 1);
3008                 btrfs_put_ordered_extent(ordered);
3009         }
3010
3011         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3012                             bio_flags, rw);
3013         return ret;
3014 }
3015
3016 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3017                             get_extent_t *get_extent, int mirror_num)
3018 {
3019         struct bio *bio = NULL;
3020         unsigned long bio_flags = 0;
3021         int ret;
3022
3023         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3024                                       &bio_flags, READ);
3025         if (bio)
3026                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3027         return ret;
3028 }
3029
3030 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3031                                  get_extent_t *get_extent, int mirror_num)
3032 {
3033         struct bio *bio = NULL;
3034         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3035         int ret;
3036
3037         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3038                                       &bio_flags, READ);
3039         if (bio)
3040                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3041         return ret;
3042 }
3043
3044 static noinline void update_nr_written(struct page *page,
3045                                       struct writeback_control *wbc,
3046                                       unsigned long nr_written)
3047 {
3048         wbc->nr_to_write -= nr_written;
3049         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3050             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3051                 page->mapping->writeback_index = page->index + nr_written;
3052 }
3053
3054 /*
3055  * the writepage semantics are similar to regular writepage.  extent
3056  * records are inserted to lock ranges in the tree, and as dirty areas
3057  * are found, they are marked writeback.  Then the lock bits are removed
3058  * and the end_io handler clears the writeback ranges
3059  */
3060 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3061                               void *data)
3062 {
3063         struct inode *inode = page->mapping->host;
3064         struct extent_page_data *epd = data;
3065         struct extent_io_tree *tree = epd->tree;
3066         u64 start = page_offset(page);
3067         u64 delalloc_start;
3068         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3069         u64 end;
3070         u64 cur = start;
3071         u64 extent_offset;
3072         u64 last_byte = i_size_read(inode);
3073         u64 block_start;
3074         u64 iosize;
3075         sector_t sector;
3076         struct extent_state *cached_state = NULL;
3077         struct extent_map *em;
3078         struct block_device *bdev;
3079         int ret;
3080         int nr = 0;
3081         size_t pg_offset = 0;
3082         size_t blocksize;
3083         loff_t i_size = i_size_read(inode);
3084         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3085         u64 nr_delalloc;
3086         u64 delalloc_end;
3087         int page_started;
3088         int compressed;
3089         int write_flags;
3090         unsigned long nr_written = 0;
3091         bool fill_delalloc = true;
3092
3093         if (wbc->sync_mode == WB_SYNC_ALL)
3094                 write_flags = WRITE_SYNC;
3095         else
3096                 write_flags = WRITE;
3097
3098         trace___extent_writepage(page, inode, wbc);
3099
3100         WARN_ON(!PageLocked(page));
3101
3102         ClearPageError(page);
3103
3104         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3105         if (page->index > end_index ||
3106            (page->index == end_index && !pg_offset)) {
3107                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3108                 unlock_page(page);
3109                 return 0;
3110         }
3111
3112         if (page->index == end_index) {
3113                 char *userpage;
3114
3115                 userpage = kmap_atomic(page);
3116                 memset(userpage + pg_offset, 0,
3117                        PAGE_CACHE_SIZE - pg_offset);
3118                 kunmap_atomic(userpage);
3119                 flush_dcache_page(page);
3120         }
3121         pg_offset = 0;
3122
3123         set_page_extent_mapped(page);
3124
3125         if (!tree->ops || !tree->ops->fill_delalloc)
3126                 fill_delalloc = false;
3127
3128         delalloc_start = start;
3129         delalloc_end = 0;
3130         page_started = 0;
3131         if (!epd->extent_locked && fill_delalloc) {
3132                 u64 delalloc_to_write = 0;
3133                 /*
3134                  * make sure the wbc mapping index is at least updated
3135                  * to this page.
3136                  */
3137                 update_nr_written(page, wbc, 0);
3138
3139                 while (delalloc_end < page_end) {
3140                         nr_delalloc = find_lock_delalloc_range(inode, tree,
3141                                                        page,
3142                                                        &delalloc_start,
3143                                                        &delalloc_end,
3144                                                        128 * 1024 * 1024);
3145                         if (nr_delalloc == 0) {
3146                                 delalloc_start = delalloc_end + 1;
3147                                 continue;
3148                         }
3149                         ret = tree->ops->fill_delalloc(inode, page,
3150                                                        delalloc_start,
3151                                                        delalloc_end,
3152                                                        &page_started,
3153                                                        &nr_written);
3154                         /* File system has been set read-only */
3155                         if (ret) {
3156                                 SetPageError(page);
3157                                 goto done;
3158                         }
3159                         /*
3160                          * delalloc_end is already one less than the total
3161                          * length, so we don't subtract one from
3162                          * PAGE_CACHE_SIZE
3163                          */
3164                         delalloc_to_write += (delalloc_end - delalloc_start +
3165                                               PAGE_CACHE_SIZE) >>
3166                                               PAGE_CACHE_SHIFT;
3167                         delalloc_start = delalloc_end + 1;
3168                 }
3169                 if (wbc->nr_to_write < delalloc_to_write) {
3170                         int thresh = 8192;
3171
3172                         if (delalloc_to_write < thresh * 2)
3173                                 thresh = delalloc_to_write;
3174                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
3175                                                  thresh);
3176                 }
3177
3178                 /* did the fill delalloc function already unlock and start
3179                  * the IO?
3180                  */
3181                 if (page_started) {
3182                         ret = 0;
3183                         /*
3184                          * we've unlocked the page, so we can't update
3185                          * the mapping's writeback index, just update
3186                          * nr_to_write.
3187                          */
3188                         wbc->nr_to_write -= nr_written;
3189                         goto done_unlocked;
3190                 }
3191         }
3192         if (tree->ops && tree->ops->writepage_start_hook) {
3193                 ret = tree->ops->writepage_start_hook(page, start,
3194                                                       page_end);
3195                 if (ret) {
3196                         /* Fixup worker will requeue */
3197                         if (ret == -EBUSY)
3198                                 wbc->pages_skipped++;
3199                         else
3200                                 redirty_page_for_writepage(wbc, page);
3201                         update_nr_written(page, wbc, nr_written);
3202                         unlock_page(page);
3203                         ret = 0;
3204                         goto done_unlocked;
3205                 }
3206         }
3207
3208         /*
3209          * we don't want to touch the inode after unlocking the page,
3210          * so we update the mapping writeback index now
3211          */
3212         update_nr_written(page, wbc, nr_written + 1);
3213
3214         end = page_end;
3215         if (last_byte <= start) {
3216                 if (tree->ops && tree->ops->writepage_end_io_hook)
3217                         tree->ops->writepage_end_io_hook(page, start,
3218                                                          page_end, NULL, 1);
3219                 goto done;
3220         }
3221
3222         blocksize = inode->i_sb->s_blocksize;
3223
3224         while (cur <= end) {
3225                 if (cur >= last_byte) {
3226                         if (tree->ops && tree->ops->writepage_end_io_hook)
3227                                 tree->ops->writepage_end_io_hook(page, cur,
3228                                                          page_end, NULL, 1);
3229                         break;
3230                 }
3231                 em = epd->get_extent(inode, page, pg_offset, cur,
3232                                      end - cur + 1, 1);
3233                 if (IS_ERR_OR_NULL(em)) {
3234                         SetPageError(page);
3235                         break;
3236                 }
3237
3238                 extent_offset = cur - em->start;
3239                 BUG_ON(extent_map_end(em) <= cur);
3240                 BUG_ON(end < cur);
3241                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3242                 iosize = ALIGN(iosize, blocksize);
3243                 sector = (em->block_start + extent_offset) >> 9;
3244                 bdev = em->bdev;
3245                 block_start = em->block_start;
3246                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3247                 free_extent_map(em);
3248                 em = NULL;
3249
3250                 /*
3251                  * compressed and inline extents are written through other
3252                  * paths in the FS
3253                  */
3254                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3255                     block_start == EXTENT_MAP_INLINE) {
3256                         /*
3257                          * end_io notification does not happen here for
3258                          * compressed extents
3259                          */
3260                         if (!compressed && tree->ops &&
3261                             tree->ops->writepage_end_io_hook)
3262                                 tree->ops->writepage_end_io_hook(page, cur,
3263                                                          cur + iosize - 1,
3264                                                          NULL, 1);
3265                         else if (compressed) {
3266                                 /* we don't want to end_page_writeback on
3267                                  * a compressed extent.  this happens
3268                                  * elsewhere
3269                                  */
3270                                 nr++;
3271                         }
3272
3273                         cur += iosize;
3274                         pg_offset += iosize;
3275                         continue;
3276                 }
3277                 /* leave this out until we have a page_mkwrite call */
3278                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
3279                                    EXTENT_DIRTY, 0, NULL)) {
3280                         cur = cur + iosize;
3281                         pg_offset += iosize;
3282                         continue;
3283                 }
3284
3285                 if (tree->ops && tree->ops->writepage_io_hook) {
3286                         ret = tree->ops->writepage_io_hook(page, cur,
3287                                                 cur + iosize - 1);
3288                 } else {
3289                         ret = 0;
3290                 }
3291                 if (ret) {
3292                         SetPageError(page);
3293                 } else {
3294                         unsigned long max_nr = end_index + 1;
3295
3296                         set_range_writeback(tree, cur, cur + iosize - 1);
3297                         if (!PageWriteback(page)) {
3298                                 printk(KERN_ERR "btrfs warning page %lu not "
3299                                        "writeback, cur %llu end %llu\n",
3300                                        page->index, cur, end);
3301                         }
3302
3303                         ret = submit_extent_page(write_flags, tree, page,
3304                                                  sector, iosize, pg_offset,
3305                                                  bdev, &epd->bio, max_nr,
3306                                                  end_bio_extent_writepage,
3307                                                  0, 0, 0);
3308                         if (ret)
3309                                 SetPageError(page);
3310                 }
3311                 cur = cur + iosize;
3312                 pg_offset += iosize;
3313                 nr++;
3314         }
3315 done:
3316         if (nr == 0) {
3317                 /* make sure the mapping tag for page dirty gets cleared */
3318                 set_page_writeback(page);
3319                 end_page_writeback(page);
3320         }
3321         unlock_page(page);
3322
3323 done_unlocked:
3324
3325         /* drop our reference on any cached states */
3326         free_extent_state(cached_state);
3327         return 0;
3328 }
3329
3330 static int eb_wait(void *word)
3331 {
3332         io_schedule();
3333         return 0;
3334 }
3335
3336 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3337 {
3338         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3339                     TASK_UNINTERRUPTIBLE);
3340 }
3341
3342 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3343                                      struct btrfs_fs_info *fs_info,
3344                                      struct extent_page_data *epd)
3345 {
3346         unsigned long i, num_pages;
3347         int flush = 0;
3348         int ret = 0;
3349
3350         if (!btrfs_try_tree_write_lock(eb)) {
3351                 flush = 1;
3352                 flush_write_bio(epd);
3353                 btrfs_tree_lock(eb);
3354         }
3355
3356         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3357                 btrfs_tree_unlock(eb);
3358                 if (!epd->sync_io)
3359                         return 0;
3360                 if (!flush) {
3361                         flush_write_bio(epd);
3362                         flush = 1;
3363                 }
3364                 while (1) {
3365                         wait_on_extent_buffer_writeback(eb);
3366                         btrfs_tree_lock(eb);
3367                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3368                                 break;
3369                         btrfs_tree_unlock(eb);
3370                 }
3371         }
3372
3373         /*
3374          * We need to do this to prevent races in people who check if the eb is
3375          * under IO since we can end up having no IO bits set for a short period
3376          * of time.
3377          */
3378         spin_lock(&eb->refs_lock);
3379         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3380                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3381                 spin_unlock(&eb->refs_lock);
3382                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3383                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3384                                      -eb->len,
3385                                      fs_info->dirty_metadata_batch);
3386                 ret = 1;
3387         } else {
3388                 spin_unlock(&eb->refs_lock);
3389         }
3390
3391         btrfs_tree_unlock(eb);
3392
3393         if (!ret)
3394                 return ret;
3395
3396         num_pages = num_extent_pages(eb->start, eb->len);
3397         for (i = 0; i < num_pages; i++) {
3398                 struct page *p = extent_buffer_page(eb, i);
3399
3400                 if (!trylock_page(p)) {
3401                         if (!flush) {
3402                                 flush_write_bio(epd);
3403                                 flush = 1;
3404                         }
3405                         lock_page(p);
3406                 }
3407         }
3408
3409         return ret;
3410 }
3411
3412 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3413 {
3414         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3415         smp_mb__after_clear_bit();
3416         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3417 }
3418
3419 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3420 {
3421         int uptodate = err == 0;
3422         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
3423         struct extent_buffer *eb;
3424         int done;
3425
3426         do {
3427                 struct page *page = bvec->bv_page;
3428
3429                 bvec--;
3430                 eb = (struct extent_buffer *)page->private;
3431                 BUG_ON(!eb);
3432                 done = atomic_dec_and_test(&eb->io_pages);
3433
3434                 if (!uptodate || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3435                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3436                         ClearPageUptodate(page);
3437                         SetPageError(page);
3438                 }
3439
3440                 end_page_writeback(page);
3441
3442                 if (!done)
3443                         continue;
3444
3445                 end_extent_buffer_writeback(eb);
3446         } while (bvec >= bio->bi_io_vec);
3447
3448         bio_put(bio);
3449
3450 }
3451
3452 static int write_one_eb(struct extent_buffer *eb,
3453                         struct btrfs_fs_info *fs_info,
3454                         struct writeback_control *wbc,
3455                         struct extent_page_data *epd)
3456 {
3457         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3458         u64 offset = eb->start;
3459         unsigned long i, num_pages;
3460         unsigned long bio_flags = 0;
3461         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3462         int ret = 0;
3463
3464         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3465         num_pages = num_extent_pages(eb->start, eb->len);
3466         atomic_set(&eb->io_pages, num_pages);
3467         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3468                 bio_flags = EXTENT_BIO_TREE_LOG;
3469
3470         for (i = 0; i < num_pages; i++) {
3471                 struct page *p = extent_buffer_page(eb, i);
3472
3473                 clear_page_dirty_for_io(p);
3474                 set_page_writeback(p);
3475                 ret = submit_extent_page(rw, eb->tree, p, offset >> 9,
3476                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3477                                          -1, end_bio_extent_buffer_writepage,
3478                                          0, epd->bio_flags, bio_flags);
3479                 epd->bio_flags = bio_flags;
3480                 if (ret) {
3481                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3482                         SetPageError(p);
3483                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3484                                 end_extent_buffer_writeback(eb);
3485                         ret = -EIO;
3486                         break;
3487                 }
3488                 offset += PAGE_CACHE_SIZE;
3489                 update_nr_written(p, wbc, 1);
3490                 unlock_page(p);
3491         }
3492
3493         if (unlikely(ret)) {
3494                 for (; i < num_pages; i++) {
3495                         struct page *p = extent_buffer_page(eb, i);
3496                         unlock_page(p);
3497                 }
3498         }
3499
3500         return ret;
3501 }
3502
3503 int btree_write_cache_pages(struct address_space *mapping,
3504                                    struct writeback_control *wbc)
3505 {
3506         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3507         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3508         struct extent_buffer *eb, *prev_eb = NULL;
3509         struct extent_page_data epd = {
3510                 .bio = NULL,
3511                 .tree = tree,
3512                 .extent_locked = 0,
3513                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3514                 .bio_flags = 0,
3515         };
3516         int ret = 0;
3517         int done = 0;
3518         int nr_to_write_done = 0;
3519         struct pagevec pvec;
3520         int nr_pages;
3521         pgoff_t index;
3522         pgoff_t end;            /* Inclusive */
3523         int scanned = 0;
3524         int tag;
3525
3526         pagevec_init(&pvec, 0);
3527         if (wbc->range_cyclic) {
3528                 index = mapping->writeback_index; /* Start from prev offset */
3529                 end = -1;
3530         } else {
3531                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3532                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3533                 scanned = 1;
3534         }
3535         if (wbc->sync_mode == WB_SYNC_ALL)
3536                 tag = PAGECACHE_TAG_TOWRITE;
3537         else
3538                 tag = PAGECACHE_TAG_DIRTY;
3539 retry:
3540         if (wbc->sync_mode == WB_SYNC_ALL)
3541                 tag_pages_for_writeback(mapping, index, end);
3542         while (!done && !nr_to_write_done && (index <= end) &&
3543                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3544                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3545                 unsigned i;
3546
3547                 scanned = 1;
3548                 for (i = 0; i < nr_pages; i++) {
3549                         struct page *page = pvec.pages[i];
3550
3551                         if (!PagePrivate(page))
3552                                 continue;
3553
3554                         if (!wbc->range_cyclic && page->index > end) {
3555                                 done = 1;
3556                                 break;
3557                         }
3558
3559                         spin_lock(&mapping->private_lock);
3560                         if (!PagePrivate(page)) {
3561                                 spin_unlock(&mapping->private_lock);
3562                                 continue;
3563                         }
3564
3565                         eb = (struct extent_buffer *)page->private;
3566
3567                         /*
3568                          * Shouldn't happen and normally this would be a BUG_ON
3569                          * but no sense in crashing the users box for something
3570                          * we can survive anyway.
3571                          */
3572                         if (!eb) {
3573                                 spin_unlock(&mapping->private_lock);
3574                                 WARN_ON(1);
3575                                 continue;
3576                         }
3577
3578                         if (eb == prev_eb) {
3579                                 spin_unlock(&mapping->private_lock);
3580                                 continue;
3581                         }
3582
3583                         ret = atomic_inc_not_zero(&eb->refs);
3584                         spin_unlock(&mapping->private_lock);
3585                         if (!ret)
3586                                 continue;
3587
3588                         prev_eb = eb;
3589                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3590                         if (!ret) {
3591                                 free_extent_buffer(eb);
3592                                 continue;
3593                         }
3594
3595                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3596                         if (ret) {
3597                                 done = 1;
3598                                 free_extent_buffer(eb);
3599                                 break;
3600                         }
3601                         free_extent_buffer(eb);
3602
3603                         /*
3604                          * the filesystem may choose to bump up nr_to_write.
3605                          * We have to make sure to honor the new nr_to_write
3606                          * at any time
3607                          */
3608                         nr_to_write_done = wbc->nr_to_write <= 0;
3609                 }
3610                 pagevec_release(&pvec);
3611                 cond_resched();
3612         }
3613         if (!scanned && !done) {
3614                 /*
3615                  * We hit the last page and there is more work to be done: wrap
3616                  * back to the start of the file
3617                  */
3618                 scanned = 1;
3619                 index = 0;
3620                 goto retry;
3621         }
3622         flush_write_bio(&epd);
3623         return ret;
3624 }
3625
3626 /**
3627  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3628  * @mapping: address space structure to write
3629  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3630  * @writepage: function called for each page
3631  * @data: data passed to writepage function
3632  *
3633  * If a page is already under I/O, write_cache_pages() skips it, even
3634  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3635  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3636  * and msync() need to guarantee that all the data which was dirty at the time
3637  * the call was made get new I/O started against them.  If wbc->sync_mode is
3638  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3639  * existing IO to complete.
3640  */
3641 static int extent_write_cache_pages(struct extent_io_tree *tree,
3642                              struct address_space *mapping,
3643                              struct writeback_control *wbc,
3644                              writepage_t writepage, void *data,
3645                              void (*flush_fn)(void *))
3646 {
3647         struct inode *inode = mapping->host;
3648         int ret = 0;
3649         int done = 0;
3650         int nr_to_write_done = 0;
3651         struct pagevec pvec;
3652         int nr_pages;
3653         pgoff_t index;
3654         pgoff_t end;            /* Inclusive */
3655         int scanned = 0;
3656         int tag;
3657
3658         /*
3659          * We have to hold onto the inode so that ordered extents can do their
3660          * work when the IO finishes.  The alternative to this is failing to add
3661          * an ordered extent if the igrab() fails there and that is a huge pain
3662          * to deal with, so instead just hold onto the inode throughout the
3663          * writepages operation.  If it fails here we are freeing up the inode
3664          * anyway and we'd rather not waste our time writing out stuff that is
3665          * going to be truncated anyway.
3666          */
3667         if (!igrab(inode))
3668                 return 0;
3669
3670         pagevec_init(&pvec, 0);
3671         if (wbc->range_cyclic) {
3672                 index = mapping->writeback_index; /* Start from prev offset */
3673                 end = -1;
3674         } else {
3675                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3676                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3677                 scanned = 1;
3678         }
3679         if (wbc->sync_mode == WB_SYNC_ALL)
3680                 tag = PAGECACHE_TAG_TOWRITE;
3681         else
3682                 tag = PAGECACHE_TAG_DIRTY;
3683 retry:
3684         if (wbc->sync_mode == WB_SYNC_ALL)
3685                 tag_pages_for_writeback(mapping, index, end);
3686         while (!done && !nr_to_write_done && (index <= end) &&
3687                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3688                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3689                 unsigned i;
3690
3691                 scanned = 1;
3692                 for (i = 0; i < nr_pages; i++) {
3693                         struct page *page = pvec.pages[i];
3694
3695                         /*
3696                          * At this point we hold neither mapping->tree_lock nor
3697                          * lock on the page itself: the page may be truncated or
3698                          * invalidated (changing page->mapping to NULL), or even
3699                          * swizzled back from swapper_space to tmpfs file
3700                          * mapping
3701                          */
3702                         if (!trylock_page(page)) {
3703                                 flush_fn(data);
3704                                 lock_page(page);
3705                         }
3706
3707                         if (unlikely(page->mapping != mapping)) {
3708                                 unlock_page(page);
3709                                 continue;
3710                         }
3711
3712                         if (!wbc->range_cyclic && page->index > end) {
3713                                 done = 1;
3714                                 unlock_page(page);
3715                                 continue;
3716                         }
3717
3718                         if (wbc->sync_mode != WB_SYNC_NONE) {
3719                                 if (PageWriteback(page))
3720                                         flush_fn(data);
3721                                 wait_on_page_writeback(page);
3722                         }
3723
3724                         if (PageWriteback(page) ||
3725                             !clear_page_dirty_for_io(page)) {
3726                                 unlock_page(page);
3727                                 continue;
3728                         }
3729
3730                         ret = (*writepage)(page, wbc, data);
3731
3732                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3733                                 unlock_page(page);
3734                                 ret = 0;
3735                         }
3736                         if (ret)
3737                                 done = 1;
3738
3739                         /*
3740                          * the filesystem may choose to bump up nr_to_write.
3741                          * We have to make sure to honor the new nr_to_write
3742                          * at any time
3743                          */
3744                         nr_to_write_done = wbc->nr_to_write <= 0;
3745                 }
3746                 pagevec_release(&pvec);
3747                 cond_resched();
3748         }
3749         if (!scanned && !done) {
3750                 /*
3751                  * We hit the last page and there is more work to be done: wrap
3752                  * back to the start of the file
3753                  */
3754                 scanned = 1;
3755                 index = 0;
3756                 goto retry;
3757         }
3758         btrfs_add_delayed_iput(inode);
3759         return ret;
3760 }
3761
3762 static void flush_epd_write_bio(struct extent_page_data *epd)
3763 {
3764         if (epd->bio) {
3765                 int rw = WRITE;
3766                 int ret;
3767
3768                 if (epd->sync_io)
3769                         rw = WRITE_SYNC;
3770
3771                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3772                 BUG_ON(ret < 0); /* -ENOMEM */
3773                 epd->bio = NULL;
3774         }
3775 }
3776
3777 static noinline void flush_write_bio(void *data)
3778 {
3779         struct extent_page_data *epd = data;
3780         flush_epd_write_bio(epd);
3781 }
3782
3783 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3784                           get_extent_t *get_extent,
3785                           struct writeback_control *wbc)
3786 {
3787         int ret;
3788         struct extent_page_data epd = {
3789                 .bio = NULL,
3790                 .tree = tree,
3791                 .get_extent = get_extent,
3792                 .extent_locked = 0,
3793                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3794                 .bio_flags = 0,
3795         };
3796
3797         ret = __extent_writepage(page, wbc, &epd);
3798
3799         flush_epd_write_bio(&epd);
3800         return ret;
3801 }
3802
3803 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3804                               u64 start, u64 end, get_extent_t *get_extent,
3805                               int mode)
3806 {
3807         int ret = 0;
3808         struct address_space *mapping = inode->i_mapping;
3809         struct page *page;
3810         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3811                 PAGE_CACHE_SHIFT;
3812
3813         struct extent_page_data epd = {
3814                 .bio = NULL,
3815                 .tree = tree,
3816                 .get_extent = get_extent,
3817                 .extent_locked = 1,
3818                 .sync_io = mode == WB_SYNC_ALL,
3819                 .bio_flags = 0,
3820         };
3821         struct writeback_control wbc_writepages = {
3822                 .sync_mode      = mode,
3823                 .nr_to_write    = nr_pages * 2,
3824                 .range_start    = start,
3825                 .range_end      = end + 1,
3826         };
3827
3828         while (start <= end) {
3829                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3830                 if (clear_page_dirty_for_io(page))
3831                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3832                 else {
3833                         if (tree->ops && tree->ops->writepage_end_io_hook)
3834                                 tree->ops->writepage_end_io_hook(page, start,
3835                                                  start + PAGE_CACHE_SIZE - 1,
3836                                                  NULL, 1);
3837                         unlock_page(page);
3838                 }
3839                 page_cache_release(page);
3840                 start += PAGE_CACHE_SIZE;
3841         }
3842
3843         flush_epd_write_bio(&epd);
3844         return ret;
3845 }
3846
3847 int extent_writepages(struct extent_io_tree *tree,
3848                       struct address_space *mapping,
3849                       get_extent_t *get_extent,
3850                       struct writeback_control *wbc)
3851 {
3852         int ret = 0;
3853         struct extent_page_data epd = {
3854                 .bio = NULL,
3855                 .tree = tree,
3856                 .get_extent = get_extent,
3857                 .extent_locked = 0,
3858                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3859                 .bio_flags = 0,
3860         };
3861
3862         ret = extent_write_cache_pages(tree, mapping, wbc,
3863                                        __extent_writepage, &epd,
3864                                        flush_write_bio);
3865         flush_epd_write_bio(&epd);
3866         return ret;
3867 }
3868
3869 int extent_readpages(struct extent_io_tree *tree,
3870                      struct address_space *mapping,
3871                      struct list_head *pages, unsigned nr_pages,
3872                      get_extent_t get_extent)
3873 {
3874         struct bio *bio = NULL;
3875         unsigned page_idx;
3876         unsigned long bio_flags = 0;
3877         struct page *pagepool[16];
3878         struct page *page;
3879         struct extent_map *em_cached = NULL;
3880         int nr = 0;
3881
3882         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3883                 page = list_entry(pages->prev, struct page, lru);
3884
3885                 prefetchw(&page->flags);
3886                 list_del(&page->lru);
3887                 if (add_to_page_cache_lru(page, mapping,
3888                                         page->index, GFP_NOFS)) {
3889                         page_cache_release(page);
3890                         continue;
3891                 }
3892
3893                 pagepool[nr++] = page;
3894                 if (nr < ARRAY_SIZE(pagepool))
3895                         continue;
3896                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3897                                    &bio, 0, &bio_flags, READ);
3898                 nr = 0;
3899         }
3900         if (nr)
3901                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3902                                    &bio, 0, &bio_flags, READ);
3903
3904         if (em_cached)
3905                 free_extent_map(em_cached);
3906
3907         BUG_ON(!list_empty(pages));
3908         if (bio)
3909                 return submit_one_bio(READ, bio, 0, bio_flags);
3910         return 0;
3911 }
3912
3913 /*
3914  * basic invalidatepage code, this waits on any locked or writeback
3915  * ranges corresponding to the page, and then deletes any extent state
3916  * records from the tree
3917  */
3918 int extent_invalidatepage(struct extent_io_tree *tree,
3919                           struct page *page, unsigned long offset)
3920 {
3921         struct extent_state *cached_state = NULL;
3922         u64 start = page_offset(page);
3923         u64 end = start + PAGE_CACHE_SIZE - 1;
3924         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3925
3926         start += ALIGN(offset, blocksize);
3927         if (start > end)
3928                 return 0;
3929
3930         lock_extent_bits(tree, start, end, 0, &cached_state);
3931         wait_on_page_writeback(page);
3932         clear_extent_bit(tree, start, end,
3933                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3934                          EXTENT_DO_ACCOUNTING,
3935                          1, 1, &cached_state, GFP_NOFS);
3936         return 0;
3937 }
3938
3939 /*
3940  * a helper for releasepage, this tests for areas of the page that
3941  * are locked or under IO and drops the related state bits if it is safe
3942  * to drop the page.
3943  */
3944 static int try_release_extent_state(struct extent_map_tree *map,
3945                                     struct extent_io_tree *tree,
3946                                     struct page *page, gfp_t mask)
3947 {
3948         u64 start = page_offset(page);
3949         u64 end = start + PAGE_CACHE_SIZE - 1;
3950         int ret = 1;
3951
3952         if (test_range_bit(tree, start, end,
3953                            EXTENT_IOBITS, 0, NULL))
3954                 ret = 0;
3955         else {
3956                 if ((mask & GFP_NOFS) == GFP_NOFS)
3957                         mask = GFP_NOFS;
3958                 /*
3959                  * at this point we can safely clear everything except the
3960                  * locked bit and the nodatasum bit
3961                  */
3962                 ret = clear_extent_bit(tree, start, end,
3963                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3964                                  0, 0, NULL, mask);
3965
3966                 /* if clear_extent_bit failed for enomem reasons,
3967                  * we can't allow the release to continue.
3968                  */
3969                 if (ret < 0)
3970                         ret = 0;
3971                 else
3972                         ret = 1;
3973         }
3974         return ret;
3975 }
3976
3977 /*
3978  * a helper for releasepage.  As long as there are no locked extents
3979  * in the range corresponding to the page, both state records and extent
3980  * map records are removed
3981  */
3982 int try_release_extent_mapping(struct extent_map_tree *map,
3983                                struct extent_io_tree *tree, struct page *page,
3984                                gfp_t mask)
3985 {
3986         struct extent_map *em;
3987         u64 start = page_offset(page);
3988         u64 end = start + PAGE_CACHE_SIZE - 1;
3989
3990         if ((mask & __GFP_WAIT) &&
3991             page->mapping->host->i_size > 16 * 1024 * 1024) {
3992                 u64 len;
3993                 while (start <= end) {
3994                         len = end - start + 1;
3995                         write_lock(&map->lock);
3996                         em = lookup_extent_mapping(map, start, len);
3997                         if (!em) {
3998                                 write_unlock(&map->lock);
3999                                 break;
4000                         }
4001                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4002                             em->start != start) {
4003                                 write_unlock(&map->lock);
4004                                 free_extent_map(em);
4005                                 break;
4006                         }
4007                         if (!test_range_bit(tree, em->start,
4008                                             extent_map_end(em) - 1,
4009                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4010                                             0, NULL)) {
4011                                 remove_extent_mapping(map, em);
4012                                 /* once for the rb tree */
4013                                 free_extent_map(em);
4014                         }
4015                         start = extent_map_end(em);
4016                         write_unlock(&map->lock);
4017
4018                         /* once for us */
4019                         free_extent_map(em);
4020                 }
4021         }
4022         return try_release_extent_state(map, tree, page, mask);
4023 }
4024
4025 /*
4026  * helper function for fiemap, which doesn't want to see any holes.
4027  * This maps until we find something past 'last'
4028  */
4029 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4030                                                 u64 offset,
4031                                                 u64 last,
4032                                                 get_extent_t *get_extent)
4033 {
4034         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4035         struct extent_map *em;
4036         u64 len;
4037
4038         if (offset >= last)
4039                 return NULL;
4040
4041         while(1) {
4042                 len = last - offset;
4043                 if (len == 0)
4044                         break;
4045                 len = ALIGN(len, sectorsize);
4046                 em = get_extent(inode, NULL, 0, offset, len, 0);
4047                 if (IS_ERR_OR_NULL(em))
4048                         return em;
4049
4050                 /* if this isn't a hole return it */
4051                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4052                     em->block_start != EXTENT_MAP_HOLE) {
4053                         return em;
4054                 }
4055
4056                 /* this is a hole, advance to the next extent */
4057                 offset = extent_map_end(em);
4058                 free_extent_map(em);
4059                 if (offset >= last)
4060                         break;
4061         }
4062         return NULL;
4063 }
4064
4065 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4066                 __u64 start, __u64 len, get_extent_t *get_extent)
4067 {
4068         int ret = 0;
4069         u64 off = start;
4070         u64 max = start + len;
4071         u32 flags = 0;
4072         u32 found_type;
4073         u64 last;
4074         u64 last_for_get_extent = 0;
4075         u64 disko = 0;
4076         u64 isize = i_size_read(inode);
4077         struct btrfs_key found_key;
4078         struct extent_map *em = NULL;
4079         struct extent_state *cached_state = NULL;
4080         struct btrfs_path *path;
4081         struct btrfs_file_extent_item *item;
4082         int end = 0;
4083         u64 em_start = 0;
4084         u64 em_len = 0;
4085         u64 em_end = 0;
4086         unsigned long emflags;
4087
4088         if (len == 0)
4089                 return -EINVAL;
4090
4091         path = btrfs_alloc_path();
4092         if (!path)
4093                 return -ENOMEM;
4094         path->leave_spinning = 1;
4095
4096         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
4097         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
4098
4099         /*
4100          * lookup the last file extent.  We're not using i_size here
4101          * because there might be preallocation past i_size
4102          */
4103         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
4104                                        path, btrfs_ino(inode), -1, 0);
4105         if (ret < 0) {
4106                 btrfs_free_path(path);
4107                 return ret;
4108         }
4109         WARN_ON(!ret);
4110         path->slots[0]--;
4111         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
4112                               struct btrfs_file_extent_item);
4113         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4114         found_type = btrfs_key_type(&found_key);
4115
4116         /* No extents, but there might be delalloc bits */
4117         if (found_key.objectid != btrfs_ino(inode) ||
4118             found_type != BTRFS_EXTENT_DATA_KEY) {
4119                 /* have to trust i_size as the end */
4120                 last = (u64)-1;
4121                 last_for_get_extent = isize;
4122         } else {
4123                 /*
4124                  * remember the start of the last extent.  There are a
4125                  * bunch of different factors that go into the length of the
4126                  * extent, so its much less complex to remember where it started
4127                  */
4128                 last = found_key.offset;
4129                 last_for_get_extent = last + 1;
4130         }
4131         btrfs_free_path(path);
4132
4133         /*
4134          * we might have some extents allocated but more delalloc past those
4135          * extents.  so, we trust isize unless the start of the last extent is
4136          * beyond isize
4137          */
4138         if (last < isize) {
4139                 last = (u64)-1;
4140                 last_for_get_extent = isize;
4141         }
4142
4143         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4144                          &cached_state);
4145
4146         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4147                                    get_extent);
4148         if (!em)
4149                 goto out;
4150         if (IS_ERR(em)) {
4151                 ret = PTR_ERR(em);
4152                 goto out;
4153         }
4154
4155         while (!end) {
4156                 u64 offset_in_extent = 0;
4157
4158                 /* break if the extent we found is outside the range */
4159                 if (em->start >= max || extent_map_end(em) < off)
4160                         break;
4161
4162                 /*
4163                  * get_extent may return an extent that starts before our
4164                  * requested range.  We have to make sure the ranges
4165                  * we return to fiemap always move forward and don't
4166                  * overlap, so adjust the offsets here
4167                  */
4168                 em_start = max(em->start, off);
4169
4170                 /*
4171                  * record the offset from the start of the extent
4172                  * for adjusting the disk offset below.  Only do this if the
4173                  * extent isn't compressed since our in ram offset may be past
4174                  * what we have actually allocated on disk.
4175                  */
4176                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4177                         offset_in_extent = em_start - em->start;
4178                 em_end = extent_map_end(em);
4179                 em_len = em_end - em_start;
4180                 emflags = em->flags;
4181                 disko = 0;
4182                 flags = 0;
4183
4184                 /*
4185                  * bump off for our next call to get_extent
4186                  */
4187                 off = extent_map_end(em);
4188                 if (off >= max)
4189                         end = 1;
4190
4191                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4192                         end = 1;
4193                         flags |= FIEMAP_EXTENT_LAST;
4194                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4195                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4196                                   FIEMAP_EXTENT_NOT_ALIGNED);
4197                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4198                         flags |= (FIEMAP_EXTENT_DELALLOC |
4199                                   FIEMAP_EXTENT_UNKNOWN);
4200                 } else {
4201                         disko = em->block_start + offset_in_extent;
4202                 }
4203                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4204                         flags |= FIEMAP_EXTENT_ENCODED;
4205
4206                 free_extent_map(em);
4207                 em = NULL;
4208                 if ((em_start >= last) || em_len == (u64)-1 ||
4209                    (last == (u64)-1 && isize <= em_end)) {
4210                         flags |= FIEMAP_EXTENT_LAST;
4211                         end = 1;
4212                 }
4213
4214                 /* now scan forward to see if this is really the last extent. */
4215                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4216                                            get_extent);
4217                 if (IS_ERR(em)) {
4218                         ret = PTR_ERR(em);
4219                         goto out;
4220                 }
4221                 if (!em) {
4222                         flags |= FIEMAP_EXTENT_LAST;
4223                         end = 1;
4224                 }
4225                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4226                                               em_len, flags);
4227                 if (ret)
4228                         goto out_free;
4229         }
4230 out_free:
4231         free_extent_map(em);
4232 out:
4233         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4234                              &cached_state, GFP_NOFS);
4235         return ret;
4236 }
4237
4238 static void __free_extent_buffer(struct extent_buffer *eb)
4239 {
4240         btrfs_leak_debug_del(&eb->leak_list);
4241         kmem_cache_free(extent_buffer_cache, eb);
4242 }
4243
4244 static int extent_buffer_under_io(struct extent_buffer *eb)
4245 {
4246         return (atomic_read(&eb->io_pages) ||
4247                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4248                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4249 }
4250
4251 /*
4252  * Helper for releasing extent buffer page.
4253  */
4254 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4255                                                 unsigned long start_idx)
4256 {
4257         unsigned long index;
4258         unsigned long num_pages;
4259         struct page *page;
4260         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4261
4262         BUG_ON(extent_buffer_under_io(eb));
4263
4264         num_pages = num_extent_pages(eb->start, eb->len);
4265         index = start_idx + num_pages;
4266         if (start_idx >= index)
4267                 return;
4268
4269         do {
4270                 index--;
4271                 page = extent_buffer_page(eb, index);
4272                 if (page && mapped) {
4273                         spin_lock(&page->mapping->private_lock);
4274                         /*
4275                          * We do this since we'll remove the pages after we've
4276                          * removed the eb from the radix tree, so we could race
4277                          * and have this page now attached to the new eb.  So
4278                          * only clear page_private if it's still connected to
4279                          * this eb.
4280                          */
4281                         if (PagePrivate(page) &&
4282                             page->private == (unsigned long)eb) {
4283                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4284                                 BUG_ON(PageDirty(page));
4285                                 BUG_ON(PageWriteback(page));
4286                                 /*
4287                                  * We need to make sure we haven't be attached
4288                                  * to a new eb.
4289                                  */
4290                                 ClearPagePrivate(page);
4291                                 set_page_private(page, 0);
4292                                 /* One for the page private */
4293                                 page_cache_release(page);
4294                         }
4295                         spin_unlock(&page->mapping->private_lock);
4296
4297                 }
4298                 if (page) {
4299                         /* One for when we alloced the page */
4300                         page_cache_release(page);
4301                 }
4302         } while (index != start_idx);
4303 }
4304
4305 /*
4306  * Helper for releasing the extent buffer.
4307  */
4308 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4309 {
4310         btrfs_release_extent_buffer_page(eb, 0);
4311         __free_extent_buffer(eb);
4312 }
4313
4314 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
4315                                                    u64 start,
4316                                                    unsigned long len,
4317                                                    gfp_t mask)
4318 {
4319         struct extent_buffer *eb = NULL;
4320
4321         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4322         if (eb == NULL)
4323                 return NULL;
4324         eb->start = start;
4325         eb->len = len;
4326         eb->tree = tree;
4327         eb->bflags = 0;
4328         rwlock_init(&eb->lock);
4329         atomic_set(&eb->write_locks, 0);
4330         atomic_set(&eb->read_locks, 0);
4331         atomic_set(&eb->blocking_readers, 0);
4332         atomic_set(&eb->blocking_writers, 0);
4333         atomic_set(&eb->spinning_readers, 0);
4334         atomic_set(&eb->spinning_writers, 0);
4335         eb->lock_nested = 0;
4336         init_waitqueue_head(&eb->write_lock_wq);
4337         init_waitqueue_head(&eb->read_lock_wq);
4338
4339         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4340
4341         spin_lock_init(&eb->refs_lock);
4342         atomic_set(&eb->refs, 1);
4343         atomic_set(&eb->io_pages, 0);
4344
4345         /*
4346          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4347          */
4348         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4349                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4350         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4351
4352         return eb;
4353 }
4354
4355 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4356 {
4357         unsigned long i;
4358         struct page *p;
4359         struct extent_buffer *new;
4360         unsigned long num_pages = num_extent_pages(src->start, src->len);
4361
4362         new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_NOFS);
4363         if (new == NULL)
4364                 return NULL;
4365
4366         for (i = 0; i < num_pages; i++) {
4367                 p = alloc_page(GFP_NOFS);
4368                 if (!p) {
4369                         btrfs_release_extent_buffer(new);
4370                         return NULL;
4371                 }
4372                 attach_extent_buffer_page(new, p);
4373                 WARN_ON(PageDirty(p));
4374                 SetPageUptodate(p);
4375                 new->pages[i] = p;
4376         }
4377
4378         copy_extent_buffer(new, src, 0, 0, src->len);
4379         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4380         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4381
4382         return new;
4383 }
4384
4385 struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4386 {
4387         struct extent_buffer *eb;
4388         unsigned long num_pages = num_extent_pages(0, len);
4389         unsigned long i;
4390
4391         eb = __alloc_extent_buffer(NULL, start, len, GFP_NOFS);
4392         if (!eb)
4393                 return NULL;
4394
4395         for (i = 0; i < num_pages; i++) {
4396                 eb->pages[i] = alloc_page(GFP_NOFS);
4397                 if (!eb->pages[i])
4398                         goto err;
4399         }
4400         set_extent_buffer_uptodate(eb);
4401         btrfs_set_header_nritems(eb, 0);
4402         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4403
4404         return eb;
4405 err:
4406         for (; i > 0; i--)
4407                 __free_page(eb->pages[i - 1]);
4408         __free_extent_buffer(eb);
4409         return NULL;
4410 }
4411
4412 static void check_buffer_tree_ref(struct extent_buffer *eb)
4413 {
4414         int refs;
4415         /* the ref bit is tricky.  We have to make sure it is set
4416          * if we have the buffer dirty.   Otherwise the
4417          * code to free a buffer can end up dropping a dirty
4418          * page
4419          *
4420          * Once the ref bit is set, it won't go away while the
4421          * buffer is dirty or in writeback, and it also won't
4422          * go away while we have the reference count on the
4423          * eb bumped.
4424          *
4425          * We can't just set the ref bit without bumping the
4426          * ref on the eb because free_extent_buffer might
4427          * see the ref bit and try to clear it.  If this happens
4428          * free_extent_buffer might end up dropping our original
4429          * ref by mistake and freeing the page before we are able
4430          * to add one more ref.
4431          *
4432          * So bump the ref count first, then set the bit.  If someone
4433          * beat us to it, drop the ref we added.
4434          */
4435         refs = atomic_read(&eb->refs);
4436         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4437                 return;
4438
4439         spin_lock(&eb->refs_lock);
4440         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4441                 atomic_inc(&eb->refs);
4442         spin_unlock(&eb->refs_lock);
4443 }
4444
4445 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4446 {
4447         unsigned long num_pages, i;
4448
4449         check_buffer_tree_ref(eb);
4450
4451         num_pages = num_extent_pages(eb->start, eb->len);
4452         for (i = 0; i < num_pages; i++) {
4453                 struct page *p = extent_buffer_page(eb, i);
4454                 mark_page_accessed(p);
4455         }
4456 }
4457
4458 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
4459                                           u64 start, unsigned long len)
4460 {
4461         unsigned long num_pages = num_extent_pages(start, len);
4462         unsigned long i;
4463         unsigned long index = start >> PAGE_CACHE_SHIFT;
4464         struct extent_buffer *eb;
4465         struct extent_buffer *exists = NULL;
4466         struct page *p;
4467         struct address_space *mapping = tree->mapping;
4468         int uptodate = 1;
4469         int ret;
4470
4471         rcu_read_lock();
4472         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4473         if (eb && atomic_inc_not_zero(&eb->refs)) {
4474                 rcu_read_unlock();
4475                 mark_extent_buffer_accessed(eb);
4476                 return eb;
4477         }
4478         rcu_read_unlock();
4479
4480         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
4481         if (!eb)
4482                 return NULL;
4483
4484         for (i = 0; i < num_pages; i++, index++) {
4485                 p = find_or_create_page(mapping, index, GFP_NOFS);
4486                 if (!p)
4487                         goto free_eb;
4488
4489                 spin_lock(&mapping->private_lock);
4490                 if (PagePrivate(p)) {
4491                         /*
4492                          * We could have already allocated an eb for this page
4493                          * and attached one so lets see if we can get a ref on
4494                          * the existing eb, and if we can we know it's good and
4495                          * we can just return that one, else we know we can just
4496                          * overwrite page->private.
4497                          */
4498                         exists = (struct extent_buffer *)p->private;
4499                         if (atomic_inc_not_zero(&exists->refs)) {
4500                                 spin_unlock(&mapping->private_lock);
4501                                 unlock_page(p);
4502                                 page_cache_release(p);
4503                                 mark_extent_buffer_accessed(exists);
4504                                 goto free_eb;
4505                         }
4506
4507                         /*
4508                          * Do this so attach doesn't complain and we need to
4509                          * drop the ref the old guy had.
4510                          */
4511                         ClearPagePrivate(p);
4512                         WARN_ON(PageDirty(p));
4513                         page_cache_release(p);
4514                 }
4515                 attach_extent_buffer_page(eb, p);
4516                 spin_unlock(&mapping->private_lock);
4517                 WARN_ON(PageDirty(p));
4518                 mark_page_accessed(p);
4519                 eb->pages[i] = p;
4520                 if (!PageUptodate(p))
4521                         uptodate = 0;
4522
4523                 /*
4524                  * see below about how we avoid a nasty race with release page
4525                  * and why we unlock later
4526                  */
4527         }
4528         if (uptodate)
4529                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4530 again:
4531         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4532         if (ret)
4533                 goto free_eb;
4534
4535         spin_lock(&tree->buffer_lock);
4536         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
4537         if (ret == -EEXIST) {
4538                 exists = radix_tree_lookup(&tree->buffer,
4539                                                 start >> PAGE_CACHE_SHIFT);
4540                 if (!atomic_inc_not_zero(&exists->refs)) {
4541                         spin_unlock(&tree->buffer_lock);
4542                         radix_tree_preload_end();
4543                         exists = NULL;
4544                         goto again;
4545                 }
4546                 spin_unlock(&tree->buffer_lock);
4547                 radix_tree_preload_end();
4548                 mark_extent_buffer_accessed(exists);
4549                 goto free_eb;
4550         }
4551         /* add one reference for the tree */
4552         check_buffer_tree_ref(eb);
4553         spin_unlock(&tree->buffer_lock);
4554         radix_tree_preload_end();
4555
4556         /*
4557          * there is a race where release page may have
4558          * tried to find this extent buffer in the radix
4559          * but failed.  It will tell the VM it is safe to
4560          * reclaim the, and it will clear the page private bit.
4561          * We must make sure to set the page private bit properly
4562          * after the extent buffer is in the radix tree so
4563          * it doesn't get lost
4564          */
4565         SetPageChecked(eb->pages[0]);
4566         for (i = 1; i < num_pages; i++) {
4567                 p = extent_buffer_page(eb, i);
4568                 ClearPageChecked(p);
4569                 unlock_page(p);
4570         }
4571         unlock_page(eb->pages[0]);
4572         return eb;
4573
4574 free_eb:
4575         for (i = 0; i < num_pages; i++) {
4576                 if (eb->pages[i])
4577                         unlock_page(eb->pages[i]);
4578         }
4579
4580         WARN_ON(!atomic_dec_and_test(&eb->refs));
4581         btrfs_release_extent_buffer(eb);
4582         return exists;
4583 }
4584
4585 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
4586                                          u64 start, unsigned long len)
4587 {
4588         struct extent_buffer *eb;
4589
4590         rcu_read_lock();
4591         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4592         if (eb && atomic_inc_not_zero(&eb->refs)) {
4593                 rcu_read_unlock();
4594                 mark_extent_buffer_accessed(eb);
4595                 return eb;
4596         }
4597         rcu_read_unlock();
4598
4599         return NULL;
4600 }
4601
4602 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4603 {
4604         struct extent_buffer *eb =
4605                         container_of(head, struct extent_buffer, rcu_head);
4606
4607         __free_extent_buffer(eb);
4608 }
4609
4610 /* Expects to have eb->eb_lock already held */
4611 static int release_extent_buffer(struct extent_buffer *eb)
4612 {
4613         WARN_ON(atomic_read(&eb->refs) == 0);
4614         if (atomic_dec_and_test(&eb->refs)) {
4615                 if (test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags)) {
4616                         spin_unlock(&eb->refs_lock);
4617                 } else {
4618                         struct extent_io_tree *tree = eb->tree;
4619
4620                         spin_unlock(&eb->refs_lock);
4621
4622                         spin_lock(&tree->buffer_lock);
4623                         radix_tree_delete(&tree->buffer,
4624                                           eb->start >> PAGE_CACHE_SHIFT);
4625                         spin_unlock(&tree->buffer_lock);
4626                 }
4627
4628                 /* Should be safe to release our pages at this point */
4629                 btrfs_release_extent_buffer_page(eb, 0);
4630                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4631                 return 1;
4632         }
4633         spin_unlock(&eb->refs_lock);
4634
4635         return 0;
4636 }
4637
4638 void free_extent_buffer(struct extent_buffer *eb)
4639 {
4640         int refs;
4641         int old;
4642         if (!eb)
4643                 return;
4644
4645         while (1) {
4646                 refs = atomic_read(&eb->refs);
4647                 if (refs <= 3)
4648                         break;
4649                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
4650                 if (old == refs)
4651                         return;
4652         }
4653
4654         spin_lock(&eb->refs_lock);
4655         if (atomic_read(&eb->refs) == 2 &&
4656             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4657                 atomic_dec(&eb->refs);
4658
4659         if (atomic_read(&eb->refs) == 2 &&
4660             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4661             !extent_buffer_under_io(eb) &&
4662             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4663                 atomic_dec(&eb->refs);
4664
4665         /*
4666          * I know this is terrible, but it's temporary until we stop tracking
4667          * the uptodate bits and such for the extent buffers.
4668          */
4669         release_extent_buffer(eb);
4670 }
4671
4672 void free_extent_buffer_stale(struct extent_buffer *eb)
4673 {
4674         if (!eb)
4675                 return;
4676
4677         spin_lock(&eb->refs_lock);
4678         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4679
4680         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4681             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4682                 atomic_dec(&eb->refs);
4683         release_extent_buffer(eb);
4684 }
4685
4686 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4687 {
4688         unsigned long i;
4689         unsigned long num_pages;
4690         struct page *page;
4691
4692         num_pages = num_extent_pages(eb->start, eb->len);
4693
4694         for (i = 0; i < num_pages; i++) {
4695                 page = extent_buffer_page(eb, i);
4696                 if (!PageDirty(page))
4697                         continue;
4698
4699                 lock_page(page);
4700                 WARN_ON(!PagePrivate(page));
4701
4702                 clear_page_dirty_for_io(page);
4703                 spin_lock_irq(&page->mapping->tree_lock);
4704                 if (!PageDirty(page)) {
4705                         radix_tree_tag_clear(&page->mapping->page_tree,
4706                                                 page_index(page),
4707                                                 PAGECACHE_TAG_DIRTY);
4708                 }
4709                 spin_unlock_irq(&page->mapping->tree_lock);
4710                 ClearPageError(page);
4711                 unlock_page(page);
4712         }
4713         WARN_ON(atomic_read(&eb->refs) == 0);
4714 }
4715
4716 int set_extent_buffer_dirty(struct extent_buffer *eb)
4717 {
4718         unsigned long i;
4719         unsigned long num_pages;
4720         int was_dirty = 0;
4721
4722         check_buffer_tree_ref(eb);
4723
4724         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4725
4726         num_pages = num_extent_pages(eb->start, eb->len);
4727         WARN_ON(atomic_read(&eb->refs) == 0);
4728         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4729
4730         for (i = 0; i < num_pages; i++)
4731                 set_page_dirty(extent_buffer_page(eb, i));
4732         return was_dirty;
4733 }
4734
4735 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4736 {
4737         unsigned long i;
4738         struct page *page;
4739         unsigned long num_pages;
4740
4741         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4742         num_pages = num_extent_pages(eb->start, eb->len);
4743         for (i = 0; i < num_pages; i++) {
4744                 page = extent_buffer_page(eb, i);
4745                 if (page)
4746                         ClearPageUptodate(page);
4747         }
4748         return 0;
4749 }
4750
4751 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4752 {
4753         unsigned long i;
4754         struct page *page;
4755         unsigned long num_pages;
4756
4757         set_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                 SetPageUptodate(page);
4762         }
4763         return 0;
4764 }
4765
4766 int extent_buffer_uptodate(struct extent_buffer *eb)
4767 {
4768         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4769 }
4770
4771 int read_extent_buffer_pages(struct extent_io_tree *tree,
4772                              struct extent_buffer *eb, u64 start, int wait,
4773                              get_extent_t *get_extent, int mirror_num)
4774 {
4775         unsigned long i;
4776         unsigned long start_i;
4777         struct page *page;
4778         int err;
4779         int ret = 0;
4780         int locked_pages = 0;
4781         int all_uptodate = 1;
4782         unsigned long num_pages;
4783         unsigned long num_reads = 0;
4784         struct bio *bio = NULL;
4785         unsigned long bio_flags = 0;
4786
4787         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4788                 return 0;
4789
4790         if (start) {
4791                 WARN_ON(start < eb->start);
4792                 start_i = (start >> PAGE_CACHE_SHIFT) -
4793                         (eb->start >> PAGE_CACHE_SHIFT);
4794         } else {
4795                 start_i = 0;
4796         }
4797
4798         num_pages = num_extent_pages(eb->start, eb->len);
4799         for (i = start_i; i < num_pages; i++) {
4800                 page = extent_buffer_page(eb, i);
4801                 if (wait == WAIT_NONE) {
4802                         if (!trylock_page(page))
4803                                 goto unlock_exit;
4804                 } else {
4805                         lock_page(page);
4806                 }
4807                 locked_pages++;
4808                 if (!PageUptodate(page)) {
4809                         num_reads++;
4810                         all_uptodate = 0;
4811                 }
4812         }
4813         if (all_uptodate) {
4814                 if (start_i == 0)
4815                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4816                 goto unlock_exit;
4817         }
4818
4819         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4820         eb->read_mirror = 0;
4821         atomic_set(&eb->io_pages, num_reads);
4822         for (i = start_i; i < num_pages; i++) {
4823                 page = extent_buffer_page(eb, i);
4824                 if (!PageUptodate(page)) {
4825                         ClearPageError(page);
4826                         err = __extent_read_full_page(tree, page,
4827                                                       get_extent, &bio,
4828                                                       mirror_num, &bio_flags,
4829                                                       READ | REQ_META);
4830                         if (err)
4831                                 ret = err;
4832                 } else {
4833                         unlock_page(page);
4834                 }
4835         }
4836
4837         if (bio) {
4838                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
4839                                      bio_flags);
4840                 if (err)
4841                         return err;
4842         }
4843
4844         if (ret || wait != WAIT_COMPLETE)
4845                 return ret;
4846
4847         for (i = start_i; i < num_pages; i++) {
4848                 page = extent_buffer_page(eb, i);
4849                 wait_on_page_locked(page);
4850                 if (!PageUptodate(page))
4851                         ret = -EIO;
4852         }
4853
4854         return ret;
4855
4856 unlock_exit:
4857         i = start_i;
4858         while (locked_pages > 0) {
4859                 page = extent_buffer_page(eb, i);
4860                 i++;
4861                 unlock_page(page);
4862                 locked_pages--;
4863         }
4864         return ret;
4865 }
4866
4867 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4868                         unsigned long start,
4869                         unsigned long len)
4870 {
4871         size_t cur;
4872         size_t offset;
4873         struct page *page;
4874         char *kaddr;
4875         char *dst = (char *)dstv;
4876         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4877         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4878
4879         WARN_ON(start > eb->len);
4880         WARN_ON(start + len > eb->start + eb->len);
4881
4882         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
4883
4884         while (len > 0) {
4885                 page = extent_buffer_page(eb, i);
4886
4887                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4888                 kaddr = page_address(page);
4889                 memcpy(dst, kaddr + offset, cur);
4890
4891                 dst += cur;
4892                 len -= cur;
4893                 offset = 0;
4894                 i++;
4895         }
4896 }
4897
4898 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4899                                unsigned long min_len, char **map,
4900                                unsigned long *map_start,
4901                                unsigned long *map_len)
4902 {
4903         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4904         char *kaddr;
4905         struct page *p;
4906         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4907         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4908         unsigned long end_i = (start_offset + start + min_len - 1) >>
4909                 PAGE_CACHE_SHIFT;
4910
4911         if (i != end_i)
4912                 return -EINVAL;
4913
4914         if (i == 0) {
4915                 offset = start_offset;
4916                 *map_start = 0;
4917         } else {
4918                 offset = 0;
4919                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4920         }
4921
4922         if (start + min_len > eb->len) {
4923                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4924                        "wanted %lu %lu\n",
4925                        eb->start, eb->len, start, min_len);
4926                 return -EINVAL;
4927         }
4928
4929         p = extent_buffer_page(eb, i);
4930         kaddr = page_address(p);
4931         *map = kaddr + offset;
4932         *map_len = PAGE_CACHE_SIZE - offset;
4933         return 0;
4934 }
4935
4936 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4937                           unsigned long start,
4938                           unsigned long len)
4939 {
4940         size_t cur;
4941         size_t offset;
4942         struct page *page;
4943         char *kaddr;
4944         char *ptr = (char *)ptrv;
4945         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4946         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4947         int ret = 0;
4948
4949         WARN_ON(start > eb->len);
4950         WARN_ON(start + len > eb->start + eb->len);
4951
4952         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
4953
4954         while (len > 0) {
4955                 page = extent_buffer_page(eb, i);
4956
4957                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4958
4959                 kaddr = page_address(page);
4960                 ret = memcmp(ptr, kaddr + offset, cur);
4961                 if (ret)
4962                         break;
4963
4964                 ptr += cur;
4965                 len -= cur;
4966                 offset = 0;
4967                 i++;
4968         }
4969         return ret;
4970 }
4971
4972 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4973                          unsigned long start, unsigned long len)
4974 {
4975         size_t cur;
4976         size_t offset;
4977         struct page *page;
4978         char *kaddr;
4979         char *src = (char *)srcv;
4980         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4981         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4982
4983         WARN_ON(start > eb->len);
4984         WARN_ON(start + len > eb->start + eb->len);
4985
4986         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
4987
4988         while (len > 0) {
4989                 page = extent_buffer_page(eb, i);
4990                 WARN_ON(!PageUptodate(page));
4991
4992                 cur = min(len, PAGE_CACHE_SIZE - offset);
4993                 kaddr = page_address(page);
4994                 memcpy(kaddr + offset, src, cur);
4995
4996                 src += cur;
4997                 len -= cur;
4998                 offset = 0;
4999                 i++;
5000         }
5001 }
5002
5003 void memset_extent_buffer(struct extent_buffer *eb, char c,
5004                           unsigned long start, unsigned long len)
5005 {
5006         size_t cur;
5007         size_t offset;
5008         struct page *page;
5009         char *kaddr;
5010         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5011         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5012
5013         WARN_ON(start > eb->len);
5014         WARN_ON(start + len > eb->start + eb->len);
5015
5016         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5017
5018         while (len > 0) {
5019                 page = extent_buffer_page(eb, i);
5020                 WARN_ON(!PageUptodate(page));
5021
5022                 cur = min(len, PAGE_CACHE_SIZE - offset);
5023                 kaddr = page_address(page);
5024                 memset(kaddr + offset, c, cur);
5025
5026                 len -= cur;
5027                 offset = 0;
5028                 i++;
5029         }
5030 }
5031
5032 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5033                         unsigned long dst_offset, unsigned long src_offset,
5034                         unsigned long len)
5035 {
5036         u64 dst_len = dst->len;
5037         size_t cur;
5038         size_t offset;
5039         struct page *page;
5040         char *kaddr;
5041         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5042         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5043
5044         WARN_ON(src->len != dst_len);
5045
5046         offset = (start_offset + dst_offset) &
5047                 (PAGE_CACHE_SIZE - 1);
5048
5049         while (len > 0) {
5050                 page = extent_buffer_page(dst, i);
5051                 WARN_ON(!PageUptodate(page));
5052
5053                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5054
5055                 kaddr = page_address(page);
5056                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5057
5058                 src_offset += cur;
5059                 len -= cur;
5060                 offset = 0;
5061                 i++;
5062         }
5063 }
5064
5065 static void move_pages(struct page *dst_page, struct page *src_page,
5066                        unsigned long dst_off, unsigned long src_off,
5067                        unsigned long len)
5068 {
5069         char *dst_kaddr = page_address(dst_page);
5070         if (dst_page == src_page) {
5071                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
5072         } else {
5073                 char *src_kaddr = page_address(src_page);
5074                 char *p = dst_kaddr + dst_off + len;
5075                 char *s = src_kaddr + src_off + len;
5076
5077                 while (len--)
5078                         *--p = *--s;
5079         }
5080 }
5081
5082 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5083 {
5084         unsigned long distance = (src > dst) ? src - dst : dst - src;
5085         return distance < len;
5086 }
5087
5088 static void copy_pages(struct page *dst_page, struct page *src_page,
5089                        unsigned long dst_off, unsigned long src_off,
5090                        unsigned long len)
5091 {
5092         char *dst_kaddr = page_address(dst_page);
5093         char *src_kaddr;
5094         int must_memmove = 0;
5095
5096         if (dst_page != src_page) {
5097                 src_kaddr = page_address(src_page);
5098         } else {
5099                 src_kaddr = dst_kaddr;
5100                 if (areas_overlap(src_off, dst_off, len))
5101                         must_memmove = 1;
5102         }
5103
5104         if (must_memmove)
5105                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5106         else
5107                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5108 }
5109
5110 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5111                            unsigned long src_offset, unsigned long len)
5112 {
5113         size_t cur;
5114         size_t dst_off_in_page;
5115         size_t src_off_in_page;
5116         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5117         unsigned long dst_i;
5118         unsigned long src_i;
5119
5120         if (src_offset + len > dst->len) {
5121                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
5122                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5123                 BUG_ON(1);
5124         }
5125         if (dst_offset + len > dst->len) {
5126                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
5127                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5128                 BUG_ON(1);
5129         }
5130
5131         while (len > 0) {
5132                 dst_off_in_page = (start_offset + dst_offset) &
5133                         (PAGE_CACHE_SIZE - 1);
5134                 src_off_in_page = (start_offset + src_offset) &
5135                         (PAGE_CACHE_SIZE - 1);
5136
5137                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5138                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5139
5140                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5141                                                src_off_in_page));
5142                 cur = min_t(unsigned long, cur,
5143                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5144
5145                 copy_pages(extent_buffer_page(dst, dst_i),
5146                            extent_buffer_page(dst, src_i),
5147                            dst_off_in_page, src_off_in_page, cur);
5148
5149                 src_offset += cur;
5150                 dst_offset += cur;
5151                 len -= cur;
5152         }
5153 }
5154
5155 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5156                            unsigned long src_offset, unsigned long len)
5157 {
5158         size_t cur;
5159         size_t dst_off_in_page;
5160         size_t src_off_in_page;
5161         unsigned long dst_end = dst_offset + len - 1;
5162         unsigned long src_end = src_offset + len - 1;
5163         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5164         unsigned long dst_i;
5165         unsigned long src_i;
5166
5167         if (src_offset + len > dst->len) {
5168                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
5169                        "len %lu len %lu\n", src_offset, len, dst->len);
5170                 BUG_ON(1);
5171         }
5172         if (dst_offset + len > dst->len) {
5173                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
5174                        "len %lu len %lu\n", dst_offset, len, dst->len);
5175                 BUG_ON(1);
5176         }
5177         if (dst_offset < src_offset) {
5178                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5179                 return;
5180         }
5181         while (len > 0) {
5182                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5183                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5184
5185                 dst_off_in_page = (start_offset + dst_end) &
5186                         (PAGE_CACHE_SIZE - 1);
5187                 src_off_in_page = (start_offset + src_end) &
5188                         (PAGE_CACHE_SIZE - 1);
5189
5190                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5191                 cur = min(cur, dst_off_in_page + 1);
5192                 move_pages(extent_buffer_page(dst, dst_i),
5193                            extent_buffer_page(dst, src_i),
5194                            dst_off_in_page - cur + 1,
5195                            src_off_in_page - cur + 1, cur);
5196
5197                 dst_end -= cur;
5198                 src_end -= cur;
5199                 len -= cur;
5200         }
5201 }
5202
5203 int try_release_extent_buffer(struct page *page)
5204 {
5205         struct extent_buffer *eb;
5206
5207         /*
5208          * We need to make sure noboody is attaching this page to an eb right
5209          * now.
5210          */
5211         spin_lock(&page->mapping->private_lock);
5212         if (!PagePrivate(page)) {
5213                 spin_unlock(&page->mapping->private_lock);
5214                 return 1;
5215         }
5216
5217         eb = (struct extent_buffer *)page->private;
5218         BUG_ON(!eb);
5219
5220         /*
5221          * This is a little awful but should be ok, we need to make sure that
5222          * the eb doesn't disappear out from under us while we're looking at
5223          * this page.
5224          */
5225         spin_lock(&eb->refs_lock);
5226         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5227                 spin_unlock(&eb->refs_lock);
5228                 spin_unlock(&page->mapping->private_lock);
5229                 return 0;
5230         }
5231         spin_unlock(&page->mapping->private_lock);
5232
5233         /*
5234          * If tree ref isn't set then we know the ref on this eb is a real ref,
5235          * so just return, this page will likely be freed soon anyway.
5236          */
5237         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5238                 spin_unlock(&eb->refs_lock);
5239                 return 0;
5240         }
5241
5242         return release_extent_buffer(eb);
5243 }