fs/btrfs/extent_io.c

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