fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26 static struct bio_set *btrfs_bioset;
  27
  28 #ifdef CONFIG_BTRFS_DEBUG
  29 static LIST_HEAD(buffers);
  30 static LIST_HEAD(states);
  31
  32 static DEFINE_SPINLOCK(leak_lock);
  33
  34 static inline
  35 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  36 {
  37         unsigned long flags;
  38
  39         spin_lock_irqsave(&leak_lock, flags);
  40         list_add(new, head);
  41         spin_unlock_irqrestore(&leak_lock, flags);
  42 }
  43
  44 static inline
  45 void btrfs_leak_debug_del(struct list_head *entry)
  46 {
  47         unsigned long flags;
  48
  49         spin_lock_irqsave(&leak_lock, flags);
  50         list_del(entry);
  51         spin_unlock_irqrestore(&leak_lock, flags);
  52 }
  53
  54 static inline
  55 void btrfs_leak_debug_check(void)
  56 {
  57         struct extent_state *state;
  58         struct extent_buffer *eb;
  59
  60         while (!list_empty(&states)) {
  61                 state = list_entry(states.next, struct extent_state, leak_list);
  62                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  63                        "state %lu in tree %p refs %d\n",
  64                        state->start, state->end, state->state, state->tree,
  65                        atomic_read(&state->refs));
  66                 list_del(&state->leak_list);
  67                 kmem_cache_free(extent_state_cache, state);
  68         }
  69
  70         while (!list_empty(&buffers)) {
  71                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  72                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  73                        "refs %d\n",
  74                        eb->start, eb->len, atomic_read(&eb->refs));
  75                 list_del(&eb->leak_list);
  76                 kmem_cache_free(extent_buffer_cache, eb);
  77         }
  78 }
  79
  80 #define btrfs_debug_check_extent_io_range(inode, start, end)            \
  81         __btrfs_debug_check_extent_io_range(__func__, (inode), (start), (end))
  82 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  83                 struct inode *inode, u64 start, u64 end)
  84 {
  85         u64 isize = i_size_read(inode);
  86
  87         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  88                 printk_ratelimited(KERN_DEBUG
  89                     "btrfs: %s: ino %llu isize %llu odd range [%llu,%llu]\n",
  90                                 caller, btrfs_ino(inode), isize, start, end);
  91         }
  92 }
  93 #else
  94 #define btrfs_leak_debug_add(new, head) do {} while (0)
  95 #define btrfs_leak_debug_del(entry)     do {} while (0)
  96 #define btrfs_leak_debug_check()        do {} while (0)
  97 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
  98 #endif
  99
 100 #define BUFFER_LRU_MAX 64
 101
 102 struct tree_entry {
 103         u64 start;
 104         u64 end;
 105         struct rb_node rb_node;
 106 };
 107
 108 struct extent_page_data {
 109         struct bio *bio;
 110         struct extent_io_tree *tree;
 111         get_extent_t *get_extent;
 112         unsigned long bio_flags;
 113
 114         /* tells writepage not to lock the state bits for this range
 115          * it still does the unlocking
 116          */
 117         unsigned int extent_locked:1;
 118
 119         /* tells the submit_bio code to use a WRITE_SYNC */
 120         unsigned int sync_io:1;
 121 };
 122
 123 static noinline void flush_write_bio(void *data);
 124 static inline struct btrfs_fs_info *
 125 tree_fs_info(struct extent_io_tree *tree)
 126 {
 127         return btrfs_sb(tree->mapping->host->i_sb);
 128 }
 129
 130 int __init extent_io_init(void)
 131 {
 132         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 133                         sizeof(struct extent_state), 0,
 134                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 135         if (!extent_state_cache)
 136                 return -ENOMEM;
 137
 138         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 139                         sizeof(struct extent_buffer), 0,
 140                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 141         if (!extent_buffer_cache)
 142                 goto free_state_cache;
 143
 144         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 145                                      offsetof(struct btrfs_io_bio, bio));
 146         if (!btrfs_bioset)
 147                 goto free_buffer_cache;
 148
 149         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 150                 goto free_bioset;
 151
 152         return 0;
 153
 154 free_bioset:
 155         bioset_free(btrfs_bioset);
 156         btrfs_bioset = NULL;
 157
 158 free_buffer_cache:
 159         kmem_cache_destroy(extent_buffer_cache);
 160         extent_buffer_cache = NULL;
 161
 162 free_state_cache:
 163         kmem_cache_destroy(extent_state_cache);
 164         extent_state_cache = NULL;
 165         return -ENOMEM;
 166 }
 167
 168 void extent_io_exit(void)
 169 {
 170         btrfs_leak_debug_check();
 171
 172         /*
 173          * Make sure all delayed rcu free are flushed before we
 174          * destroy caches.
 175          */
 176         rcu_barrier();
 177         if (extent_state_cache)
 178                 kmem_cache_destroy(extent_state_cache);
 179         if (extent_buffer_cache)
 180                 kmem_cache_destroy(extent_buffer_cache);
 181         if (btrfs_bioset)
 182                 bioset_free(btrfs_bioset);
 183 }
 184
 185 void extent_io_tree_init(struct extent_io_tree *tree,
 186                          struct address_space *mapping)
 187 {
 188         tree->state = RB_ROOT;
 189         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 190         tree->ops = NULL;
 191         tree->dirty_bytes = 0;
 192         spin_lock_init(&tree->lock);
 193         spin_lock_init(&tree->buffer_lock);
 194         tree->mapping = mapping;
 195 }
 196
 197 static struct extent_state *alloc_extent_state(gfp_t mask)
 198 {
 199         struct extent_state *state;
 200
 201         state = kmem_cache_alloc(extent_state_cache, mask);
 202         if (!state)
 203                 return state;
 204         state->state = 0;
 205         state->private = 0;
 206         state->tree = NULL;
 207         btrfs_leak_debug_add(&state->leak_list, &states);
 208         atomic_set(&state->refs, 1);
 209         init_waitqueue_head(&state->wq);
 210         trace_alloc_extent_state(state, mask, _RET_IP_);
 211         return state;
 212 }
 213
 214 void free_extent_state(struct extent_state *state)
 215 {
 216         if (!state)
 217                 return;
 218         if (atomic_dec_and_test(&state->refs)) {
 219                 WARN_ON(state->tree);
 220                 btrfs_leak_debug_del(&state->leak_list);
 221                 trace_free_extent_state(state, _RET_IP_);
 222                 kmem_cache_free(extent_state_cache, state);
 223         }
 224 }
 225
 226 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 227                                    struct rb_node *node)
 228 {
 229         struct rb_node **p = &root->rb_node;
 230         struct rb_node *parent = NULL;
 231         struct tree_entry *entry;
 232
 233         while (*p) {
 234                 parent = *p;
 235                 entry = rb_entry(parent, struct tree_entry, rb_node);
 236
 237                 if (offset < entry->start)
 238                         p = &(*p)->rb_left;
 239                 else if (offset > entry->end)
 240                         p = &(*p)->rb_right;
 241                 else
 242                         return parent;
 243         }
 244
 245         rb_link_node(node, parent, p);
 246         rb_insert_color(node, root);
 247         return NULL;
 248 }
 249
 250 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 251                                      struct rb_node **prev_ret,
 252                                      struct rb_node **next_ret)
 253 {
 254         struct rb_root *root = &tree->state;
 255         struct rb_node *n = root->rb_node;
 256         struct rb_node *prev = NULL;
 257         struct rb_node *orig_prev = NULL;
 258         struct tree_entry *entry;
 259         struct tree_entry *prev_entry = NULL;
 260
 261         while (n) {
 262                 entry = rb_entry(n, struct tree_entry, rb_node);
 263                 prev = n;
 264                 prev_entry = entry;
 265
 266                 if (offset < entry->start)
 267                         n = n->rb_left;
 268                 else if (offset > entry->end)
 269                         n = n->rb_right;
 270                 else
 271                         return n;
 272         }
 273
 274         if (prev_ret) {
 275                 orig_prev = prev;
 276                 while (prev && offset > prev_entry->end) {
 277                         prev = rb_next(prev);
 278                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 279                 }
 280                 *prev_ret = prev;
 281                 prev = orig_prev;
 282         }
 283
 284         if (next_ret) {
 285                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 286                 while (prev && offset < prev_entry->start) {
 287                         prev = rb_prev(prev);
 288                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 289                 }
 290                 *next_ret = prev;
 291         }
 292         return NULL;
 293 }
 294
 295 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 296                                           u64 offset)
 297 {
 298         struct rb_node *prev = NULL;
 299         struct rb_node *ret;
 300
 301         ret = __etree_search(tree, offset, &prev, NULL);
 302         if (!ret)
 303                 return prev;
 304         return ret;
 305 }
 306
 307 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 308                      struct extent_state *other)
 309 {
 310         if (tree->ops && tree->ops->merge_extent_hook)
 311                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 312                                              other);
 313 }
 314
 315 /*
 316  * utility function to look for merge candidates inside a given range.
 317  * Any extents with matching state are merged together into a single
 318  * extent in the tree.  Extents with EXTENT_IO in their state field
 319  * are not merged because the end_io handlers need to be able to do
 320  * operations on them without sleeping (or doing allocations/splits).
 321  *
 322  * This should be called with the tree lock held.
 323  */
 324 static void merge_state(struct extent_io_tree *tree,
 325                         struct extent_state *state)
 326 {
 327         struct extent_state *other;
 328         struct rb_node *other_node;
 329
 330         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 331                 return;
 332
 333         other_node = rb_prev(&state->rb_node);
 334         if (other_node) {
 335                 other = rb_entry(other_node, struct extent_state, rb_node);
 336                 if (other->end == state->start - 1 &&
 337                     other->state == state->state) {
 338                         merge_cb(tree, state, other);
 339                         state->start = other->start;
 340                         other->tree = NULL;
 341                         rb_erase(&other->rb_node, &tree->state);
 342                         free_extent_state(other);
 343                 }
 344         }
 345         other_node = rb_next(&state->rb_node);
 346         if (other_node) {
 347                 other = rb_entry(other_node, struct extent_state, rb_node);
 348                 if (other->start == state->end + 1 &&
 349                     other->state == state->state) {
 350                         merge_cb(tree, state, other);
 351                         state->end = other->end;
 352                         other->tree = NULL;
 353                         rb_erase(&other->rb_node, &tree->state);
 354                         free_extent_state(other);
 355                 }
 356         }
 357 }
 358
 359 static void set_state_cb(struct extent_io_tree *tree,
 360                          struct extent_state *state, unsigned long *bits)
 361 {
 362         if (tree->ops && tree->ops->set_bit_hook)
 363                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 364 }
 365
 366 static void clear_state_cb(struct extent_io_tree *tree,
 367                            struct extent_state *state, unsigned long *bits)
 368 {
 369         if (tree->ops && tree->ops->clear_bit_hook)
 370                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 371 }
 372
 373 static void set_state_bits(struct extent_io_tree *tree,
 374                            struct extent_state *state, unsigned long *bits);
 375
 376 /*
 377  * insert an extent_state struct into the tree.  'bits' are set on the
 378  * struct before it is inserted.
 379  *
 380  * This may return -EEXIST if the extent is already there, in which case the
 381  * state struct is freed.
 382  *
 383  * The tree lock is not taken internally.  This is a utility function and
 384  * probably isn't what you want to call (see set/clear_extent_bit).
 385  */
 386 static int insert_state(struct extent_io_tree *tree,
 387                         struct extent_state *state, u64 start, u64 end,
 388                         unsigned long *bits)
 389 {
 390         struct rb_node *node;
 391
 392         if (end < start)
 393                 WARN(1, KERN_ERR "btrfs end < start %llu %llu\n",
 394                        end, start);
 395         state->start = start;
 396         state->end = end;
 397
 398         set_state_bits(tree, state, bits);
 399
 400         node = tree_insert(&tree->state, end, &state->rb_node);
 401         if (node) {
 402                 struct extent_state *found;
 403                 found = rb_entry(node, struct extent_state, rb_node);
 404                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 405                        "%llu %llu\n",
 406                        found->start, found->end, start, end);
 407                 return -EEXIST;
 408         }
 409         state->tree = tree;
 410         merge_state(tree, state);
 411         return 0;
 412 }
 413
 414 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 415                      u64 split)
 416 {
 417         if (tree->ops && tree->ops->split_extent_hook)
 418                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 419 }
 420
 421 /*
 422  * split a given extent state struct in two, inserting the preallocated
 423  * struct 'prealloc' as the newly created second half.  'split' indicates an
 424  * offset inside 'orig' where it should be split.
 425  *
 426  * Before calling,
 427  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 428  * are two extent state structs in the tree:
 429  * prealloc: [orig->start, split - 1]
 430  * orig: [ split, orig->end ]
 431  *
 432  * The tree locks are not taken by this function. They need to be held
 433  * by the caller.
 434  */
 435 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 436                        struct extent_state *prealloc, u64 split)
 437 {
 438         struct rb_node *node;
 439
 440         split_cb(tree, orig, split);
 441
 442         prealloc->start = orig->start;
 443         prealloc->end = split - 1;
 444         prealloc->state = orig->state;
 445         orig->start = split;
 446
 447         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 448         if (node) {
 449                 free_extent_state(prealloc);
 450                 return -EEXIST;
 451         }
 452         prealloc->tree = tree;
 453         return 0;
 454 }
 455
 456 static struct extent_state *next_state(struct extent_state *state)
 457 {
 458         struct rb_node *next = rb_next(&state->rb_node);
 459         if (next)
 460                 return rb_entry(next, struct extent_state, rb_node);
 461         else
 462                 return NULL;
 463 }
 464
 465 /*
 466  * utility function to clear some bits in an extent state struct.
 467  * it will optionally wake up any one waiting on this state (wake == 1).
 468  *
 469  * If no bits are set on the state struct after clearing things, the
 470  * struct is freed and removed from the tree
 471  */
 472 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 473                                             struct extent_state *state,
 474                                             unsigned long *bits, int wake)
 475 {
 476         struct extent_state *next;
 477         unsigned long bits_to_clear = *bits & ~EXTENT_CTLBITS;
 478
 479         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 480                 u64 range = state->end - state->start + 1;
 481                 WARN_ON(range > tree->dirty_bytes);
 482                 tree->dirty_bytes -= range;
 483         }
 484         clear_state_cb(tree, state, bits);
 485         state->state &= ~bits_to_clear;
 486         if (wake)
 487                 wake_up(&state->wq);
 488         if (state->state == 0) {
 489                 next = next_state(state);
 490                 if (state->tree) {
 491                         rb_erase(&state->rb_node, &tree->state);
 492                         state->tree = NULL;
 493                         free_extent_state(state);
 494                 } else {
 495                         WARN_ON(1);
 496                 }
 497         } else {
 498                 merge_state(tree, state);
 499                 next = next_state(state);
 500         }
 501         return next;
 502 }
 503
 504 static struct extent_state *
 505 alloc_extent_state_atomic(struct extent_state *prealloc)
 506 {
 507         if (!prealloc)
 508                 prealloc = alloc_extent_state(GFP_ATOMIC);
 509
 510         return prealloc;
 511 }
 512
 513 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 514 {
 515         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 516                     "Extent tree was modified by another "
 517                     "thread while locked.");
 518 }
 519
 520 /*
 521  * clear some bits on a range in the tree.  This may require splitting
 522  * or inserting elements in the tree, so the gfp mask is used to
 523  * indicate which allocations or sleeping are allowed.
 524  *
 525  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 526  * the given range from the tree regardless of state (ie for truncate).
 527  *
 528  * the range [start, end] is inclusive.
 529  *
 530  * This takes the tree lock, and returns 0 on success and < 0 on error.
 531  */
 532 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 533                      unsigned long bits, int wake, int delete,
 534                      struct extent_state **cached_state,
 535                      gfp_t mask)
 536 {
 537         struct extent_state *state;
 538         struct extent_state *cached;
 539         struct extent_state *prealloc = NULL;
 540         struct rb_node *node;
 541         u64 last_end;
 542         int err;
 543         int clear = 0;
 544
 545         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
 546
 547         if (bits & EXTENT_DELALLOC)
 548                 bits |= EXTENT_NORESERVE;
 549
 550         if (delete)
 551                 bits |= ~EXTENT_CTLBITS;
 552         bits |= EXTENT_FIRST_DELALLOC;
 553
 554         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 555                 clear = 1;
 556 again:
 557         if (!prealloc && (mask & __GFP_WAIT)) {
 558                 prealloc = alloc_extent_state(mask);
 559                 if (!prealloc)
 560                         return -ENOMEM;
 561         }
 562
 563         spin_lock(&tree->lock);
 564         if (cached_state) {
 565                 cached = *cached_state;
 566
 567                 if (clear) {
 568                         *cached_state = NULL;
 569                         cached_state = NULL;
 570                 }
 571
 572                 if (cached && cached->tree && cached->start <= start &&
 573                     cached->end > start) {
 574                         if (clear)
 575                                 atomic_dec(&cached->refs);
 576                         state = cached;
 577                         goto hit_next;
 578                 }
 579                 if (clear)
 580                         free_extent_state(cached);
 581         }
 582         /*
 583          * this search will find the extents that end after
 584          * our range starts
 585          */
 586         node = tree_search(tree, start);
 587         if (!node)
 588                 goto out;
 589         state = rb_entry(node, struct extent_state, rb_node);
 590 hit_next:
 591         if (state->start > end)
 592                 goto out;
 593         WARN_ON(state->end < start);
 594         last_end = state->end;
 595
 596         /* the state doesn't have the wanted bits, go ahead */
 597         if (!(state->state & bits)) {
 598                 state = next_state(state);
 599                 goto next;
 600         }
 601
 602         /*
 603          *     | ---- desired range ---- |
 604          *  | state | or
 605          *  | ------------- state -------------- |
 606          *
 607          * We need to split the extent we found, and may flip
 608          * bits on second half.
 609          *
 610          * If the extent we found extends past our range, we
 611          * just split and search again.  It'll get split again
 612          * the next time though.
 613          *
 614          * If the extent we found is inside our range, we clear
 615          * the desired bit on it.
 616          */
 617
 618         if (state->start < start) {
 619                 prealloc = alloc_extent_state_atomic(prealloc);
 620                 BUG_ON(!prealloc);
 621                 err = split_state(tree, state, prealloc, start);
 622                 if (err)
 623                         extent_io_tree_panic(tree, err);
 624
 625                 prealloc = NULL;
 626                 if (err)
 627                         goto out;
 628                 if (state->end <= end) {
 629                         state = clear_state_bit(tree, state, &bits, wake);
 630                         goto next;
 631                 }
 632                 goto search_again;
 633         }
 634         /*
 635          * | ---- desired range ---- |
 636          *                        | state |
 637          * We need to split the extent, and clear the bit
 638          * on the first half
 639          */
 640         if (state->start <= end && state->end > end) {
 641                 prealloc = alloc_extent_state_atomic(prealloc);
 642                 BUG_ON(!prealloc);
 643                 err = split_state(tree, state, prealloc, end + 1);
 644                 if (err)
 645                         extent_io_tree_panic(tree, err);
 646
 647                 if (wake)
 648                         wake_up(&state->wq);
 649
 650                 clear_state_bit(tree, prealloc, &bits, wake);
 651
 652                 prealloc = NULL;
 653                 goto out;
 654         }
 655
 656         state = clear_state_bit(tree, state, &bits, wake);
 657 next:
 658         if (last_end == (u64)-1)
 659                 goto out;
 660         start = last_end + 1;
 661         if (start <= end && state && !need_resched())
 662                 goto hit_next;
 663         goto search_again;
 664
 665 out:
 666         spin_unlock(&tree->lock);
 667         if (prealloc)
 668                 free_extent_state(prealloc);
 669
 670         return 0;
 671
 672 search_again:
 673         if (start > end)
 674                 goto out;
 675         spin_unlock(&tree->lock);
 676         if (mask & __GFP_WAIT)
 677                 cond_resched();
 678         goto again;
 679 }
 680
 681 static void wait_on_state(struct extent_io_tree *tree,
 682                           struct extent_state *state)
 683                 __releases(tree->lock)
 684                 __acquires(tree->lock)
 685 {
 686         DEFINE_WAIT(wait);
 687         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 688         spin_unlock(&tree->lock);
 689         schedule();
 690         spin_lock(&tree->lock);
 691         finish_wait(&state->wq, &wait);
 692 }
 693
 694 /*
 695  * waits for one or more bits to clear on a range in the state tree.
 696  * The range [start, end] is inclusive.
 697  * The tree lock is taken by this function
 698  */
 699 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 700                             unsigned long bits)
 701 {
 702         struct extent_state *state;
 703         struct rb_node *node;
 704
 705         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
 706
 707         spin_lock(&tree->lock);
 708 again:
 709         while (1) {
 710                 /*
 711                  * this search will find all the extents that end after
 712                  * our range starts
 713                  */
 714                 node = tree_search(tree, start);
 715                 if (!node)
 716                         break;
 717
 718                 state = rb_entry(node, struct extent_state, rb_node);
 719
 720                 if (state->start > end)
 721                         goto out;
 722
 723                 if (state->state & bits) {
 724                         start = state->start;
 725                         atomic_inc(&state->refs);
 726                         wait_on_state(tree, state);
 727                         free_extent_state(state);
 728                         goto again;
 729                 }
 730                 start = state->end + 1;
 731
 732                 if (start > end)
 733                         break;
 734
 735                 cond_resched_lock(&tree->lock);
 736         }
 737 out:
 738         spin_unlock(&tree->lock);
 739 }
 740
 741 static void set_state_bits(struct extent_io_tree *tree,
 742                            struct extent_state *state,
 743                            unsigned long *bits)
 744 {
 745         unsigned long bits_to_set = *bits & ~EXTENT_CTLBITS;
 746
 747         set_state_cb(tree, state, bits);
 748         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 749                 u64 range = state->end - state->start + 1;
 750                 tree->dirty_bytes += range;
 751         }
 752         state->state |= bits_to_set;
 753 }
 754
 755 static void cache_state(struct extent_state *state,
 756                         struct extent_state **cached_ptr)
 757 {
 758         if (cached_ptr && !(*cached_ptr)) {
 759                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 760                         *cached_ptr = state;
 761                         atomic_inc(&state->refs);
 762                 }
 763         }
 764 }
 765
 766 /*
 767  * set some bits on a range in the tree.  This may require allocations or
 768  * sleeping, so the gfp mask is used to indicate what is allowed.
 769  *
 770  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 771  * part of the range already has the desired bits set.  The start of the
 772  * existing range is returned in failed_start in this case.
 773  *
 774  * [start, end] is inclusive This takes the tree lock.
 775  */
 776
 777 static int __must_check
 778 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 779                  unsigned long bits, unsigned long exclusive_bits,
 780                  u64 *failed_start, struct extent_state **cached_state,
 781                  gfp_t mask)
 782 {
 783         struct extent_state *state;
 784         struct extent_state *prealloc = NULL;
 785         struct rb_node *node;
 786         int err = 0;
 787         u64 last_start;
 788         u64 last_end;
 789
 790         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
 791
 792         bits |= EXTENT_FIRST_DELALLOC;
 793 again:
 794         if (!prealloc && (mask & __GFP_WAIT)) {
 795                 prealloc = alloc_extent_state(mask);
 796                 BUG_ON(!prealloc);
 797         }
 798
 799         spin_lock(&tree->lock);
 800         if (cached_state && *cached_state) {
 801                 state = *cached_state;
 802                 if (state->start <= start && state->end > start &&
 803                     state->tree) {
 804                         node = &state->rb_node;
 805                         goto hit_next;
 806                 }
 807         }
 808         /*
 809          * this search will find all the extents that end after
 810          * our range starts.
 811          */
 812         node = tree_search(tree, start);
 813         if (!node) {
 814                 prealloc = alloc_extent_state_atomic(prealloc);
 815                 BUG_ON(!prealloc);
 816                 err = insert_state(tree, prealloc, start, end, &bits);
 817                 if (err)
 818                         extent_io_tree_panic(tree, err);
 819
 820                 prealloc = NULL;
 821                 goto out;
 822         }
 823         state = rb_entry(node, struct extent_state, rb_node);
 824 hit_next:
 825         last_start = state->start;
 826         last_end = state->end;
 827
 828         /*
 829          * | ---- desired range ---- |
 830          * | state |
 831          *
 832          * Just lock what we found and keep going
 833          */
 834         if (state->start == start && state->end <= end) {
 835                 if (state->state & exclusive_bits) {
 836                         *failed_start = state->start;
 837                         err = -EEXIST;
 838                         goto out;
 839                 }
 840
 841                 set_state_bits(tree, state, &bits);
 842                 cache_state(state, cached_state);
 843                 merge_state(tree, state);
 844                 if (last_end == (u64)-1)
 845                         goto out;
 846                 start = last_end + 1;
 847                 state = next_state(state);
 848                 if (start < end && state && state->start == start &&
 849                     !need_resched())
 850                         goto hit_next;
 851                 goto search_again;
 852         }
 853
 854         /*
 855          *     | ---- desired range ---- |
 856          * | state |
 857          *   or
 858          * | ------------- state -------------- |
 859          *
 860          * We need to split the extent we found, and may flip bits on
 861          * second half.
 862          *
 863          * If the extent we found extends past our
 864          * range, we just split and search again.  It'll get split
 865          * again the next time though.
 866          *
 867          * If the extent we found is inside our range, we set the
 868          * desired bit on it.
 869          */
 870         if (state->start < start) {
 871                 if (state->state & exclusive_bits) {
 872                         *failed_start = start;
 873                         err = -EEXIST;
 874                         goto out;
 875                 }
 876
 877                 prealloc = alloc_extent_state_atomic(prealloc);
 878                 BUG_ON(!prealloc);
 879                 err = split_state(tree, state, prealloc, start);
 880                 if (err)
 881                         extent_io_tree_panic(tree, err);
 882
 883                 prealloc = NULL;
 884                 if (err)
 885                         goto out;
 886                 if (state->end <= end) {
 887                         set_state_bits(tree, state, &bits);
 888                         cache_state(state, cached_state);
 889                         merge_state(tree, state);
 890                         if (last_end == (u64)-1)
 891                                 goto out;
 892                         start = last_end + 1;
 893                         state = next_state(state);
 894                         if (start < end && state && state->start == start &&
 895                             !need_resched())
 896                                 goto hit_next;
 897                 }
 898                 goto search_again;
 899         }
 900         /*
 901          * | ---- desired range ---- |
 902          *     | state | or               | state |
 903          *
 904          * There's a hole, we need to insert something in it and
 905          * ignore the extent we found.
 906          */
 907         if (state->start > start) {
 908                 u64 this_end;
 909                 if (end < last_start)
 910                         this_end = end;
 911                 else
 912                         this_end = last_start - 1;
 913
 914                 prealloc = alloc_extent_state_atomic(prealloc);
 915                 BUG_ON(!prealloc);
 916
 917                 /*
 918                  * Avoid to free 'prealloc' if it can be merged with
 919                  * the later extent.
 920                  */
 921                 err = insert_state(tree, prealloc, start, this_end,
 922                                    &bits);
 923                 if (err)
 924                         extent_io_tree_panic(tree, err);
 925
 926                 cache_state(prealloc, cached_state);
 927                 prealloc = NULL;
 928                 start = this_end + 1;
 929                 goto search_again;
 930         }
 931         /*
 932          * | ---- desired range ---- |
 933          *                        | state |
 934          * We need to split the extent, and set the bit
 935          * on the first half
 936          */
 937         if (state->start <= end && state->end > end) {
 938                 if (state->state & exclusive_bits) {
 939                         *failed_start = start;
 940                         err = -EEXIST;
 941                         goto out;
 942                 }
 943
 944                 prealloc = alloc_extent_state_atomic(prealloc);
 945                 BUG_ON(!prealloc);
 946                 err = split_state(tree, state, prealloc, end + 1);
 947                 if (err)
 948                         extent_io_tree_panic(tree, err);
 949
 950                 set_state_bits(tree, prealloc, &bits);
 951                 cache_state(prealloc, cached_state);
 952                 merge_state(tree, prealloc);
 953                 prealloc = NULL;
 954                 goto out;
 955         }
 956
 957         goto search_again;
 958
 959 out:
 960         spin_unlock(&tree->lock);
 961         if (prealloc)
 962                 free_extent_state(prealloc);
 963
 964         return err;
 965
 966 search_again:
 967         if (start > end)
 968                 goto out;
 969         spin_unlock(&tree->lock);
 970         if (mask & __GFP_WAIT)
 971                 cond_resched();
 972         goto again;
 973 }
 974
 975 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 976                    unsigned long bits, u64 * failed_start,
 977                    struct extent_state **cached_state, gfp_t mask)
 978 {
 979         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
 980                                 cached_state, mask);
 981 }
 982
 983
 984 /**
 985  * convert_extent_bit - convert all bits in a given range from one bit to
 986  *                      another
 987  * @tree:       the io tree to search
 988  * @start:      the start offset in bytes
 989  * @end:        the end offset in bytes (inclusive)
 990  * @bits:       the bits to set in this range
 991  * @clear_bits: the bits to clear in this range
 992  * @cached_state:       state that we're going to cache
 993  * @mask:       the allocation mask
 994  *
 995  * This will go through and set bits for the given range.  If any states exist
 996  * already in this range they are set with the given bit and cleared of the
 997  * clear_bits.  This is only meant to be used by things that are mergeable, ie
 998  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
 999  * boundary bits like LOCK.
1000  */
1001 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1002                        unsigned long bits, unsigned long clear_bits,
1003                        struct extent_state **cached_state, gfp_t mask)
1004 {
1005         struct extent_state *state;
1006         struct extent_state *prealloc = NULL;
1007         struct rb_node *node;
1008         int err = 0;
1009         u64 last_start;
1010         u64 last_end;
1011
1012         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
1013
1014 again:
1015         if (!prealloc && (mask & __GFP_WAIT)) {
1016                 prealloc = alloc_extent_state(mask);
1017                 if (!prealloc)
1018                         return -ENOMEM;
1019         }
1020
1021         spin_lock(&tree->lock);
1022         if (cached_state && *cached_state) {
1023                 state = *cached_state;
1024                 if (state->start <= start && state->end > start &&
1025                     state->tree) {
1026                         node = &state->rb_node;
1027                         goto hit_next;
1028                 }
1029         }
1030
1031         /*
1032          * this search will find all the extents that end after
1033          * our range starts.
1034          */
1035         node = tree_search(tree, start);
1036         if (!node) {
1037                 prealloc = alloc_extent_state_atomic(prealloc);
1038                 if (!prealloc) {
1039                         err = -ENOMEM;
1040                         goto out;
1041                 }
1042                 err = insert_state(tree, prealloc, start, end, &bits);
1043                 prealloc = NULL;
1044                 if (err)
1045                         extent_io_tree_panic(tree, err);
1046                 goto out;
1047         }
1048         state = rb_entry(node, struct extent_state, rb_node);
1049 hit_next:
1050         last_start = state->start;
1051         last_end = state->end;
1052
1053         /*
1054          * | ---- desired range ---- |
1055          * | state |
1056          *
1057          * Just lock what we found and keep going
1058          */
1059         if (state->start == start && state->end <= end) {
1060                 set_state_bits(tree, state, &bits);
1061                 cache_state(state, cached_state);
1062                 state = clear_state_bit(tree, state, &clear_bits, 0);
1063                 if (last_end == (u64)-1)
1064                         goto out;
1065                 start = last_end + 1;
1066                 if (start < end && state && state->start == start &&
1067                     !need_resched())
1068                         goto hit_next;
1069                 goto search_again;
1070         }
1071
1072         /*
1073          *     | ---- desired range ---- |
1074          * | state |
1075          *   or
1076          * | ------------- state -------------- |
1077          *
1078          * We need to split the extent we found, and may flip bits on
1079          * second half.
1080          *
1081          * If the extent we found extends past our
1082          * range, we just split and search again.  It'll get split
1083          * again the next time though.
1084          *
1085          * If the extent we found is inside our range, we set the
1086          * desired bit on it.
1087          */
1088         if (state->start < start) {
1089                 prealloc = alloc_extent_state_atomic(prealloc);
1090                 if (!prealloc) {
1091                         err = -ENOMEM;
1092                         goto out;
1093                 }
1094                 err = split_state(tree, state, prealloc, start);
1095                 if (err)
1096                         extent_io_tree_panic(tree, err);
1097                 prealloc = NULL;
1098                 if (err)
1099                         goto out;
1100                 if (state->end <= end) {
1101                         set_state_bits(tree, state, &bits);
1102                         cache_state(state, cached_state);
1103                         state = clear_state_bit(tree, state, &clear_bits, 0);
1104                         if (last_end == (u64)-1)
1105                                 goto out;
1106                         start = last_end + 1;
1107                         if (start < end && state && state->start == start &&
1108                             !need_resched())
1109                                 goto hit_next;
1110                 }
1111                 goto search_again;
1112         }
1113         /*
1114          * | ---- desired range ---- |
1115          *     | state | or               | state |
1116          *
1117          * There's a hole, we need to insert something in it and
1118          * ignore the extent we found.
1119          */
1120         if (state->start > start) {
1121                 u64 this_end;
1122                 if (end < last_start)
1123                         this_end = end;
1124                 else
1125                         this_end = last_start - 1;
1126
1127                 prealloc = alloc_extent_state_atomic(prealloc);
1128                 if (!prealloc) {
1129                         err = -ENOMEM;
1130                         goto out;
1131                 }
1132
1133                 /*
1134                  * Avoid to free 'prealloc' if it can be merged with
1135                  * the later extent.
1136                  */
1137                 err = insert_state(tree, prealloc, start, this_end,
1138                                    &bits);
1139                 if (err)
1140                         extent_io_tree_panic(tree, err);
1141                 cache_state(prealloc, cached_state);
1142                 prealloc = NULL;
1143                 start = this_end + 1;
1144                 goto search_again;
1145         }
1146         /*
1147          * | ---- desired range ---- |
1148          *                        | state |
1149          * We need to split the extent, and set the bit
1150          * on the first half
1151          */
1152         if (state->start <= end && state->end > end) {
1153                 prealloc = alloc_extent_state_atomic(prealloc);
1154                 if (!prealloc) {
1155                         err = -ENOMEM;
1156                         goto out;
1157                 }
1158
1159                 err = split_state(tree, state, prealloc, end + 1);
1160                 if (err)
1161                         extent_io_tree_panic(tree, err);
1162
1163                 set_state_bits(tree, prealloc, &bits);
1164                 cache_state(prealloc, cached_state);
1165                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1166                 prealloc = NULL;
1167                 goto out;
1168         }
1169
1170         goto search_again;
1171
1172 out:
1173         spin_unlock(&tree->lock);
1174         if (prealloc)
1175                 free_extent_state(prealloc);
1176
1177         return err;
1178
1179 search_again:
1180         if (start > end)
1181                 goto out;
1182         spin_unlock(&tree->lock);
1183         if (mask & __GFP_WAIT)
1184                 cond_resched();
1185         goto again;
1186 }
1187
1188 /* wrappers around set/clear extent bit */
1189 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1190                      gfp_t mask)
1191 {
1192         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1193                               NULL, mask);
1194 }
1195
1196 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1197                     unsigned long bits, gfp_t mask)
1198 {
1199         return set_extent_bit(tree, start, end, bits, NULL,
1200                               NULL, mask);
1201 }
1202
1203 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1204                       unsigned long bits, gfp_t mask)
1205 {
1206         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1207 }
1208
1209 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1210                         struct extent_state **cached_state, gfp_t mask)
1211 {
1212         return set_extent_bit(tree, start, end,
1213                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1214                               NULL, cached_state, mask);
1215 }
1216
1217 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1218                       struct extent_state **cached_state, gfp_t mask)
1219 {
1220         return set_extent_bit(tree, start, end,
1221                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1222                               NULL, cached_state, mask);
1223 }
1224
1225 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1226                        gfp_t mask)
1227 {
1228         return clear_extent_bit(tree, start, end,
1229                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1230                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1231 }
1232
1233 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1234                      gfp_t mask)
1235 {
1236         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1237                               NULL, mask);
1238 }
1239
1240 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1241                         struct extent_state **cached_state, gfp_t mask)
1242 {
1243         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1244                               cached_state, mask);
1245 }
1246
1247 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1248                           struct extent_state **cached_state, gfp_t mask)
1249 {
1250         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1251                                 cached_state, mask);
1252 }
1253
1254 /*
1255  * either insert or lock state struct between start and end use mask to tell
1256  * us if waiting is desired.
1257  */
1258 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1259                      unsigned long bits, struct extent_state **cached_state)
1260 {
1261         int err;
1262         u64 failed_start;
1263         while (1) {
1264                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1265                                        EXTENT_LOCKED, &failed_start,
1266                                        cached_state, GFP_NOFS);
1267                 if (err == -EEXIST) {
1268                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1269                         start = failed_start;
1270                 } else
1271                         break;
1272                 WARN_ON(start > end);
1273         }
1274         return err;
1275 }
1276
1277 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1278 {
1279         return lock_extent_bits(tree, start, end, 0, NULL);
1280 }
1281
1282 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1283 {
1284         int err;
1285         u64 failed_start;
1286
1287         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1288                                &failed_start, NULL, GFP_NOFS);
1289         if (err == -EEXIST) {
1290                 if (failed_start > start)
1291                         clear_extent_bit(tree, start, failed_start - 1,
1292                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1293                 return 0;
1294         }
1295         return 1;
1296 }
1297
1298 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1299                          struct extent_state **cached, gfp_t mask)
1300 {
1301         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1302                                 mask);
1303 }
1304
1305 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1306 {
1307         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1308                                 GFP_NOFS);
1309 }
1310
1311 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1312 {
1313         unsigned long index = start >> PAGE_CACHE_SHIFT;
1314         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1315         struct page *page;
1316
1317         while (index <= end_index) {
1318                 page = find_get_page(inode->i_mapping, index);
1319                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1320                 clear_page_dirty_for_io(page);
1321                 page_cache_release(page);
1322                 index++;
1323         }
1324         return 0;
1325 }
1326
1327 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1328 {
1329         unsigned long index = start >> PAGE_CACHE_SHIFT;
1330         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1331         struct page *page;
1332
1333         while (index <= end_index) {
1334                 page = find_get_page(inode->i_mapping, index);
1335                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1336                 account_page_redirty(page);
1337                 __set_page_dirty_nobuffers(page);
1338                 page_cache_release(page);
1339                 index++;
1340         }
1341         return 0;
1342 }
1343
1344 /*
1345  * helper function to set both pages and extents in the tree writeback
1346  */
1347 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1348 {
1349         unsigned long index = start >> PAGE_CACHE_SHIFT;
1350         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1351         struct page *page;
1352
1353         while (index <= end_index) {
1354                 page = find_get_page(tree->mapping, index);
1355                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1356                 set_page_writeback(page);
1357                 page_cache_release(page);
1358                 index++;
1359         }
1360         return 0;
1361 }
1362
1363 /* find the first state struct with 'bits' set after 'start', and
1364  * return it.  tree->lock must be held.  NULL will returned if
1365  * nothing was found after 'start'
1366  */
1367 static struct extent_state *
1368 find_first_extent_bit_state(struct extent_io_tree *tree,
1369                             u64 start, unsigned long bits)
1370 {
1371         struct rb_node *node;
1372         struct extent_state *state;
1373
1374         /*
1375          * this search will find all the extents that end after
1376          * our range starts.
1377          */
1378         node = tree_search(tree, start);
1379         if (!node)
1380                 goto out;
1381
1382         while (1) {
1383                 state = rb_entry(node, struct extent_state, rb_node);
1384                 if (state->end >= start && (state->state & bits))
1385                         return state;
1386
1387                 node = rb_next(node);
1388                 if (!node)
1389                         break;
1390         }
1391 out:
1392         return NULL;
1393 }
1394
1395 /*
1396  * find the first offset in the io tree with 'bits' set. zero is
1397  * returned if we find something, and *start_ret and *end_ret are
1398  * set to reflect the state struct that was found.
1399  *
1400  * If nothing was found, 1 is returned. If found something, return 0.
1401  */
1402 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1403                           u64 *start_ret, u64 *end_ret, unsigned long bits,
1404                           struct extent_state **cached_state)
1405 {
1406         struct extent_state *state;
1407         struct rb_node *n;
1408         int ret = 1;
1409
1410         spin_lock(&tree->lock);
1411         if (cached_state && *cached_state) {
1412                 state = *cached_state;
1413                 if (state->end == start - 1 && state->tree) {
1414                         n = rb_next(&state->rb_node);
1415                         while (n) {
1416                                 state = rb_entry(n, struct extent_state,
1417                                                  rb_node);
1418                                 if (state->state & bits)
1419                                         goto got_it;
1420                                 n = rb_next(n);
1421                         }
1422                         free_extent_state(*cached_state);
1423                         *cached_state = NULL;
1424                         goto out;
1425                 }
1426                 free_extent_state(*cached_state);
1427                 *cached_state = NULL;
1428         }
1429
1430         state = find_first_extent_bit_state(tree, start, bits);
1431 got_it:
1432         if (state) {
1433                 cache_state(state, cached_state);
1434                 *start_ret = state->start;
1435                 *end_ret = state->end;
1436                 ret = 0;
1437         }
1438 out:
1439         spin_unlock(&tree->lock);
1440         return ret;
1441 }
1442
1443 /*
1444  * find a contiguous range of bytes in the file marked as delalloc, not
1445  * more than 'max_bytes'.  start and end are used to return the range,
1446  *
1447  * 1 is returned if we find something, 0 if nothing was in the tree
1448  */
1449 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1450                                         u64 *start, u64 *end, u64 max_bytes,
1451                                         struct extent_state **cached_state)
1452 {
1453         struct rb_node *node;
1454         struct extent_state *state;
1455         u64 cur_start = *start;
1456         u64 found = 0;
1457         u64 total_bytes = 0;
1458
1459         spin_lock(&tree->lock);
1460
1461         /*
1462          * this search will find all the extents that end after
1463          * our range starts.
1464          */
1465         node = tree_search(tree, cur_start);
1466         if (!node) {
1467                 if (!found)
1468                         *end = (u64)-1;
1469                 goto out;
1470         }
1471
1472         while (1) {
1473                 state = rb_entry(node, struct extent_state, rb_node);
1474                 if (found && (state->start != cur_start ||
1475                               (state->state & EXTENT_BOUNDARY))) {
1476                         goto out;
1477                 }
1478                 if (!(state->state & EXTENT_DELALLOC)) {
1479                         if (!found)
1480                                 *end = state->end;
1481                         goto out;
1482                 }
1483                 if (!found) {
1484                         *start = state->start;
1485                         *cached_state = state;
1486                         atomic_inc(&state->refs);
1487                 }
1488                 found++;
1489                 *end = state->end;
1490                 cur_start = state->end + 1;
1491                 node = rb_next(node);
1492                 total_bytes += state->end - state->start + 1;
1493                 if (total_bytes >= max_bytes)
1494                         break;
1495                 if (!node)
1496                         break;
1497         }
1498 out:
1499         spin_unlock(&tree->lock);
1500         return found;
1501 }
1502
1503 static noinline void __unlock_for_delalloc(struct inode *inode,
1504                                            struct page *locked_page,
1505                                            u64 start, u64 end)
1506 {
1507         int ret;
1508         struct page *pages[16];
1509         unsigned long index = start >> PAGE_CACHE_SHIFT;
1510         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1511         unsigned long nr_pages = end_index - index + 1;
1512         int i;
1513
1514         if (index == locked_page->index && end_index == index)
1515                 return;
1516
1517         while (nr_pages > 0) {
1518                 ret = find_get_pages_contig(inode->i_mapping, index,
1519                                      min_t(unsigned long, nr_pages,
1520                                      ARRAY_SIZE(pages)), pages);
1521                 for (i = 0; i < ret; i++) {
1522                         if (pages[i] != locked_page)
1523                                 unlock_page(pages[i]);
1524                         page_cache_release(pages[i]);
1525                 }
1526                 nr_pages -= ret;
1527                 index += ret;
1528                 cond_resched();
1529         }
1530 }
1531
1532 static noinline int lock_delalloc_pages(struct inode *inode,
1533                                         struct page *locked_page,
1534                                         u64 delalloc_start,
1535                                         u64 delalloc_end)
1536 {
1537         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1538         unsigned long start_index = index;
1539         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1540         unsigned long pages_locked = 0;
1541         struct page *pages[16];
1542         unsigned long nrpages;
1543         int ret;
1544         int i;
1545
1546         /* the caller is responsible for locking the start index */
1547         if (index == locked_page->index && index == end_index)
1548                 return 0;
1549
1550         /* skip the page at the start index */
1551         nrpages = end_index - index + 1;
1552         while (nrpages > 0) {
1553                 ret = find_get_pages_contig(inode->i_mapping, index,
1554                                      min_t(unsigned long,
1555                                      nrpages, ARRAY_SIZE(pages)), pages);
1556                 if (ret == 0) {
1557                         ret = -EAGAIN;
1558                         goto done;
1559                 }
1560                 /* now we have an array of pages, lock them all */
1561                 for (i = 0; i < ret; i++) {
1562                         /*
1563                          * the caller is taking responsibility for
1564                          * locked_page
1565                          */
1566                         if (pages[i] != locked_page) {
1567                                 lock_page(pages[i]);
1568                                 if (!PageDirty(pages[i]) ||
1569                                     pages[i]->mapping != inode->i_mapping) {
1570                                         ret = -EAGAIN;
1571                                         unlock_page(pages[i]);
1572                                         page_cache_release(pages[i]);
1573                                         goto done;
1574                                 }
1575                         }
1576                         page_cache_release(pages[i]);
1577                         pages_locked++;
1578                 }
1579                 nrpages -= ret;
1580                 index += ret;
1581                 cond_resched();
1582         }
1583         ret = 0;
1584 done:
1585         if (ret && pages_locked) {
1586                 __unlock_for_delalloc(inode, locked_page,
1587                               delalloc_start,
1588                               ((u64)(start_index + pages_locked - 1)) <<
1589                               PAGE_CACHE_SHIFT);
1590         }
1591         return ret;
1592 }
1593
1594 /*
1595  * find a contiguous range of bytes in the file marked as delalloc, not
1596  * more than 'max_bytes'.  start and end are used to return the range,
1597  *
1598  * 1 is returned if we find something, 0 if nothing was in the tree
1599  */
1600 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1601                                     struct extent_io_tree *tree,
1602                                     struct page *locked_page, u64 *start,
1603                                     u64 *end, u64 max_bytes)
1604 {
1605         u64 delalloc_start;
1606         u64 delalloc_end;
1607         u64 found;
1608         struct extent_state *cached_state = NULL;
1609         int ret;
1610         int loops = 0;
1611
1612 again:
1613         /* step one, find a bunch of delalloc bytes starting at start */
1614         delalloc_start = *start;
1615         delalloc_end = 0;
1616         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1617                                     max_bytes, &cached_state);
1618         if (!found || delalloc_end <= *start) {
1619                 *start = delalloc_start;
1620                 *end = delalloc_end;
1621                 free_extent_state(cached_state);
1622                 return 0;
1623         }
1624
1625         /*
1626          * start comes from the offset of locked_page.  We have to lock
1627          * pages in order, so we can't process delalloc bytes before
1628          * locked_page
1629          */
1630         if (delalloc_start < *start)
1631                 delalloc_start = *start;
1632
1633         /*
1634          * make sure to limit the number of pages we try to lock down
1635          */
1636         if (delalloc_end + 1 - delalloc_start > max_bytes)
1637                 delalloc_end = delalloc_start + max_bytes - 1;
1638
1639         /* step two, lock all the pages after the page that has start */
1640         ret = lock_delalloc_pages(inode, locked_page,
1641                                   delalloc_start, delalloc_end);
1642         if (ret == -EAGAIN) {
1643                 /* some of the pages are gone, lets avoid looping by
1644                  * shortening the size of the delalloc range we're searching
1645                  */
1646                 free_extent_state(cached_state);
1647                 if (!loops) {
1648                         max_bytes = PAGE_CACHE_SIZE;
1649                         loops = 1;
1650                         goto again;
1651                 } else {
1652                         found = 0;
1653                         goto out_failed;
1654                 }
1655         }
1656         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1657
1658         /* step three, lock the state bits for the whole range */
1659         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1660
1661         /* then test to make sure it is all still delalloc */
1662         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1663                              EXTENT_DELALLOC, 1, cached_state);
1664         if (!ret) {
1665                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1666                                      &cached_state, GFP_NOFS);
1667                 __unlock_for_delalloc(inode, locked_page,
1668                               delalloc_start, delalloc_end);
1669                 cond_resched();
1670                 goto again;
1671         }
1672         free_extent_state(cached_state);
1673         *start = delalloc_start;
1674         *end = delalloc_end;
1675 out_failed:
1676         return found;
1677 }
1678
1679 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1680                                  struct page *locked_page,
1681                                  unsigned long clear_bits,
1682                                  unsigned long page_ops)
1683 {
1684         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1685         int ret;
1686         struct page *pages[16];
1687         unsigned long index = start >> PAGE_CACHE_SHIFT;
1688         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1689         unsigned long nr_pages = end_index - index + 1;
1690         int i;
1691
1692         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1693         if (page_ops == 0)
1694                 return 0;
1695
1696         while (nr_pages > 0) {
1697                 ret = find_get_pages_contig(inode->i_mapping, index,
1698                                      min_t(unsigned long,
1699                                      nr_pages, ARRAY_SIZE(pages)), pages);
1700                 for (i = 0; i < ret; i++) {
1701
1702                         if (page_ops & PAGE_SET_PRIVATE2)
1703                                 SetPagePrivate2(pages[i]);
1704
1705                         if (pages[i] == locked_page) {
1706                                 page_cache_release(pages[i]);
1707                                 continue;
1708                         }
1709                         if (page_ops & PAGE_CLEAR_DIRTY)
1710                                 clear_page_dirty_for_io(pages[i]);
1711                         if (page_ops & PAGE_SET_WRITEBACK)
1712                                 set_page_writeback(pages[i]);
1713                         if (page_ops & PAGE_END_WRITEBACK)
1714                                 end_page_writeback(pages[i]);
1715                         if (page_ops & PAGE_UNLOCK)
1716                                 unlock_page(pages[i]);
1717                         page_cache_release(pages[i]);
1718                 }
1719                 nr_pages -= ret;
1720                 index += ret;
1721                 cond_resched();
1722         }
1723         return 0;
1724 }
1725
1726 /*
1727  * count the number of bytes in the tree that have a given bit(s)
1728  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1729  * cached.  The total number found is returned.
1730  */
1731 u64 count_range_bits(struct extent_io_tree *tree,
1732                      u64 *start, u64 search_end, u64 max_bytes,
1733                      unsigned long bits, int contig)
1734 {
1735         struct rb_node *node;
1736         struct extent_state *state;
1737         u64 cur_start = *start;
1738         u64 total_bytes = 0;
1739         u64 last = 0;
1740         int found = 0;
1741
1742         if (WARN_ON(search_end <= cur_start))
1743                 return 0;
1744
1745         spin_lock(&tree->lock);
1746         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1747                 total_bytes = tree->dirty_bytes;
1748                 goto out;
1749         }
1750         /*
1751          * this search will find all the extents that end after
1752          * our range starts.
1753          */
1754         node = tree_search(tree, cur_start);
1755         if (!node)
1756                 goto out;
1757
1758         while (1) {
1759                 state = rb_entry(node, struct extent_state, rb_node);
1760                 if (state->start > search_end)
1761                         break;
1762                 if (contig && found && state->start > last + 1)
1763                         break;
1764                 if (state->end >= cur_start && (state->state & bits) == bits) {
1765                         total_bytes += min(search_end, state->end) + 1 -
1766                                        max(cur_start, state->start);
1767                         if (total_bytes >= max_bytes)
1768                                 break;
1769                         if (!found) {
1770                                 *start = max(cur_start, state->start);
1771                                 found = 1;
1772                         }
1773                         last = state->end;
1774                 } else if (contig && found) {
1775                         break;
1776                 }
1777                 node = rb_next(node);
1778                 if (!node)
1779                         break;
1780         }
1781 out:
1782         spin_unlock(&tree->lock);
1783         return total_bytes;
1784 }
1785
1786 /*
1787  * set the private field for a given byte offset in the tree.  If there isn't
1788  * an extent_state there already, this does nothing.
1789  */
1790 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1791 {
1792         struct rb_node *node;
1793         struct extent_state *state;
1794         int ret = 0;
1795
1796         spin_lock(&tree->lock);
1797         /*
1798          * this search will find all the extents that end after
1799          * our range starts.
1800          */
1801         node = tree_search(tree, start);
1802         if (!node) {
1803                 ret = -ENOENT;
1804                 goto out;
1805         }
1806         state = rb_entry(node, struct extent_state, rb_node);
1807         if (state->start != start) {
1808                 ret = -ENOENT;
1809                 goto out;
1810         }
1811         state->private = private;
1812 out:
1813         spin_unlock(&tree->lock);
1814         return ret;
1815 }
1816
1817 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1818 {
1819         struct rb_node *node;
1820         struct extent_state *state;
1821         int ret = 0;
1822
1823         spin_lock(&tree->lock);
1824         /*
1825          * this search will find all the extents that end after
1826          * our range starts.
1827          */
1828         node = tree_search(tree, start);
1829         if (!node) {
1830                 ret = -ENOENT;
1831                 goto out;
1832         }
1833         state = rb_entry(node, struct extent_state, rb_node);
1834         if (state->start != start) {
1835                 ret = -ENOENT;
1836                 goto out;
1837         }
1838         *private = state->private;
1839 out:
1840         spin_unlock(&tree->lock);
1841         return ret;
1842 }
1843
1844 /*
1845  * searches a range in the state tree for a given mask.
1846  * If 'filled' == 1, this returns 1 only if every extent in the tree
1847  * has the bits set.  Otherwise, 1 is returned if any bit in the
1848  * range is found set.
1849  */
1850 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1851                    unsigned long bits, int filled, struct extent_state *cached)
1852 {
1853         struct extent_state *state = NULL;
1854         struct rb_node *node;
1855         int bitset = 0;
1856
1857         spin_lock(&tree->lock);
1858         if (cached && cached->tree && cached->start <= start &&
1859             cached->end > start)
1860                 node = &cached->rb_node;
1861         else
1862                 node = tree_search(tree, start);
1863         while (node && start <= end) {
1864                 state = rb_entry(node, struct extent_state, rb_node);
1865
1866                 if (filled && state->start > start) {
1867                         bitset = 0;
1868                         break;
1869                 }
1870
1871                 if (state->start > end)
1872                         break;
1873
1874                 if (state->state & bits) {
1875                         bitset = 1;
1876                         if (!filled)
1877                                 break;
1878                 } else if (filled) {
1879                         bitset = 0;
1880                         break;
1881                 }
1882
1883                 if (state->end == (u64)-1)
1884                         break;
1885
1886                 start = state->end + 1;
1887                 if (start > end)
1888                         break;
1889                 node = rb_next(node);
1890                 if (!node) {
1891                         if (filled)
1892                                 bitset = 0;
1893                         break;
1894                 }
1895         }
1896         spin_unlock(&tree->lock);
1897         return bitset;
1898 }
1899
1900 /*
1901  * helper function to set a given page up to date if all the
1902  * extents in the tree for that page are up to date
1903  */
1904 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1905 {
1906         u64 start = page_offset(page);
1907         u64 end = start + PAGE_CACHE_SIZE - 1;
1908         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1909                 SetPageUptodate(page);
1910 }
1911
1912 /*
1913  * When IO fails, either with EIO or csum verification fails, we
1914  * try other mirrors that might have a good copy of the data.  This
1915  * io_failure_record is used to record state as we go through all the
1916  * mirrors.  If another mirror has good data, the page is set up to date
1917  * and things continue.  If a good mirror can't be found, the original
1918  * bio end_io callback is called to indicate things have failed.
1919  */
1920 struct io_failure_record {
1921         struct page *page;
1922         u64 start;
1923         u64 len;
1924         u64 logical;
1925         unsigned long bio_flags;
1926         int this_mirror;
1927         int failed_mirror;
1928         int in_validation;
1929 };
1930
1931 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1932                                 int did_repair)
1933 {
1934         int ret;
1935         int err = 0;
1936         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1937
1938         set_state_private(failure_tree, rec->start, 0);
1939         ret = clear_extent_bits(failure_tree, rec->start,
1940                                 rec->start + rec->len - 1,
1941                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1942         if (ret)
1943                 err = ret;
1944
1945         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1946                                 rec->start + rec->len - 1,
1947                                 EXTENT_DAMAGED, GFP_NOFS);
1948         if (ret && !err)
1949                 err = ret;
1950
1951         kfree(rec);
1952         return err;
1953 }
1954
1955 static void repair_io_failure_callback(struct bio *bio, int err)
1956 {
1957         complete(bio->bi_private);
1958 }
1959
1960 /*
1961  * this bypasses the standard btrfs submit functions deliberately, as
1962  * the standard behavior is to write all copies in a raid setup. here we only
1963  * want to write the one bad copy. so we do the mapping for ourselves and issue
1964  * submit_bio directly.
1965  * to avoid any synchronization issues, wait for the data after writing, which
1966  * actually prevents the read that triggered the error from finishing.
1967  * currently, there can be no more than two copies of every data bit. thus,
1968  * exactly one rewrite is required.
1969  */
1970 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 start,
1971                         u64 length, u64 logical, struct page *page,
1972                         int mirror_num)
1973 {
1974         struct bio *bio;
1975         struct btrfs_device *dev;
1976         DECLARE_COMPLETION_ONSTACK(compl);
1977         u64 map_length = 0;
1978         u64 sector;
1979         struct btrfs_bio *bbio = NULL;
1980         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
1981         int ret;
1982
1983         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
1984         BUG_ON(!mirror_num);
1985
1986         /* we can't repair anything in raid56 yet */
1987         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
1988                 return 0;
1989
1990         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1991         if (!bio)
1992                 return -EIO;
1993         bio->bi_private = &compl;
1994         bio->bi_end_io = repair_io_failure_callback;
1995         bio->bi_size = 0;
1996         map_length = length;
1997
1998         ret = btrfs_map_block(fs_info, WRITE, logical,
1999                               &map_length, &bbio, mirror_num);
2000         if (ret) {
2001                 bio_put(bio);
2002                 return -EIO;
2003         }
2004         BUG_ON(mirror_num != bbio->mirror_num);
2005         sector = bbio->stripes[mirror_num-1].physical >> 9;
2006         bio->bi_sector = sector;
2007         dev = bbio->stripes[mirror_num-1].dev;
2008         kfree(bbio);
2009         if (!dev || !dev->bdev || !dev->writeable) {
2010                 bio_put(bio);
2011                 return -EIO;
2012         }
2013         bio->bi_bdev = dev->bdev;
2014         bio_add_page(bio, page, length, start - page_offset(page));
2015         btrfsic_submit_bio(WRITE_SYNC, bio);
2016         wait_for_completion(&compl);
2017
2018         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2019                 /* try to remap that extent elsewhere? */
2020                 bio_put(bio);
2021                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2022                 return -EIO;
2023         }
2024
2025         printk_ratelimited_in_rcu(KERN_INFO "btrfs read error corrected: ino %lu off %llu "
2026                       "(dev %s sector %llu)\n", page->mapping->host->i_ino,
2027                       start, rcu_str_deref(dev->name), sector);
2028
2029         bio_put(bio);
2030         return 0;
2031 }
2032
2033 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2034                          int mirror_num)
2035 {
2036         u64 start = eb->start;
2037         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2038         int ret = 0;
2039
2040         if (root->fs_info->sb->s_flags & MS_RDONLY)
2041                 return -EROFS;
2042
2043         for (i = 0; i < num_pages; i++) {
2044                 struct page *p = extent_buffer_page(eb, i);
2045                 ret = repair_io_failure(root->fs_info, start, PAGE_CACHE_SIZE,
2046                                         start, p, mirror_num);
2047                 if (ret)
2048                         break;
2049                 start += PAGE_CACHE_SIZE;
2050         }
2051
2052         return ret;
2053 }
2054
2055 /*
2056  * each time an IO finishes, we do a fast check in the IO failure tree
2057  * to see if we need to process or clean up an io_failure_record
2058  */
2059 static int clean_io_failure(u64 start, struct page *page)
2060 {
2061         u64 private;
2062         u64 private_failure;
2063         struct io_failure_record *failrec;
2064         struct inode *inode = page->mapping->host;
2065         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2066         struct extent_state *state;
2067         int num_copies;
2068         int did_repair = 0;
2069         int ret;
2070
2071         private = 0;
2072         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2073                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2074         if (!ret)
2075                 return 0;
2076
2077         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2078                                 &private_failure);
2079         if (ret)
2080                 return 0;
2081
2082         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2083         BUG_ON(!failrec->this_mirror);
2084
2085         if (failrec->in_validation) {
2086                 /* there was no real error, just free the record */
2087                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2088                          failrec->start);
2089                 did_repair = 1;
2090                 goto out;
2091         }
2092         if (fs_info->sb->s_flags & MS_RDONLY)
2093                 goto out;
2094
2095         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2096         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2097                                             failrec->start,
2098                                             EXTENT_LOCKED);
2099         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2100
2101         if (state && state->start <= failrec->start &&
2102             state->end >= failrec->start + failrec->len - 1) {
2103                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2104                                               failrec->len);
2105                 if (num_copies > 1)  {
2106                         ret = repair_io_failure(fs_info, start, failrec->len,
2107                                                 failrec->logical, page,
2108                                                 failrec->failed_mirror);
2109                         did_repair = !ret;
2110                 }
2111                 ret = 0;
2112         }
2113
2114 out:
2115         if (!ret)
2116                 ret = free_io_failure(inode, failrec, did_repair);
2117
2118         return ret;
2119 }
2120
2121 /*
2122  * this is a generic handler for readpage errors (default
2123  * readpage_io_failed_hook). if other copies exist, read those and write back
2124  * good data to the failed position. does not investigate in remapping the
2125  * failed extent elsewhere, hoping the device will be smart enough to do this as
2126  * needed
2127  */
2128
2129 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2130                               struct page *page, u64 start, u64 end,
2131                               int failed_mirror)
2132 {
2133         struct io_failure_record *failrec = NULL;
2134         u64 private;
2135         struct extent_map *em;
2136         struct inode *inode = page->mapping->host;
2137         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2138         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2139         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2140         struct bio *bio;
2141         struct btrfs_io_bio *btrfs_failed_bio;
2142         struct btrfs_io_bio *btrfs_bio;
2143         int num_copies;
2144         int ret;
2145         int read_mode;
2146         u64 logical;
2147
2148         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2149
2150         ret = get_state_private(failure_tree, start, &private);
2151         if (ret) {
2152                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2153                 if (!failrec)
2154                         return -ENOMEM;
2155                 failrec->start = start;
2156                 failrec->len = end - start + 1;
2157                 failrec->this_mirror = 0;
2158                 failrec->bio_flags = 0;
2159                 failrec->in_validation = 0;
2160
2161                 read_lock(&em_tree->lock);
2162                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2163                 if (!em) {
2164                         read_unlock(&em_tree->lock);
2165                         kfree(failrec);
2166                         return -EIO;
2167                 }
2168
2169                 if (em->start > start || em->start + em->len < start) {
2170                         free_extent_map(em);
2171                         em = NULL;
2172                 }
2173                 read_unlock(&em_tree->lock);
2174
2175                 if (!em) {
2176                         kfree(failrec);
2177                         return -EIO;
2178                 }
2179                 logical = start - em->start;
2180                 logical = em->block_start + logical;
2181                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2182                         logical = em->block_start;
2183                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2184                         extent_set_compress_type(&failrec->bio_flags,
2185                                                  em->compress_type);
2186                 }
2187                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2188                          "len=%llu\n", logical, start, failrec->len);
2189                 failrec->logical = logical;
2190                 free_extent_map(em);
2191
2192                 /* set the bits in the private failure tree */
2193                 ret = set_extent_bits(failure_tree, start, end,
2194                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2195                 if (ret >= 0)
2196                         ret = set_state_private(failure_tree, start,
2197                                                 (u64)(unsigned long)failrec);
2198                 /* set the bits in the inode's tree */
2199                 if (ret >= 0)
2200                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2201                                                 GFP_NOFS);
2202                 if (ret < 0) {
2203                         kfree(failrec);
2204                         return ret;
2205                 }
2206         } else {
2207                 failrec = (struct io_failure_record *)(unsigned long)private;
2208                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2209                          "start=%llu, len=%llu, validation=%d\n",
2210                          failrec->logical, failrec->start, failrec->len,
2211                          failrec->in_validation);
2212                 /*
2213                  * when data can be on disk more than twice, add to failrec here
2214                  * (e.g. with a list for failed_mirror) to make
2215                  * clean_io_failure() clean all those errors at once.
2216                  */
2217         }
2218         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2219                                       failrec->logical, failrec->len);
2220         if (num_copies == 1) {
2221                 /*
2222                  * we only have a single copy of the data, so don't bother with
2223                  * all the retry and error correction code that follows. no
2224                  * matter what the error is, it is very likely to persist.
2225                  */
2226                 pr_debug("bio_readpage_error: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2227                          num_copies, failrec->this_mirror, failed_mirror);
2228                 free_io_failure(inode, failrec, 0);
2229                 return -EIO;
2230         }
2231
2232         /*
2233          * there are two premises:
2234          *      a) deliver good data to the caller
2235          *      b) correct the bad sectors on disk
2236          */
2237         if (failed_bio->bi_vcnt > 1) {
2238                 /*
2239                  * to fulfill b), we need to know the exact failing sectors, as
2240                  * we don't want to rewrite any more than the failed ones. thus,
2241                  * we need separate read requests for the failed bio
2242                  *
2243                  * if the following BUG_ON triggers, our validation request got
2244                  * merged. we need separate requests for our algorithm to work.
2245                  */
2246                 BUG_ON(failrec->in_validation);
2247                 failrec->in_validation = 1;
2248                 failrec->this_mirror = failed_mirror;
2249                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2250         } else {
2251                 /*
2252                  * we're ready to fulfill a) and b) alongside. get a good copy
2253                  * of the failed sector and if we succeed, we have setup
2254                  * everything for repair_io_failure to do the rest for us.
2255                  */
2256                 if (failrec->in_validation) {
2257                         BUG_ON(failrec->this_mirror != failed_mirror);
2258                         failrec->in_validation = 0;
2259                         failrec->this_mirror = 0;
2260                 }
2261                 failrec->failed_mirror = failed_mirror;
2262                 failrec->this_mirror++;
2263                 if (failrec->this_mirror == failed_mirror)
2264                         failrec->this_mirror++;
2265                 read_mode = READ_SYNC;
2266         }
2267
2268         if (failrec->this_mirror > num_copies) {
2269                 pr_debug("bio_readpage_error: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2270                          num_copies, failrec->this_mirror, failed_mirror);
2271                 free_io_failure(inode, failrec, 0);
2272                 return -EIO;
2273         }
2274
2275         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2276         if (!bio) {
2277                 free_io_failure(inode, failrec, 0);
2278                 return -EIO;
2279         }
2280         bio->bi_end_io = failed_bio->bi_end_io;
2281         bio->bi_sector = failrec->logical >> 9;
2282         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2283         bio->bi_size = 0;
2284
2285         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2286         if (btrfs_failed_bio->csum) {
2287                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2288                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2289
2290                 btrfs_bio = btrfs_io_bio(bio);
2291                 btrfs_bio->csum = btrfs_bio->csum_inline;
2292                 phy_offset >>= inode->i_sb->s_blocksize_bits;
2293                 phy_offset *= csum_size;
2294                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + phy_offset,
2295                        csum_size);
2296         }
2297
2298         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2299
2300         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2301                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2302                  failrec->this_mirror, num_copies, failrec->in_validation);
2303
2304         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2305                                          failrec->this_mirror,
2306                                          failrec->bio_flags, 0);
2307         return ret;
2308 }
2309
2310 /* lots and lots of room for performance fixes in the end_bio funcs */
2311
2312 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2313 {
2314         int uptodate = (err == 0);
2315         struct extent_io_tree *tree;
2316         int ret;
2317
2318         tree = &BTRFS_I(page->mapping->host)->io_tree;
2319
2320         if (tree->ops && tree->ops->writepage_end_io_hook) {
2321                 ret = tree->ops->writepage_end_io_hook(page, start,
2322                                                end, NULL, uptodate);
2323                 if (ret)
2324                         uptodate = 0;
2325         }
2326
2327         if (!uptodate) {
2328                 ClearPageUptodate(page);
2329                 SetPageError(page);
2330         }
2331         return 0;
2332 }
2333
2334 /*
2335  * after a writepage IO is done, we need to:
2336  * clear the uptodate bits on error
2337  * clear the writeback bits in the extent tree for this IO
2338  * end_page_writeback if the page has no more pending IO
2339  *
2340  * Scheduling is not allowed, so the extent state tree is expected
2341  * to have one and only one object corresponding to this IO.
2342  */
2343 static void end_bio_extent_writepage(struct bio *bio, int err)
2344 {
2345         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2346         struct extent_io_tree *tree;
2347         u64 start;
2348         u64 end;
2349
2350         do {
2351                 struct page *page = bvec->bv_page;
2352                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2353
2354                 /* We always issue full-page reads, but if some block
2355                  * in a page fails to read, blk_update_request() will
2356                  * advance bv_offset and adjust bv_len to compensate.
2357                  * Print a warning for nonzero offsets, and an error
2358                  * if they don't add up to a full page.  */
2359                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE)
2360                         printk("%s page write in btrfs with offset %u and length %u\n",
2361                                bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE
2362                                ? KERN_ERR "partial" : KERN_INFO "incomplete",
2363                                bvec->bv_offset, bvec->bv_len);
2364
2365                 start = page_offset(page);
2366                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2367
2368                 if (--bvec >= bio->bi_io_vec)
2369                         prefetchw(&bvec->bv_page->flags);
2370
2371                 if (end_extent_writepage(page, err, start, end))
2372                         continue;
2373
2374                 end_page_writeback(page);
2375         } while (bvec >= bio->bi_io_vec);
2376
2377         bio_put(bio);
2378 }
2379
2380 static void
2381 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2382                               int uptodate)
2383 {
2384         struct extent_state *cached = NULL;
2385         u64 end = start + len - 1;
2386
2387         if (uptodate && tree->track_uptodate)
2388                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2389         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2390 }
2391
2392 /*
2393  * after a readpage IO is done, we need to:
2394  * clear the uptodate bits on error
2395  * set the uptodate bits if things worked
2396  * set the page up to date if all extents in the tree are uptodate
2397  * clear the lock bit in the extent tree
2398  * unlock the page if there are no other extents locked for it
2399  *
2400  * Scheduling is not allowed, so the extent state tree is expected
2401  * to have one and only one object corresponding to this IO.
2402  */
2403 static void end_bio_extent_readpage(struct bio *bio, int err)
2404 {
2405         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2406         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2407         struct bio_vec *bvec = bio->bi_io_vec;
2408         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2409         struct extent_io_tree *tree;
2410         u64 offset = 0;
2411         u64 start;
2412         u64 end;
2413         u64 len;
2414         u64 extent_start = 0;
2415         u64 extent_len = 0;
2416         int mirror;
2417         int ret;
2418
2419         if (err)
2420                 uptodate = 0;
2421
2422         do {
2423                 struct page *page = bvec->bv_page;
2424                 struct inode *inode = page->mapping->host;
2425
2426                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2427                          "mirror=%lu\n", (u64)bio->bi_sector, err,
2428                          io_bio->mirror_num);
2429                 tree = &BTRFS_I(inode)->io_tree;
2430
2431                 /* We always issue full-page reads, but if some block
2432                  * in a page fails to read, blk_update_request() will
2433                  * advance bv_offset and adjust bv_len to compensate.
2434                  * Print a warning for nonzero offsets, and an error
2435                  * if they don't add up to a full page.  */
2436                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE)
2437                         printk("%s page read in btrfs with offset %u and length %u\n",
2438                                bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE
2439                                ? KERN_ERR "partial" : KERN_INFO "incomplete",
2440                                bvec->bv_offset, bvec->bv_len);
2441
2442                 start = page_offset(page);
2443                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2444                 len = bvec->bv_len;
2445
2446                 if (++bvec <= bvec_end)
2447                         prefetchw(&bvec->bv_page->flags);
2448
2449                 mirror = io_bio->mirror_num;
2450                 if (likely(uptodate && tree->ops &&
2451                            tree->ops->readpage_end_io_hook)) {
2452                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2453                                                               page, start, end,
2454                                                               mirror);
2455                         if (ret)
2456                                 uptodate = 0;
2457                         else
2458                                 clean_io_failure(start, page);
2459                 }
2460
2461                 if (likely(uptodate))
2462                         goto readpage_ok;
2463
2464                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2465                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2466                         if (!ret && !err &&
2467                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2468                                 uptodate = 1;
2469                 } else {
2470                         /*
2471                          * The generic bio_readpage_error handles errors the
2472                          * following way: If possible, new read requests are
2473                          * created and submitted and will end up in
2474                          * end_bio_extent_readpage as well (if we're lucky, not
2475                          * in the !uptodate case). In that case it returns 0 and
2476                          * we just go on with the next page in our bio. If it
2477                          * can't handle the error it will return -EIO and we
2478                          * remain responsible for that page.
2479                          */
2480                         ret = bio_readpage_error(bio, offset, page, start, end,
2481                                                  mirror);
2482                         if (ret == 0) {
2483                                 uptodate =
2484                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2485                                 if (err)
2486                                         uptodate = 0;
2487                                 continue;
2488                         }
2489                 }
2490 readpage_ok:
2491                 if (likely(uptodate)) {
2492                         loff_t i_size = i_size_read(inode);
2493                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2494                         unsigned offset;
2495
2496                         /* Zero out the end if this page straddles i_size */
2497                         offset = i_size & (PAGE_CACHE_SIZE-1);
2498                         if (page->index == end_index && offset)
2499                                 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2500                         SetPageUptodate(page);
2501                 } else {
2502                         ClearPageUptodate(page);
2503                         SetPageError(page);
2504                 }
2505                 unlock_page(page);
2506                 offset += len;
2507
2508                 if (unlikely(!uptodate)) {
2509                         if (extent_len) {
2510                                 endio_readpage_release_extent(tree,
2511                                                               extent_start,
2512                                                               extent_len, 1);
2513                                 extent_start = 0;
2514                                 extent_len = 0;
2515                         }
2516                         endio_readpage_release_extent(tree, start,
2517                                                       end - start + 1, 0);
2518                 } else if (!extent_len) {
2519                         extent_start = start;
2520                         extent_len = end + 1 - start;
2521                 } else if (extent_start + extent_len == start) {
2522                         extent_len += end + 1 - start;
2523                 } else {
2524                         endio_readpage_release_extent(tree, extent_start,
2525                                                       extent_len, uptodate);
2526                         extent_start = start;
2527                         extent_len = end + 1 - start;
2528                 }
2529         } while (bvec <= bvec_end);
2530
2531         if (extent_len)
2532                 endio_readpage_release_extent(tree, extent_start, extent_len,
2533                                               uptodate);
2534         if (io_bio->end_io)
2535                 io_bio->end_io(io_bio, err);
2536         bio_put(bio);
2537 }
2538
2539 /*
2540  * this allocates from the btrfs_bioset.  We're returning a bio right now
2541  * but you can call btrfs_io_bio for the appropriate container_of magic
2542  */
2543 struct bio *
2544 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2545                 gfp_t gfp_flags)
2546 {
2547         struct btrfs_io_bio *btrfs_bio;
2548         struct bio *bio;
2549
2550         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2551
2552         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2553                 while (!bio && (nr_vecs /= 2)) {
2554                         bio = bio_alloc_bioset(gfp_flags,
2555                                                nr_vecs, btrfs_bioset);
2556                 }
2557         }
2558
2559         if (bio) {
2560                 bio->bi_size = 0;
2561                 bio->bi_bdev = bdev;
2562                 bio->bi_sector = first_sector;
2563                 btrfs_bio = btrfs_io_bio(bio);
2564                 btrfs_bio->csum = NULL;
2565                 btrfs_bio->csum_allocated = NULL;
2566                 btrfs_bio->end_io = NULL;
2567         }
2568         return bio;
2569 }
2570
2571 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2572 {
2573         return bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2574 }
2575
2576
2577 /* this also allocates from the btrfs_bioset */
2578 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2579 {
2580         struct btrfs_io_bio *btrfs_bio;
2581         struct bio *bio;
2582
2583         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2584         if (bio) {
2585                 btrfs_bio = btrfs_io_bio(bio);
2586                 btrfs_bio->csum = NULL;
2587                 btrfs_bio->csum_allocated = NULL;
2588                 btrfs_bio->end_io = NULL;
2589         }
2590         return bio;
2591 }
2592
2593
2594 static int __must_check submit_one_bio(int rw, struct bio *bio,
2595                                        int mirror_num, unsigned long bio_flags)
2596 {
2597         int ret = 0;
2598         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2599         struct page *page = bvec->bv_page;
2600         struct extent_io_tree *tree = bio->bi_private;
2601         u64 start;
2602
2603         start = page_offset(page) + bvec->bv_offset;
2604
2605         bio->bi_private = NULL;
2606
2607         bio_get(bio);
2608
2609         if (tree->ops && tree->ops->submit_bio_hook)
2610                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2611                                            mirror_num, bio_flags, start);
2612         else
2613                 btrfsic_submit_bio(rw, bio);
2614
2615         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2616                 ret = -EOPNOTSUPP;
2617         bio_put(bio);
2618         return ret;
2619 }
2620
2621 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2622                      unsigned long offset, size_t size, struct bio *bio,
2623                      unsigned long bio_flags)
2624 {
2625         int ret = 0;
2626         if (tree->ops && tree->ops->merge_bio_hook)
2627                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2628                                                 bio_flags);
2629         BUG_ON(ret < 0);
2630         return ret;
2631
2632 }
2633
2634 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2635                               struct page *page, sector_t sector,
2636                               size_t size, unsigned long offset,
2637                               struct block_device *bdev,
2638                               struct bio **bio_ret,
2639                               unsigned long max_pages,
2640                               bio_end_io_t end_io_func,
2641                               int mirror_num,
2642                               unsigned long prev_bio_flags,
2643                               unsigned long bio_flags)
2644 {
2645         int ret = 0;
2646         struct bio *bio;
2647         int nr;
2648         int contig = 0;
2649         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2650         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2651         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2652
2653         if (bio_ret && *bio_ret) {
2654                 bio = *bio_ret;
2655                 if (old_compressed)
2656                         contig = bio->bi_sector == sector;
2657                 else
2658                         contig = bio_end_sector(bio) == sector;
2659
2660                 if (prev_bio_flags != bio_flags || !contig ||
2661                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2662                     bio_add_page(bio, page, page_size, offset) < page_size) {
2663                         ret = submit_one_bio(rw, bio, mirror_num,
2664                                              prev_bio_flags);
2665                         if (ret < 0)
2666                                 return ret;
2667                         bio = NULL;
2668                 } else {
2669                         return 0;
2670                 }
2671         }
2672         if (this_compressed)
2673                 nr = BIO_MAX_PAGES;
2674         else
2675                 nr = bio_get_nr_vecs(bdev);
2676
2677         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2678         if (!bio)
2679                 return -ENOMEM;
2680
2681         bio_add_page(bio, page, page_size, offset);
2682         bio->bi_end_io = end_io_func;
2683         bio->bi_private = tree;
2684
2685         if (bio_ret)
2686                 *bio_ret = bio;
2687         else
2688                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2689
2690         return ret;
2691 }
2692
2693 static void attach_extent_buffer_page(struct extent_buffer *eb,
2694                                       struct page *page)
2695 {
2696         if (!PagePrivate(page)) {
2697                 SetPagePrivate(page);
2698                 page_cache_get(page);
2699                 set_page_private(page, (unsigned long)eb);
2700         } else {
2701                 WARN_ON(page->private != (unsigned long)eb);
2702         }
2703 }
2704
2705 void set_page_extent_mapped(struct page *page)
2706 {
2707         if (!PagePrivate(page)) {
2708                 SetPagePrivate(page);
2709                 page_cache_get(page);
2710                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2711         }
2712 }
2713
2714 static struct extent_map *
2715 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2716                  u64 start, u64 len, get_extent_t *get_extent,
2717                  struct extent_map **em_cached)
2718 {
2719         struct extent_map *em;
2720
2721         if (em_cached && *em_cached) {
2722                 em = *em_cached;
2723                 if (em->in_tree && start >= em->start &&
2724                     start < extent_map_end(em)) {
2725                         atomic_inc(&em->refs);
2726                         return em;
2727                 }
2728
2729                 free_extent_map(em);
2730                 *em_cached = NULL;
2731         }
2732
2733         em = get_extent(inode, page, pg_offset, start, len, 0);
2734         if (em_cached && !IS_ERR_OR_NULL(em)) {
2735                 BUG_ON(*em_cached);
2736                 atomic_inc(&em->refs);
2737                 *em_cached = em;
2738         }
2739         return em;
2740 }
2741 /*
2742  * basic readpage implementation.  Locked extent state structs are inserted
2743  * into the tree that are removed when the IO is done (by the end_io
2744  * handlers)
2745  * XXX JDM: This needs looking at to ensure proper page locking
2746  */
2747 static int __do_readpage(struct extent_io_tree *tree,
2748                          struct page *page,
2749                          get_extent_t *get_extent,
2750                          struct extent_map **em_cached,
2751                          struct bio **bio, int mirror_num,
2752                          unsigned long *bio_flags, int rw)
2753 {
2754         struct inode *inode = page->mapping->host;
2755         u64 start = page_offset(page);
2756         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2757         u64 end;
2758         u64 cur = start;
2759         u64 extent_offset;
2760         u64 last_byte = i_size_read(inode);
2761         u64 block_start;
2762         u64 cur_end;
2763         sector_t sector;
2764         struct extent_map *em;
2765         struct block_device *bdev;
2766         int ret;
2767         int nr = 0;
2768         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2769         size_t pg_offset = 0;
2770         size_t iosize;
2771         size_t disk_io_size;
2772         size_t blocksize = inode->i_sb->s_blocksize;
2773         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2774
2775         set_page_extent_mapped(page);
2776
2777         end = page_end;
2778         if (!PageUptodate(page)) {
2779                 if (cleancache_get_page(page) == 0) {
2780                         BUG_ON(blocksize != PAGE_SIZE);
2781                         unlock_extent(tree, start, end);
2782                         goto out;
2783                 }
2784         }
2785
2786         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2787                 char *userpage;
2788                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2789
2790                 if (zero_offset) {
2791                         iosize = PAGE_CACHE_SIZE - zero_offset;
2792                         userpage = kmap_atomic(page);
2793                         memset(userpage + zero_offset, 0, iosize);
2794                         flush_dcache_page(page);
2795                         kunmap_atomic(userpage);
2796                 }
2797         }
2798         while (cur <= end) {
2799                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2800
2801                 if (cur >= last_byte) {
2802                         char *userpage;
2803                         struct extent_state *cached = NULL;
2804
2805                         iosize = PAGE_CACHE_SIZE - pg_offset;
2806                         userpage = kmap_atomic(page);
2807                         memset(userpage + pg_offset, 0, iosize);
2808                         flush_dcache_page(page);
2809                         kunmap_atomic(userpage);
2810                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2811                                             &cached, GFP_NOFS);
2812                         if (!parent_locked)
2813                                 unlock_extent_cached(tree, cur,
2814                                                      cur + iosize - 1,
2815                                                      &cached, GFP_NOFS);
2816                         break;
2817                 }
2818                 em = __get_extent_map(inode, page, pg_offset, cur,
2819                                       end - cur + 1, get_extent, em_cached);
2820                 if (IS_ERR_OR_NULL(em)) {
2821                         SetPageError(page);
2822                         if (!parent_locked)
2823                                 unlock_extent(tree, cur, end);
2824                         break;
2825                 }
2826                 extent_offset = cur - em->start;
2827                 BUG_ON(extent_map_end(em) <= cur);
2828                 BUG_ON(end < cur);
2829
2830                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2831                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2832                         extent_set_compress_type(&this_bio_flag,
2833                                                  em->compress_type);
2834                 }
2835
2836                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2837                 cur_end = min(extent_map_end(em) - 1, end);
2838                 iosize = ALIGN(iosize, blocksize);
2839                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2840                         disk_io_size = em->block_len;
2841                         sector = em->block_start >> 9;
2842                 } else {
2843                         sector = (em->block_start + extent_offset) >> 9;
2844                         disk_io_size = iosize;
2845                 }
2846                 bdev = em->bdev;
2847                 block_start = em->block_start;
2848                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2849                         block_start = EXTENT_MAP_HOLE;
2850                 free_extent_map(em);
2851                 em = NULL;
2852
2853                 /* we've found a hole, just zero and go on */
2854                 if (block_start == EXTENT_MAP_HOLE) {
2855                         char *userpage;
2856                         struct extent_state *cached = NULL;
2857
2858                         userpage = kmap_atomic(page);
2859                         memset(userpage + pg_offset, 0, iosize);
2860                         flush_dcache_page(page);
2861                         kunmap_atomic(userpage);
2862
2863                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2864                                             &cached, GFP_NOFS);
2865                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2866                                              &cached, GFP_NOFS);
2867                         cur = cur + iosize;
2868                         pg_offset += iosize;
2869                         continue;
2870                 }
2871                 /* the get_extent function already copied into the page */
2872                 if (test_range_bit(tree, cur, cur_end,
2873                                    EXTENT_UPTODATE, 1, NULL)) {
2874                         check_page_uptodate(tree, page);
2875                         if (!parent_locked)
2876                                 unlock_extent(tree, cur, cur + iosize - 1);
2877                         cur = cur + iosize;
2878                         pg_offset += iosize;
2879                         continue;
2880                 }
2881                 /* we have an inline extent but it didn't get marked up
2882                  * to date.  Error out
2883                  */
2884                 if (block_start == EXTENT_MAP_INLINE) {
2885                         SetPageError(page);
2886                         if (!parent_locked)
2887                                 unlock_extent(tree, cur, cur + iosize - 1);
2888                         cur = cur + iosize;
2889                         pg_offset += iosize;
2890                         continue;
2891                 }
2892
2893                 pnr -= page->index;
2894                 ret = submit_extent_page(rw, tree, page,
2895                                          sector, disk_io_size, pg_offset,
2896                                          bdev, bio, pnr,
2897                                          end_bio_extent_readpage, mirror_num,
2898                                          *bio_flags,
2899                                          this_bio_flag);
2900                 if (!ret) {
2901                         nr++;
2902                         *bio_flags = this_bio_flag;
2903                 } else {
2904                         SetPageError(page);
2905                         if (!parent_locked)
2906                                 unlock_extent(tree, cur, cur + iosize - 1);
2907                 }
2908                 cur = cur + iosize;
2909                 pg_offset += iosize;
2910         }
2911 out:
2912         if (!nr) {
2913                 if (!PageError(page))
2914                         SetPageUptodate(page);
2915                 unlock_page(page);
2916         }
2917         return 0;
2918 }
2919
2920 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
2921                                              struct page *pages[], int nr_pages,
2922                                              u64 start, u64 end,
2923                                              get_extent_t *get_extent,
2924                                              struct extent_map **em_cached,
2925                                              struct bio **bio, int mirror_num,
2926                                              unsigned long *bio_flags, int rw)
2927 {
2928         struct inode *inode;
2929         struct btrfs_ordered_extent *ordered;
2930         int index;
2931
2932         inode = pages[0]->mapping->host;
2933         while (1) {
2934                 lock_extent(tree, start, end);
2935                 ordered = btrfs_lookup_ordered_range(inode, start,
2936                                                      end - start + 1);
2937                 if (!ordered)
2938                         break;
2939                 unlock_extent(tree, start, end);
2940                 btrfs_start_ordered_extent(inode, ordered, 1);
2941                 btrfs_put_ordered_extent(ordered);
2942         }
2943
2944         for (index = 0; index < nr_pages; index++) {
2945                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
2946                               mirror_num, bio_flags, rw);
2947                 page_cache_release(pages[index]);
2948         }
2949 }
2950
2951 static void __extent_readpages(struct extent_io_tree *tree,
2952                                struct page *pages[],
2953                                int nr_pages, get_extent_t *get_extent,
2954                                struct extent_map **em_cached,
2955                                struct bio **bio, int mirror_num,
2956                                unsigned long *bio_flags, int rw)
2957 {
2958         u64 start = 0;
2959         u64 end = 0;
2960         u64 page_start;
2961         int index;
2962         int first_index = 0;
2963
2964         for (index = 0; index < nr_pages; index++) {
2965                 page_start = page_offset(pages[index]);
2966                 if (!end) {
2967                         start = page_start;
2968                         end = start + PAGE_CACHE_SIZE - 1;
2969                         first_index = index;
2970                 } else if (end + 1 == page_start) {
2971                         end += PAGE_CACHE_SIZE;
2972                 } else {
2973                         __do_contiguous_readpages(tree, &pages[first_index],
2974                                                   index - first_index, start,
2975                                                   end, get_extent, em_cached,
2976                                                   bio, mirror_num, bio_flags,
2977                                                   rw);
2978                         start = page_start;
2979                         end = start + PAGE_CACHE_SIZE - 1;
2980                         first_index = index;
2981                 }
2982         }
2983
2984         if (end)
2985                 __do_contiguous_readpages(tree, &pages[first_index],
2986                                           index - first_index, start,
2987                                           end, get_extent, em_cached, bio,
2988                                           mirror_num, bio_flags, rw);
2989 }
2990
2991 static int __extent_read_full_page(struct extent_io_tree *tree,
2992                                    struct page *page,
2993                                    get_extent_t *get_extent,
2994                                    struct bio **bio, int mirror_num,
2995                                    unsigned long *bio_flags, int rw)
2996 {
2997         struct inode *inode = page->mapping->host;
2998         struct btrfs_ordered_extent *ordered;
2999         u64 start = page_offset(page);
3000         u64 end = start + PAGE_CACHE_SIZE - 1;
3001         int ret;
3002
3003         while (1) {
3004                 lock_extent(tree, start, end);
3005                 ordered = btrfs_lookup_ordered_extent(inode, start);
3006                 if (!ordered)
3007                         break;
3008                 unlock_extent(tree, start, end);
3009                 btrfs_start_ordered_extent(inode, ordered, 1);
3010                 btrfs_put_ordered_extent(ordered);
3011         }
3012
3013         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3014                             bio_flags, rw);
3015         return ret;
3016 }
3017
3018 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3019                             get_extent_t *get_extent, int mirror_num)
3020 {
3021         struct bio *bio = NULL;
3022         unsigned long bio_flags = 0;
3023         int ret;
3024
3025         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3026                                       &bio_flags, READ);
3027         if (bio)
3028                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3029         return ret;
3030 }
3031
3032 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3033                                  get_extent_t *get_extent, int mirror_num)
3034 {
3035         struct bio *bio = NULL;
3036         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3037         int ret;
3038
3039         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3040                                       &bio_flags, READ);
3041         if (bio)
3042                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3043         return ret;
3044 }
3045
3046 static noinline void update_nr_written(struct page *page,
3047                                       struct writeback_control *wbc,
3048                                       unsigned long nr_written)
3049 {
3050         wbc->nr_to_write -= nr_written;
3051         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3052             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3053                 page->mapping->writeback_index = page->index + nr_written;
3054 }
3055
3056 /*
3057  * the writepage semantics are similar to regular writepage.  extent
3058  * records are inserted to lock ranges in the tree, and as dirty areas
3059  * are found, they are marked writeback.  Then the lock bits are removed
3060  * and the end_io handler clears the writeback ranges
3061  */
3062 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3063                               void *data)
3064 {
3065         struct inode *inode = page->mapping->host;
3066         struct extent_page_data *epd = data;
3067         struct extent_io_tree *tree = epd->tree;
3068         u64 start = page_offset(page);
3069         u64 delalloc_start;
3070         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3071         u64 end;
3072         u64 cur = start;
3073         u64 extent_offset;
3074         u64 last_byte = i_size_read(inode);
3075         u64 block_start;
3076         u64 iosize;
3077         sector_t sector;
3078         struct extent_state *cached_state = NULL;
3079         struct extent_map *em;
3080         struct block_device *bdev;
3081         int ret;
3082         int nr = 0;
3083         size_t pg_offset = 0;
3084         size_t blocksize;
3085         loff_t i_size = i_size_read(inode);
3086         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3087         u64 nr_delalloc;
3088         u64 delalloc_end;
3089         int page_started;
3090         int compressed;
3091         int write_flags;
3092         unsigned long nr_written = 0;
3093         bool fill_delalloc = true;
3094
3095         if (wbc->sync_mode == WB_SYNC_ALL)
3096                 write_flags = WRITE_SYNC;
3097         else
3098                 write_flags = WRITE;
3099
3100         trace___extent_writepage(page, inode, wbc);
3101
3102         WARN_ON(!PageLocked(page));
3103
3104         ClearPageError(page);
3105
3106         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3107         if (page->index > end_index ||
3108            (page->index == end_index && !pg_offset)) {
3109                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3110                 unlock_page(page);
3111                 return 0;
3112         }
3113
3114         if (page->index == end_index) {
3115                 char *userpage;
3116
3117                 userpage = kmap_atomic(page);
3118                 memset(userpage + pg_offset, 0,
3119                        PAGE_CACHE_SIZE - pg_offset);
3120                 kunmap_atomic(userpage);
3121                 flush_dcache_page(page);
3122         }
3123         pg_offset = 0;
3124
3125         set_page_extent_mapped(page);
3126
3127         if (!tree->ops || !tree->ops->fill_delalloc)
3128                 fill_delalloc = false;
3129
3130         delalloc_start = start;
3131         delalloc_end = 0;
3132         page_started = 0;
3133         if (!epd->extent_locked && fill_delalloc) {
3134                 u64 delalloc_to_write = 0;
3135                 /*
3136                  * make sure the wbc mapping index is at least updated
3137                  * to this page.
3138                  */
3139                 update_nr_written(page, wbc, 0);
3140
3141                 while (delalloc_end < page_end) {
3142                         nr_delalloc = find_lock_delalloc_range(inode, tree,
3143                                                        page,
3144                                                        &delalloc_start,
3145                                                        &delalloc_end,
3146                                                        128 * 1024 * 1024);
3147                         if (nr_delalloc == 0) {
3148                                 delalloc_start = delalloc_end + 1;
3149                                 continue;
3150                         }
3151                         ret = tree->ops->fill_delalloc(inode, page,
3152                                                        delalloc_start,
3153                                                        delalloc_end,
3154                                                        &page_started,
3155                                                        &nr_written);
3156                         /* File system has been set read-only */
3157                         if (ret) {
3158                                 SetPageError(page);
3159                                 goto done;
3160                         }
3161                         /*
3162                          * delalloc_end is already one less than the total
3163                          * length, so we don't subtract one from
3164                          * PAGE_CACHE_SIZE
3165                          */
3166                         delalloc_to_write += (delalloc_end - delalloc_start +
3167                                               PAGE_CACHE_SIZE) >>
3168                                               PAGE_CACHE_SHIFT;
3169                         delalloc_start = delalloc_end + 1;
3170                 }
3171                 if (wbc->nr_to_write < delalloc_to_write) {
3172                         int thresh = 8192;
3173
3174                         if (delalloc_to_write < thresh * 2)
3175                                 thresh = delalloc_to_write;
3176                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
3177                                                  thresh);
3178                 }
3179
3180                 /* did the fill delalloc function already unlock and start
3181                  * the IO?
3182                  */
3183                 if (page_started) {
3184                         ret = 0;
3185                         /*
3186                          * we've unlocked the page, so we can't update
3187                          * the mapping's writeback index, just update
3188                          * nr_to_write.
3189                          */
3190                         wbc->nr_to_write -= nr_written;
3191                         goto done_unlocked;
3192                 }
3193         }
3194         if (tree->ops && tree->ops->writepage_start_hook) {
3195                 ret = tree->ops->writepage_start_hook(page, start,
3196                                                       page_end);
3197                 if (ret) {
3198                         /* Fixup worker will requeue */
3199                         if (ret == -EBUSY)
3200                                 wbc->pages_skipped++;
3201                         else
3202                                 redirty_page_for_writepage(wbc, page);
3203                         update_nr_written(page, wbc, nr_written);
3204                         unlock_page(page);
3205                         ret = 0;
3206                         goto done_unlocked;
3207                 }
3208         }
3209
3210         /*
3211          * we don't want to touch the inode after unlocking the page,
3212          * so we update the mapping writeback index now
3213          */
3214         update_nr_written(page, wbc, nr_written + 1);
3215
3216         end = page_end;
3217         if (last_byte <= start) {
3218                 if (tree->ops && tree->ops->writepage_end_io_hook)
3219                         tree->ops->writepage_end_io_hook(page, start,
3220                                                          page_end, NULL, 1);
3221                 goto done;
3222         }
3223
3224         blocksize = inode->i_sb->s_blocksize;
3225
3226         while (cur <= end) {
3227                 if (cur >= last_byte) {
3228                         if (tree->ops && tree->ops->writepage_end_io_hook)
3229                                 tree->ops->writepage_end_io_hook(page, cur,
3230                                                          page_end, NULL, 1);
3231                         break;
3232                 }
3233                 em = epd->get_extent(inode, page, pg_offset, cur,
3234                                      end - cur + 1, 1);
3235                 if (IS_ERR_OR_NULL(em)) {
3236                         SetPageError(page);
3237                         break;
3238                 }
3239
3240                 extent_offset = cur - em->start;
3241                 BUG_ON(extent_map_end(em) <= cur);
3242                 BUG_ON(end < cur);
3243                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3244                 iosize = ALIGN(iosize, blocksize);
3245                 sector = (em->block_start + extent_offset) >> 9;
3246                 bdev = em->bdev;
3247                 block_start = em->block_start;
3248                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3249                 free_extent_map(em);
3250                 em = NULL;
3251
3252                 /*
3253                  * compressed and inline extents are written through other
3254                  * paths in the FS
3255                  */
3256                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3257                     block_start == EXTENT_MAP_INLINE) {
3258                         /*
3259                          * end_io notification does not happen here for
3260                          * compressed extents
3261                          */
3262                         if (!compressed && tree->ops &&
3263                             tree->ops->writepage_end_io_hook)
3264                                 tree->ops->writepage_end_io_hook(page, cur,
3265                                                          cur + iosize - 1,
3266                                                          NULL, 1);
3267                         else if (compressed) {
3268                                 /* we don't want to end_page_writeback on
3269                                  * a compressed extent.  this happens
3270                                  * elsewhere
3271                                  */
3272                                 nr++;
3273                         }
3274
3275                         cur += iosize;
3276                         pg_offset += iosize;
3277                         continue;
3278                 }
3279                 /* leave this out until we have a page_mkwrite call */
3280                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
3281                                    EXTENT_DIRTY, 0, NULL)) {
3282                         cur = cur + iosize;
3283                         pg_offset += iosize;
3284                         continue;
3285                 }
3286
3287                 if (tree->ops && tree->ops->writepage_io_hook) {
3288                         ret = tree->ops->writepage_io_hook(page, cur,
3289                                                 cur + iosize - 1);
3290                 } else {
3291                         ret = 0;
3292                 }
3293                 if (ret) {
3294                         SetPageError(page);
3295                 } else {
3296                         unsigned long max_nr = end_index + 1;
3297
3298                         set_range_writeback(tree, cur, cur + iosize - 1);
3299                         if (!PageWriteback(page)) {
3300                                 printk(KERN_ERR "btrfs warning page %lu not "
3301                                        "writeback, cur %llu end %llu\n",
3302                                        page->index, cur, end);
3303                         }
3304
3305                         ret = submit_extent_page(write_flags, tree, page,
3306                                                  sector, iosize, pg_offset,
3307                                                  bdev, &epd->bio, max_nr,
3308                                                  end_bio_extent_writepage,
3309                                                  0, 0, 0);
3310                         if (ret)
3311                                 SetPageError(page);
3312                 }
3313                 cur = cur + iosize;
3314                 pg_offset += iosize;
3315                 nr++;
3316         }
3317 done:
3318         if (nr == 0) {
3319                 /* make sure the mapping tag for page dirty gets cleared */
3320                 set_page_writeback(page);
3321                 end_page_writeback(page);
3322         }
3323         unlock_page(page);
3324
3325 done_unlocked:
3326
3327         /* drop our reference on any cached states */
3328         free_extent_state(cached_state);
3329         return 0;
3330 }
3331
3332 static int eb_wait(void *word)
3333 {
3334         io_schedule();
3335         return 0;
3336 }
3337
3338 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3339 {
3340         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3341                     TASK_UNINTERRUPTIBLE);
3342 }
3343
3344 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3345                                      struct btrfs_fs_info *fs_info,
3346                                      struct extent_page_data *epd)
3347 {
3348         unsigned long i, num_pages;
3349         int flush = 0;
3350         int ret = 0;
3351
3352         if (!btrfs_try_tree_write_lock(eb)) {
3353                 flush = 1;
3354                 flush_write_bio(epd);
3355                 btrfs_tree_lock(eb);
3356         }
3357
3358         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3359                 btrfs_tree_unlock(eb);
3360                 if (!epd->sync_io)
3361                         return 0;
3362                 if (!flush) {
3363                         flush_write_bio(epd);
3364                         flush = 1;
3365                 }
3366                 while (1) {
3367                         wait_on_extent_buffer_writeback(eb);
3368                         btrfs_tree_lock(eb);
3369                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3370                                 break;
3371                         btrfs_tree_unlock(eb);
3372                 }
3373         }
3374
3375         /*
3376          * We need to do this to prevent races in people who check if the eb is
3377          * under IO since we can end up having no IO bits set for a short period
3378          * of time.
3379          */
3380         spin_lock(&eb->refs_lock);
3381         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3382                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3383                 spin_unlock(&eb->refs_lock);
3384                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3385                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3386                                      -eb->len,
3387                                      fs_info->dirty_metadata_batch);
3388                 ret = 1;
3389         } else {
3390                 spin_unlock(&eb->refs_lock);
3391         }
3392
3393         btrfs_tree_unlock(eb);
3394
3395         if (!ret)
3396                 return ret;
3397
3398         num_pages = num_extent_pages(eb->start, eb->len);
3399         for (i = 0; i < num_pages; i++) {
3400                 struct page *p = extent_buffer_page(eb, i);
3401
3402                 if (!trylock_page(p)) {
3403                         if (!flush) {
3404                                 flush_write_bio(epd);
3405                                 flush = 1;
3406                         }
3407                         lock_page(p);
3408                 }
3409         }
3410
3411         return ret;
3412 }
3413
3414 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3415 {
3416         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3417         smp_mb__after_clear_bit();
3418         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3419 }
3420
3421 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3422 {
3423         int uptodate = err == 0;
3424         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
3425         struct extent_buffer *eb;
3426         int done;
3427
3428         do {
3429                 struct page *page = bvec->bv_page;
3430
3431                 bvec--;
3432                 eb = (struct extent_buffer *)page->private;
3433                 BUG_ON(!eb);
3434                 done = atomic_dec_and_test(&eb->io_pages);
3435
3436                 if (!uptodate || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3437                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3438                         ClearPageUptodate(page);
3439                         SetPageError(page);
3440                 }
3441
3442                 end_page_writeback(page);
3443
3444                 if (!done)
3445                         continue;
3446
3447                 end_extent_buffer_writeback(eb);
3448         } while (bvec >= bio->bi_io_vec);
3449
3450         bio_put(bio);
3451
3452 }
3453
3454 static int write_one_eb(struct extent_buffer *eb,
3455                         struct btrfs_fs_info *fs_info,
3456                         struct writeback_control *wbc,
3457                         struct extent_page_data *epd)
3458 {
3459         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3460         u64 offset = eb->start;
3461         unsigned long i, num_pages;
3462         unsigned long bio_flags = 0;
3463         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3464         int ret = 0;
3465
3466         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3467         num_pages = num_extent_pages(eb->start, eb->len);
3468         atomic_set(&eb->io_pages, num_pages);
3469         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3470                 bio_flags = EXTENT_BIO_TREE_LOG;
3471
3472         for (i = 0; i < num_pages; i++) {
3473                 struct page *p = extent_buffer_page(eb, i);
3474
3475                 clear_page_dirty_for_io(p);
3476                 set_page_writeback(p);
3477                 ret = submit_extent_page(rw, eb->tree, p, offset >> 9,
3478                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3479                                          -1, end_bio_extent_buffer_writepage,
3480                                          0, epd->bio_flags, bio_flags);
3481                 epd->bio_flags = bio_flags;
3482                 if (ret) {
3483                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3484                         SetPageError(p);
3485                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3486                                 end_extent_buffer_writeback(eb);
3487                         ret = -EIO;
3488                         break;
3489                 }
3490                 offset += PAGE_CACHE_SIZE;
3491                 update_nr_written(p, wbc, 1);
3492                 unlock_page(p);
3493         }
3494
3495         if (unlikely(ret)) {
3496                 for (; i < num_pages; i++) {
3497                         struct page *p = extent_buffer_page(eb, i);
3498                         unlock_page(p);
3499                 }
3500         }
3501
3502         return ret;
3503 }
3504
3505 int btree_write_cache_pages(struct address_space *mapping,
3506                                    struct writeback_control *wbc)
3507 {
3508         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3509         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3510         struct extent_buffer *eb, *prev_eb = NULL;
3511         struct extent_page_data epd = {
3512                 .bio = NULL,
3513                 .tree = tree,
3514                 .extent_locked = 0,
3515                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3516                 .bio_flags = 0,
3517         };
3518         int ret = 0;
3519         int done = 0;
3520         int nr_to_write_done = 0;
3521         struct pagevec pvec;
3522         int nr_pages;
3523         pgoff_t index;
3524         pgoff_t end;            /* Inclusive */
3525         int scanned = 0;
3526         int tag;
3527
3528         pagevec_init(&pvec, 0);
3529         if (wbc->range_cyclic) {
3530                 index = mapping->writeback_index; /* Start from prev offset */
3531                 end = -1;
3532         } else {
3533                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3534                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3535                 scanned = 1;
3536         }
3537         if (wbc->sync_mode == WB_SYNC_ALL)
3538                 tag = PAGECACHE_TAG_TOWRITE;
3539         else
3540                 tag = PAGECACHE_TAG_DIRTY;
3541 retry:
3542         if (wbc->sync_mode == WB_SYNC_ALL)
3543                 tag_pages_for_writeback(mapping, index, end);
3544         while (!done && !nr_to_write_done && (index <= end) &&
3545                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3546                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3547                 unsigned i;
3548
3549                 scanned = 1;
3550                 for (i = 0; i < nr_pages; i++) {
3551                         struct page *page = pvec.pages[i];
3552
3553                         if (!PagePrivate(page))
3554                                 continue;
3555
3556                         if (!wbc->range_cyclic && page->index > end) {
3557                                 done = 1;
3558                                 break;
3559                         }
3560
3561                         spin_lock(&mapping->private_lock);
3562                         if (!PagePrivate(page)) {
3563                                 spin_unlock(&mapping->private_lock);
3564                                 continue;
3565                         }
3566
3567                         eb = (struct extent_buffer *)page->private;
3568
3569                         /*
3570                          * Shouldn't happen and normally this would be a BUG_ON
3571                          * but no sense in crashing the users box for something
3572                          * we can survive anyway.
3573                          */
3574                         if (WARN_ON(!eb)) {
3575                                 spin_unlock(&mapping->private_lock);
3576                                 continue;
3577                         }
3578
3579                         if (eb == prev_eb) {
3580                                 spin_unlock(&mapping->private_lock);
3581                                 continue;
3582                         }
3583
3584                         ret = atomic_inc_not_zero(&eb->refs);
3585                         spin_unlock(&mapping->private_lock);
3586                         if (!ret)
3587                                 continue;
3588
3589                         prev_eb = eb;
3590                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3591                         if (!ret) {
3592                                 free_extent_buffer(eb);
3593                                 continue;
3594                         }
3595
3596                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3597                         if (ret) {
3598                                 done = 1;
3599                                 free_extent_buffer(eb);
3600                                 break;
3601                         }
3602                         free_extent_buffer(eb);
3603
3604                         /*
3605                          * the filesystem may choose to bump up nr_to_write.
3606                          * We have to make sure to honor the new nr_to_write
3607                          * at any time
3608                          */
3609                         nr_to_write_done = wbc->nr_to_write <= 0;
3610                 }
3611                 pagevec_release(&pvec);
3612                 cond_resched();
3613         }
3614         if (!scanned && !done) {
3615                 /*
3616                  * We hit the last page and there is more work to be done: wrap
3617                  * back to the start of the file
3618                  */
3619                 scanned = 1;
3620                 index = 0;
3621                 goto retry;
3622         }
3623         flush_write_bio(&epd);
3624         return ret;
3625 }
3626
3627 /**
3628  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3629  * @mapping: address space structure to write
3630  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3631  * @writepage: function called for each page
3632  * @data: data passed to writepage function
3633  *
3634  * If a page is already under I/O, write_cache_pages() skips it, even
3635  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3636  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3637  * and msync() need to guarantee that all the data which was dirty at the time
3638  * the call was made get new I/O started against them.  If wbc->sync_mode is
3639  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3640  * existing IO to complete.
3641  */
3642 static int extent_write_cache_pages(struct extent_io_tree *tree,
3643                              struct address_space *mapping,
3644                              struct writeback_control *wbc,
3645                              writepage_t writepage, void *data,
3646                              void (*flush_fn)(void *))
3647 {
3648         struct inode *inode = mapping->host;
3649         int ret = 0;
3650         int done = 0;
3651         int nr_to_write_done = 0;
3652         struct pagevec pvec;
3653         int nr_pages;
3654         pgoff_t index;
3655         pgoff_t end;            /* Inclusive */
3656         int scanned = 0;
3657         int tag;
3658
3659         /*
3660          * We have to hold onto the inode so that ordered extents can do their
3661          * work when the IO finishes.  The alternative to this is failing to add
3662          * an ordered extent if the igrab() fails there and that is a huge pain
3663          * to deal with, so instead just hold onto the inode throughout the
3664          * writepages operation.  If it fails here we are freeing up the inode
3665          * anyway and we'd rather not waste our time writing out stuff that is
3666          * going to be truncated anyway.
3667          */
3668         if (!igrab(inode))
3669                 return 0;
3670
3671         pagevec_init(&pvec, 0);
3672         if (wbc->range_cyclic) {
3673                 index = mapping->writeback_index; /* Start from prev offset */
3674                 end = -1;
3675         } else {
3676                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3677                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3678                 scanned = 1;
3679         }
3680         if (wbc->sync_mode == WB_SYNC_ALL)
3681                 tag = PAGECACHE_TAG_TOWRITE;
3682         else
3683                 tag = PAGECACHE_TAG_DIRTY;
3684 retry:
3685         if (wbc->sync_mode == WB_SYNC_ALL)
3686                 tag_pages_for_writeback(mapping, index, end);
3687         while (!done && !nr_to_write_done && (index <= end) &&
3688                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3689                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3690                 unsigned i;
3691
3692                 scanned = 1;
3693                 for (i = 0; i < nr_pages; i++) {
3694                         struct page *page = pvec.pages[i];
3695
3696                         /*
3697                          * At this point we hold neither mapping->tree_lock nor
3698                          * lock on the page itself: the page may be truncated or
3699                          * invalidated (changing page->mapping to NULL), or even
3700                          * swizzled back from swapper_space to tmpfs file
3701                          * mapping
3702                          */
3703                         if (!trylock_page(page)) {
3704                                 flush_fn(data);
3705                                 lock_page(page);
3706                         }
3707
3708                         if (unlikely(page->mapping != mapping)) {
3709                                 unlock_page(page);
3710                                 continue;
3711                         }
3712
3713                         if (!wbc->range_cyclic && page->index > end) {
3714                                 done = 1;
3715                                 unlock_page(page);
3716                                 continue;
3717                         }
3718
3719                         if (wbc->sync_mode != WB_SYNC_NONE) {
3720                                 if (PageWriteback(page))
3721                                         flush_fn(data);
3722                                 wait_on_page_writeback(page);
3723                         }
3724
3725                         if (PageWriteback(page) ||
3726                             !clear_page_dirty_for_io(page)) {
3727                                 unlock_page(page);
3728                                 continue;
3729                         }
3730
3731                         ret = (*writepage)(page, wbc, data);
3732
3733                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3734                                 unlock_page(page);
3735                                 ret = 0;
3736                         }
3737                         if (ret)
3738                                 done = 1;
3739
3740                         /*
3741                          * the filesystem may choose to bump up nr_to_write.
3742                          * We have to make sure to honor the new nr_to_write
3743                          * at any time
3744                          */
3745                         nr_to_write_done = wbc->nr_to_write <= 0;
3746                 }
3747                 pagevec_release(&pvec);
3748                 cond_resched();
3749         }
3750         if (!scanned && !done) {
3751                 /*
3752                  * We hit the last page and there is more work to be done: wrap
3753                  * back to the start of the file
3754                  */
3755                 scanned = 1;
3756                 index = 0;
3757                 goto retry;
3758         }
3759         btrfs_add_delayed_iput(inode);
3760         return ret;
3761 }
3762
3763 static void flush_epd_write_bio(struct extent_page_data *epd)
3764 {
3765         if (epd->bio) {
3766                 int rw = WRITE;
3767                 int ret;
3768
3769                 if (epd->sync_io)
3770                         rw = WRITE_SYNC;
3771
3772                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3773                 BUG_ON(ret < 0); /* -ENOMEM */
3774                 epd->bio = NULL;
3775         }
3776 }
3777
3778 static noinline void flush_write_bio(void *data)
3779 {
3780         struct extent_page_data *epd = data;
3781         flush_epd_write_bio(epd);
3782 }
3783
3784 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3785                           get_extent_t *get_extent,
3786                           struct writeback_control *wbc)
3787 {
3788         int ret;
3789         struct extent_page_data epd = {
3790                 .bio = NULL,
3791                 .tree = tree,
3792                 .get_extent = get_extent,
3793                 .extent_locked = 0,
3794                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3795                 .bio_flags = 0,
3796         };
3797
3798         ret = __extent_writepage(page, wbc, &epd);
3799
3800         flush_epd_write_bio(&epd);
3801         return ret;
3802 }
3803
3804 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3805                               u64 start, u64 end, get_extent_t *get_extent,
3806                               int mode)
3807 {
3808         int ret = 0;
3809         struct address_space *mapping = inode->i_mapping;
3810         struct page *page;
3811         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3812                 PAGE_CACHE_SHIFT;
3813
3814         struct extent_page_data epd = {
3815                 .bio = NULL,
3816                 .tree = tree,
3817                 .get_extent = get_extent,
3818                 .extent_locked = 1,
3819                 .sync_io = mode == WB_SYNC_ALL,
3820                 .bio_flags = 0,
3821         };
3822         struct writeback_control wbc_writepages = {
3823                 .sync_mode      = mode,
3824                 .nr_to_write    = nr_pages * 2,
3825                 .range_start    = start,
3826                 .range_end      = end + 1,
3827         };
3828
3829         while (start <= end) {
3830                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3831                 if (clear_page_dirty_for_io(page))
3832                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3833                 else {
3834                         if (tree->ops && tree->ops->writepage_end_io_hook)
3835                                 tree->ops->writepage_end_io_hook(page, start,
3836                                                  start + PAGE_CACHE_SIZE - 1,
3837                                                  NULL, 1);
3838                         unlock_page(page);
3839                 }
3840                 page_cache_release(page);
3841                 start += PAGE_CACHE_SIZE;
3842         }
3843
3844         flush_epd_write_bio(&epd);
3845         return ret;
3846 }
3847
3848 int extent_writepages(struct extent_io_tree *tree,
3849                       struct address_space *mapping,
3850                       get_extent_t *get_extent,
3851                       struct writeback_control *wbc)
3852 {
3853         int ret = 0;
3854         struct extent_page_data epd = {
3855                 .bio = NULL,
3856                 .tree = tree,
3857                 .get_extent = get_extent,
3858                 .extent_locked = 0,
3859                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3860                 .bio_flags = 0,
3861         };
3862
3863         ret = extent_write_cache_pages(tree, mapping, wbc,
3864                                        __extent_writepage, &epd,
3865                                        flush_write_bio);
3866         flush_epd_write_bio(&epd);
3867         return ret;
3868 }
3869
3870 int extent_readpages(struct extent_io_tree *tree,
3871                      struct address_space *mapping,
3872                      struct list_head *pages, unsigned nr_pages,
3873                      get_extent_t get_extent)
3874 {
3875         struct bio *bio = NULL;
3876         unsigned page_idx;
3877         unsigned long bio_flags = 0;
3878         struct page *pagepool[16];
3879         struct page *page;
3880         struct extent_map *em_cached = NULL;
3881         int nr = 0;
3882
3883         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3884                 page = list_entry(pages->prev, struct page, lru);
3885
3886                 prefetchw(&page->flags);
3887                 list_del(&page->lru);
3888                 if (add_to_page_cache_lru(page, mapping,
3889                                         page->index, GFP_NOFS)) {
3890                         page_cache_release(page);
3891                         continue;
3892                 }
3893
3894                 pagepool[nr++] = page;
3895                 if (nr < ARRAY_SIZE(pagepool))
3896                         continue;
3897                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3898                                    &bio, 0, &bio_flags, READ);
3899                 nr = 0;
3900         }
3901         if (nr)
3902                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3903                                    &bio, 0, &bio_flags, READ);
3904
3905         if (em_cached)
3906                 free_extent_map(em_cached);
3907
3908         BUG_ON(!list_empty(pages));
3909         if (bio)
3910                 return submit_one_bio(READ, bio, 0, bio_flags);
3911         return 0;
3912 }
3913
3914 /*
3915  * basic invalidatepage code, this waits on any locked or writeback
3916  * ranges corresponding to the page, and then deletes any extent state
3917  * records from the tree
3918  */
3919 int extent_invalidatepage(struct extent_io_tree *tree,
3920                           struct page *page, unsigned long offset)
3921 {
3922         struct extent_state *cached_state = NULL;
3923         u64 start = page_offset(page);
3924         u64 end = start + PAGE_CACHE_SIZE - 1;
3925         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3926
3927         start += ALIGN(offset, blocksize);
3928         if (start > end)
3929                 return 0;
3930
3931         lock_extent_bits(tree, start, end, 0, &cached_state);
3932         wait_on_page_writeback(page);
3933         clear_extent_bit(tree, start, end,
3934                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3935                          EXTENT_DO_ACCOUNTING,
3936                          1, 1, &cached_state, GFP_NOFS);
3937         return 0;
3938 }
3939
3940 /*
3941  * a helper for releasepage, this tests for areas of the page that
3942  * are locked or under IO and drops the related state bits if it is safe
3943  * to drop the page.
3944  */
3945 static int try_release_extent_state(struct extent_map_tree *map,
3946                                     struct extent_io_tree *tree,
3947                                     struct page *page, gfp_t mask)
3948 {
3949         u64 start = page_offset(page);
3950         u64 end = start + PAGE_CACHE_SIZE - 1;
3951         int ret = 1;
3952
3953         if (test_range_bit(tree, start, end,
3954                            EXTENT_IOBITS, 0, NULL))
3955                 ret = 0;
3956         else {
3957                 if ((mask & GFP_NOFS) == GFP_NOFS)
3958                         mask = GFP_NOFS;
3959                 /*
3960                  * at this point we can safely clear everything except the
3961                  * locked bit and the nodatasum bit
3962                  */
3963                 ret = clear_extent_bit(tree, start, end,
3964                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3965                                  0, 0, NULL, mask);
3966
3967                 /* if clear_extent_bit failed for enomem reasons,
3968                  * we can't allow the release to continue.
3969                  */
3970                 if (ret < 0)
3971                         ret = 0;
3972                 else
3973                         ret = 1;
3974         }
3975         return ret;
3976 }
3977
3978 /*
3979  * a helper for releasepage.  As long as there are no locked extents
3980  * in the range corresponding to the page, both state records and extent
3981  * map records are removed
3982  */
3983 int try_release_extent_mapping(struct extent_map_tree *map,
3984                                struct extent_io_tree *tree, struct page *page,
3985                                gfp_t mask)
3986 {
3987         struct extent_map *em;
3988         u64 start = page_offset(page);
3989         u64 end = start + PAGE_CACHE_SIZE - 1;
3990
3991         if ((mask & __GFP_WAIT) &&
3992             page->mapping->host->i_size > 16 * 1024 * 1024) {
3993                 u64 len;
3994                 while (start <= end) {
3995                         len = end - start + 1;
3996                         write_lock(&map->lock);
3997                         em = lookup_extent_mapping(map, start, len);
3998                         if (!em) {
3999                                 write_unlock(&map->lock);
4000                                 break;
4001                         }
4002                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4003                             em->start != start) {
4004                                 write_unlock(&map->lock);
4005                                 free_extent_map(em);
4006                                 break;
4007                         }
4008                         if (!test_range_bit(tree, em->start,
4009                                             extent_map_end(em) - 1,
4010                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4011                                             0, NULL)) {
4012                                 remove_extent_mapping(map, em);
4013                                 /* once for the rb tree */
4014                                 free_extent_map(em);
4015                         }
4016                         start = extent_map_end(em);
4017                         write_unlock(&map->lock);
4018
4019                         /* once for us */
4020                         free_extent_map(em);
4021                 }
4022         }
4023         return try_release_extent_state(map, tree, page, mask);
4024 }
4025
4026 /*
4027  * helper function for fiemap, which doesn't want to see any holes.
4028  * This maps until we find something past 'last'
4029  */
4030 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4031                                                 u64 offset,
4032                                                 u64 last,
4033                                                 get_extent_t *get_extent)
4034 {
4035         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4036         struct extent_map *em;
4037         u64 len;
4038
4039         if (offset >= last)
4040                 return NULL;
4041
4042         while (1) {
4043                 len = last - offset;
4044                 if (len == 0)
4045                         break;
4046                 len = ALIGN(len, sectorsize);
4047                 em = get_extent(inode, NULL, 0, offset, len, 0);
4048                 if (IS_ERR_OR_NULL(em))
4049                         return em;
4050
4051                 /* if this isn't a hole return it */
4052                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4053                     em->block_start != EXTENT_MAP_HOLE) {
4054                         return em;
4055                 }
4056
4057                 /* this is a hole, advance to the next extent */
4058                 offset = extent_map_end(em);
4059                 free_extent_map(em);
4060                 if (offset >= last)
4061                         break;
4062         }
4063         return NULL;
4064 }
4065
4066 static noinline int count_ext_ref(u64 inum, u64 offset, u64 root_id, void *ctx)
4067 {
4068         unsigned long cnt = *((unsigned long *)ctx);
4069
4070         cnt++;
4071         *((unsigned long *)ctx) = cnt;
4072
4073         /* Now we're sure that the extent is shared. */
4074         if (cnt > 1)
4075                 return 1;
4076         return 0;
4077 }
4078
4079 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4080                 __u64 start, __u64 len, get_extent_t *get_extent)
4081 {
4082         int ret = 0;
4083         u64 off = start;
4084         u64 max = start + len;
4085         u32 flags = 0;
4086         u32 found_type;
4087         u64 last;
4088         u64 last_for_get_extent = 0;
4089         u64 disko = 0;
4090         u64 isize = i_size_read(inode);
4091         struct btrfs_key found_key;
4092         struct extent_map *em = NULL;
4093         struct extent_state *cached_state = NULL;
4094         struct btrfs_path *path;
4095         struct btrfs_file_extent_item *item;
4096         int end = 0;
4097         u64 em_start = 0;
4098         u64 em_len = 0;
4099         u64 em_end = 0;
4100         unsigned long emflags;
4101
4102         if (len == 0)
4103                 return -EINVAL;
4104
4105         path = btrfs_alloc_path();
4106         if (!path)
4107                 return -ENOMEM;
4108         path->leave_spinning = 1;
4109
4110         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
4111         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
4112
4113         /*
4114          * lookup the last file extent.  We're not using i_size here
4115          * because there might be preallocation past i_size
4116          */
4117         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
4118                                        path, btrfs_ino(inode), -1, 0);
4119         if (ret < 0) {
4120                 btrfs_free_path(path);
4121                 return ret;
4122         }
4123         WARN_ON(!ret);
4124         path->slots[0]--;
4125         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
4126                               struct btrfs_file_extent_item);
4127         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4128         found_type = btrfs_key_type(&found_key);
4129
4130         /* No extents, but there might be delalloc bits */
4131         if (found_key.objectid != btrfs_ino(inode) ||
4132             found_type != BTRFS_EXTENT_DATA_KEY) {
4133                 /* have to trust i_size as the end */
4134                 last = (u64)-1;
4135                 last_for_get_extent = isize;
4136         } else {
4137                 /*
4138                  * remember the start of the last extent.  There are a
4139                  * bunch of different factors that go into the length of the
4140                  * extent, so its much less complex to remember where it started
4141                  */
4142                 last = found_key.offset;
4143                 last_for_get_extent = last + 1;
4144         }
4145         btrfs_release_path(path);
4146
4147         /*
4148          * we might have some extents allocated but more delalloc past those
4149          * extents.  so, we trust isize unless the start of the last extent is
4150          * beyond isize
4151          */
4152         if (last < isize) {
4153                 last = (u64)-1;
4154                 last_for_get_extent = isize;
4155         }
4156
4157         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4158                          &cached_state);
4159
4160         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4161                                    get_extent);
4162         if (!em)
4163                 goto out;
4164         if (IS_ERR(em)) {
4165                 ret = PTR_ERR(em);
4166                 goto out;
4167         }
4168
4169         while (!end) {
4170                 u64 offset_in_extent = 0;
4171
4172                 /* break if the extent we found is outside the range */
4173                 if (em->start >= max || extent_map_end(em) < off)
4174                         break;
4175
4176                 /*
4177                  * get_extent may return an extent that starts before our
4178                  * requested range.  We have to make sure the ranges
4179                  * we return to fiemap always move forward and don't
4180                  * overlap, so adjust the offsets here
4181                  */
4182                 em_start = max(em->start, off);
4183
4184                 /*
4185                  * record the offset from the start of the extent
4186                  * for adjusting the disk offset below.  Only do this if the
4187                  * extent isn't compressed since our in ram offset may be past
4188                  * what we have actually allocated on disk.
4189                  */
4190                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4191                         offset_in_extent = em_start - em->start;
4192                 em_end = extent_map_end(em);
4193                 em_len = em_end - em_start;
4194                 emflags = em->flags;
4195                 disko = 0;
4196                 flags = 0;
4197
4198                 /*
4199                  * bump off for our next call to get_extent
4200                  */
4201                 off = extent_map_end(em);
4202                 if (off >= max)
4203                         end = 1;
4204
4205                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4206                         end = 1;
4207                         flags |= FIEMAP_EXTENT_LAST;
4208                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4209                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4210                                   FIEMAP_EXTENT_NOT_ALIGNED);
4211                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4212                         flags |= (FIEMAP_EXTENT_DELALLOC |
4213                                   FIEMAP_EXTENT_UNKNOWN);
4214                 } else {
4215                         unsigned long ref_cnt = 0;
4216
4217                         disko = em->block_start + offset_in_extent;
4218
4219                         /*
4220                          * As btrfs supports shared space, this information
4221                          * can be exported to userspace tools via
4222                          * flag FIEMAP_EXTENT_SHARED.
4223                          */
4224                         ret = iterate_inodes_from_logical(
4225                                         em->block_start,
4226                                         BTRFS_I(inode)->root->fs_info,
4227                                         path, count_ext_ref, &ref_cnt);
4228                         if (ret < 0 && ret != -ENOENT)
4229                                 goto out_free;
4230
4231                         if (ref_cnt > 1)
4232                                 flags |= FIEMAP_EXTENT_SHARED;
4233                 }
4234                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4235                         flags |= FIEMAP_EXTENT_ENCODED;
4236
4237                 free_extent_map(em);
4238                 em = NULL;
4239                 if ((em_start >= last) || em_len == (u64)-1 ||
4240                    (last == (u64)-1 && isize <= em_end)) {
4241                         flags |= FIEMAP_EXTENT_LAST;
4242                         end = 1;
4243                 }
4244
4245                 /* now scan forward to see if this is really the last extent. */
4246                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4247                                            get_extent);
4248                 if (IS_ERR(em)) {
4249                         ret = PTR_ERR(em);
4250                         goto out;
4251                 }
4252                 if (!em) {
4253                         flags |= FIEMAP_EXTENT_LAST;
4254                         end = 1;
4255                 }
4256                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4257                                               em_len, flags);
4258                 if (ret)
4259                         goto out_free;
4260         }
4261 out_free:
4262         free_extent_map(em);
4263 out:
4264         btrfs_free_path(path);
4265         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4266                              &cached_state, GFP_NOFS);
4267         return ret;
4268 }
4269
4270 static void __free_extent_buffer(struct extent_buffer *eb)
4271 {
4272         btrfs_leak_debug_del(&eb->leak_list);
4273         kmem_cache_free(extent_buffer_cache, eb);
4274 }
4275
4276 static int extent_buffer_under_io(struct extent_buffer *eb)
4277 {
4278         return (atomic_read(&eb->io_pages) ||
4279                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4280                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4281 }
4282
4283 /*
4284  * Helper for releasing extent buffer page.
4285  */
4286 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4287                                                 unsigned long start_idx)
4288 {
4289         unsigned long index;
4290         unsigned long num_pages;
4291         struct page *page;
4292         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4293
4294         BUG_ON(extent_buffer_under_io(eb));
4295
4296         num_pages = num_extent_pages(eb->start, eb->len);
4297         index = start_idx + num_pages;
4298         if (start_idx >= index)
4299                 return;
4300
4301         do {
4302                 index--;
4303                 page = extent_buffer_page(eb, index);
4304                 if (page && mapped) {
4305                         spin_lock(&page->mapping->private_lock);
4306                         /*
4307                          * We do this since we'll remove the pages after we've
4308                          * removed the eb from the radix tree, so we could race
4309                          * and have this page now attached to the new eb.  So
4310                          * only clear page_private if it's still connected to
4311                          * this eb.
4312                          */
4313                         if (PagePrivate(page) &&
4314                             page->private == (unsigned long)eb) {
4315                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4316                                 BUG_ON(PageDirty(page));
4317                                 BUG_ON(PageWriteback(page));
4318                                 /*
4319                                  * We need to make sure we haven't be attached
4320                                  * to a new eb.
4321                                  */
4322                                 ClearPagePrivate(page);
4323                                 set_page_private(page, 0);
4324                                 /* One for the page private */
4325                                 page_cache_release(page);
4326                         }
4327                         spin_unlock(&page->mapping->private_lock);
4328
4329                 }
4330                 if (page) {
4331                         /* One for when we alloced the page */
4332                         page_cache_release(page);
4333                 }
4334         } while (index != start_idx);
4335 }
4336
4337 /*
4338  * Helper for releasing the extent buffer.
4339  */
4340 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4341 {
4342         btrfs_release_extent_buffer_page(eb, 0);
4343         __free_extent_buffer(eb);
4344 }
4345
4346 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
4347                                                    u64 start,
4348                                                    unsigned long len,
4349                                                    gfp_t mask)
4350 {
4351         struct extent_buffer *eb = NULL;
4352
4353         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4354         if (eb == NULL)
4355                 return NULL;
4356         eb->start = start;
4357         eb->len = len;
4358         eb->tree = tree;
4359         eb->bflags = 0;
4360         rwlock_init(&eb->lock);
4361         atomic_set(&eb->write_locks, 0);
4362         atomic_set(&eb->read_locks, 0);
4363         atomic_set(&eb->blocking_readers, 0);
4364         atomic_set(&eb->blocking_writers, 0);
4365         atomic_set(&eb->spinning_readers, 0);
4366         atomic_set(&eb->spinning_writers, 0);
4367         eb->lock_nested = 0;
4368         init_waitqueue_head(&eb->write_lock_wq);
4369         init_waitqueue_head(&eb->read_lock_wq);
4370
4371         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4372
4373         spin_lock_init(&eb->refs_lock);
4374         atomic_set(&eb->refs, 1);
4375         atomic_set(&eb->io_pages, 0);
4376
4377         /*
4378          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4379          */
4380         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4381                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4382         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4383
4384         return eb;
4385 }
4386
4387 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4388 {
4389         unsigned long i;
4390         struct page *p;
4391         struct extent_buffer *new;
4392         unsigned long num_pages = num_extent_pages(src->start, src->len);
4393
4394         new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_NOFS);
4395         if (new == NULL)
4396                 return NULL;
4397
4398         for (i = 0; i < num_pages; i++) {
4399                 p = alloc_page(GFP_NOFS);
4400                 if (!p) {
4401                         btrfs_release_extent_buffer(new);
4402                         return NULL;
4403                 }
4404                 attach_extent_buffer_page(new, p);
4405                 WARN_ON(PageDirty(p));
4406                 SetPageUptodate(p);
4407                 new->pages[i] = p;
4408         }
4409
4410         copy_extent_buffer(new, src, 0, 0, src->len);
4411         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4412         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4413
4414         return new;
4415 }
4416
4417 struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4418 {
4419         struct extent_buffer *eb;
4420         unsigned long num_pages = num_extent_pages(0, len);
4421         unsigned long i;
4422
4423         eb = __alloc_extent_buffer(NULL, start, len, GFP_NOFS);
4424         if (!eb)
4425                 return NULL;
4426
4427         for (i = 0; i < num_pages; i++) {
4428                 eb->pages[i] = alloc_page(GFP_NOFS);
4429                 if (!eb->pages[i])
4430                         goto err;
4431         }
4432         set_extent_buffer_uptodate(eb);
4433         btrfs_set_header_nritems(eb, 0);
4434         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4435
4436         return eb;
4437 err:
4438         for (; i > 0; i--)
4439                 __free_page(eb->pages[i - 1]);
4440         __free_extent_buffer(eb);
4441         return NULL;
4442 }
4443
4444 static void check_buffer_tree_ref(struct extent_buffer *eb)
4445 {
4446         int refs;
4447         /* the ref bit is tricky.  We have to make sure it is set
4448          * if we have the buffer dirty.   Otherwise the
4449          * code to free a buffer can end up dropping a dirty
4450          * page
4451          *
4452          * Once the ref bit is set, it won't go away while the
4453          * buffer is dirty or in writeback, and it also won't
4454          * go away while we have the reference count on the
4455          * eb bumped.
4456          *
4457          * We can't just set the ref bit without bumping the
4458          * ref on the eb because free_extent_buffer might
4459          * see the ref bit and try to clear it.  If this happens
4460          * free_extent_buffer might end up dropping our original
4461          * ref by mistake and freeing the page before we are able
4462          * to add one more ref.
4463          *
4464          * So bump the ref count first, then set the bit.  If someone
4465          * beat us to it, drop the ref we added.
4466          */
4467         refs = atomic_read(&eb->refs);
4468         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4469                 return;
4470
4471         spin_lock(&eb->refs_lock);
4472         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4473                 atomic_inc(&eb->refs);
4474         spin_unlock(&eb->refs_lock);
4475 }
4476
4477 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4478 {
4479         unsigned long num_pages, i;
4480
4481         check_buffer_tree_ref(eb);
4482
4483         num_pages = num_extent_pages(eb->start, eb->len);
4484         for (i = 0; i < num_pages; i++) {
4485                 struct page *p = extent_buffer_page(eb, i);
4486                 mark_page_accessed(p);
4487         }
4488 }
4489
4490 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
4491                                                         u64 start)
4492 {
4493         struct extent_buffer *eb;
4494
4495         rcu_read_lock();
4496         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4497         if (eb && atomic_inc_not_zero(&eb->refs)) {
4498                 rcu_read_unlock();
4499                 mark_extent_buffer_accessed(eb);
4500                 return eb;
4501         }
4502         rcu_read_unlock();
4503
4504         return NULL;
4505 }
4506
4507 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
4508                                           u64 start, unsigned long len)
4509 {
4510         unsigned long num_pages = num_extent_pages(start, len);
4511         unsigned long i;
4512         unsigned long index = start >> PAGE_CACHE_SHIFT;
4513         struct extent_buffer *eb;
4514         struct extent_buffer *exists = NULL;
4515         struct page *p;
4516         struct address_space *mapping = tree->mapping;
4517         int uptodate = 1;
4518         int ret;
4519
4520
4521         eb = find_extent_buffer(tree, start);
4522         if (eb)
4523                 return eb;
4524
4525         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
4526         if (!eb)
4527                 return NULL;
4528
4529         for (i = 0; i < num_pages; i++, index++) {
4530                 p = find_or_create_page(mapping, index, GFP_NOFS);
4531                 if (!p)
4532                         goto free_eb;
4533
4534                 spin_lock(&mapping->private_lock);
4535                 if (PagePrivate(p)) {
4536                         /*
4537                          * We could have already allocated an eb for this page
4538                          * and attached one so lets see if we can get a ref on
4539                          * the existing eb, and if we can we know it's good and
4540                          * we can just return that one, else we know we can just
4541                          * overwrite page->private.
4542                          */
4543                         exists = (struct extent_buffer *)p->private;
4544                         if (atomic_inc_not_zero(&exists->refs)) {
4545                                 spin_unlock(&mapping->private_lock);
4546                                 unlock_page(p);
4547                                 page_cache_release(p);
4548                                 mark_extent_buffer_accessed(exists);
4549                                 goto free_eb;
4550                         }
4551
4552                         /*
4553                          * Do this so attach doesn't complain and we need to
4554                          * drop the ref the old guy had.
4555                          */
4556                         ClearPagePrivate(p);
4557                         WARN_ON(PageDirty(p));
4558                         page_cache_release(p);
4559                 }
4560                 attach_extent_buffer_page(eb, p);
4561                 spin_unlock(&mapping->private_lock);
4562                 WARN_ON(PageDirty(p));
4563                 mark_page_accessed(p);
4564                 eb->pages[i] = p;
4565                 if (!PageUptodate(p))
4566                         uptodate = 0;
4567
4568                 /*
4569                  * see below about how we avoid a nasty race with release page
4570                  * and why we unlock later
4571                  */
4572         }
4573         if (uptodate)
4574                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4575 again:
4576         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4577         if (ret)
4578                 goto free_eb;
4579
4580         spin_lock(&tree->buffer_lock);
4581         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
4582         spin_unlock(&tree->buffer_lock);
4583         radix_tree_preload_end();
4584         if (ret == -EEXIST) {
4585                 exists = find_extent_buffer(tree, start);
4586                 if (exists)
4587                         goto free_eb;
4588                 else
4589                         goto again;
4590         }
4591         /* add one reference for the tree */
4592         check_buffer_tree_ref(eb);
4593
4594         /*
4595          * there is a race where release page may have
4596          * tried to find this extent buffer in the radix
4597          * but failed.  It will tell the VM it is safe to
4598          * reclaim the, and it will clear the page private bit.
4599          * We must make sure to set the page private bit properly
4600          * after the extent buffer is in the radix tree so
4601          * it doesn't get lost
4602          */
4603         SetPageChecked(eb->pages[0]);
4604         for (i = 1; i < num_pages; i++) {
4605                 p = extent_buffer_page(eb, i);
4606                 ClearPageChecked(p);
4607                 unlock_page(p);
4608         }
4609         unlock_page(eb->pages[0]);
4610         return eb;
4611
4612 free_eb:
4613         for (i = 0; i < num_pages; i++) {
4614                 if (eb->pages[i])
4615                         unlock_page(eb->pages[i]);
4616         }
4617
4618         WARN_ON(!atomic_dec_and_test(&eb->refs));
4619         btrfs_release_extent_buffer(eb);
4620         return exists;
4621 }
4622
4623 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4624 {
4625         struct extent_buffer *eb =
4626                         container_of(head, struct extent_buffer, rcu_head);
4627
4628         __free_extent_buffer(eb);
4629 }
4630
4631 /* Expects to have eb->eb_lock already held */
4632 static int release_extent_buffer(struct extent_buffer *eb)
4633 {
4634         WARN_ON(atomic_read(&eb->refs) == 0);
4635         if (atomic_dec_and_test(&eb->refs)) {
4636                 if (test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags)) {
4637                         spin_unlock(&eb->refs_lock);
4638                 } else {
4639                         struct extent_io_tree *tree = eb->tree;
4640
4641                         spin_unlock(&eb->refs_lock);
4642
4643                         spin_lock(&tree->buffer_lock);
4644                         radix_tree_delete(&tree->buffer,
4645                                           eb->start >> PAGE_CACHE_SHIFT);
4646                         spin_unlock(&tree->buffer_lock);
4647                 }
4648
4649                 /* Should be safe to release our pages at this point */
4650                 btrfs_release_extent_buffer_page(eb, 0);
4651                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4652                 return 1;
4653         }
4654         spin_unlock(&eb->refs_lock);
4655
4656         return 0;
4657 }
4658
4659 void free_extent_buffer(struct extent_buffer *eb)
4660 {
4661         int refs;
4662         int old;
4663         if (!eb)
4664                 return;
4665
4666         while (1) {
4667                 refs = atomic_read(&eb->refs);
4668                 if (refs <= 3)
4669                         break;
4670                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
4671                 if (old == refs)
4672                         return;
4673         }
4674
4675         spin_lock(&eb->refs_lock);
4676         if (atomic_read(&eb->refs) == 2 &&
4677             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4678                 atomic_dec(&eb->refs);
4679
4680         if (atomic_read(&eb->refs) == 2 &&
4681             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4682             !extent_buffer_under_io(eb) &&
4683             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4684                 atomic_dec(&eb->refs);
4685
4686         /*
4687          * I know this is terrible, but it's temporary until we stop tracking
4688          * the uptodate bits and such for the extent buffers.
4689          */
4690         release_extent_buffer(eb);
4691 }
4692
4693 void free_extent_buffer_stale(struct extent_buffer *eb)
4694 {
4695         if (!eb)
4696                 return;
4697
4698         spin_lock(&eb->refs_lock);
4699         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4700
4701         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4702             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4703                 atomic_dec(&eb->refs);
4704         release_extent_buffer(eb);
4705 }
4706
4707 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4708 {
4709         unsigned long i;
4710         unsigned long num_pages;
4711         struct page *page;
4712
4713         num_pages = num_extent_pages(eb->start, eb->len);
4714
4715         for (i = 0; i < num_pages; i++) {
4716                 page = extent_buffer_page(eb, i);
4717                 if (!PageDirty(page))
4718                         continue;
4719
4720                 lock_page(page);
4721                 WARN_ON(!PagePrivate(page));
4722
4723                 clear_page_dirty_for_io(page);
4724                 spin_lock_irq(&page->mapping->tree_lock);
4725                 if (!PageDirty(page)) {
4726                         radix_tree_tag_clear(&page->mapping->page_tree,
4727                                                 page_index(page),
4728                                                 PAGECACHE_TAG_DIRTY);
4729                 }
4730                 spin_unlock_irq(&page->mapping->tree_lock);
4731                 ClearPageError(page);
4732                 unlock_page(page);
4733         }
4734         WARN_ON(atomic_read(&eb->refs) == 0);
4735 }
4736
4737 int set_extent_buffer_dirty(struct extent_buffer *eb)
4738 {
4739         unsigned long i;
4740         unsigned long num_pages;
4741         int was_dirty = 0;
4742
4743         check_buffer_tree_ref(eb);
4744
4745         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4746
4747         num_pages = num_extent_pages(eb->start, eb->len);
4748         WARN_ON(atomic_read(&eb->refs) == 0);
4749         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4750
4751         for (i = 0; i < num_pages; i++)
4752                 set_page_dirty(extent_buffer_page(eb, i));
4753         return was_dirty;
4754 }
4755
4756 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4757 {
4758         unsigned long i;
4759         struct page *page;
4760         unsigned long num_pages;
4761
4762         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4763         num_pages = num_extent_pages(eb->start, eb->len);
4764         for (i = 0; i < num_pages; i++) {
4765                 page = extent_buffer_page(eb, i);
4766                 if (page)
4767                         ClearPageUptodate(page);
4768         }
4769         return 0;
4770 }
4771
4772 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4773 {
4774         unsigned long i;
4775         struct page *page;
4776         unsigned long num_pages;
4777
4778         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4779         num_pages = num_extent_pages(eb->start, eb->len);
4780         for (i = 0; i < num_pages; i++) {
4781                 page = extent_buffer_page(eb, i);
4782                 SetPageUptodate(page);
4783         }
4784         return 0;
4785 }
4786
4787 int extent_buffer_uptodate(struct extent_buffer *eb)
4788 {
4789         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4790 }
4791
4792 int read_extent_buffer_pages(struct extent_io_tree *tree,
4793                              struct extent_buffer *eb, u64 start, int wait,
4794                              get_extent_t *get_extent, int mirror_num)
4795 {
4796         unsigned long i;
4797         unsigned long start_i;
4798         struct page *page;
4799         int err;
4800         int ret = 0;
4801         int locked_pages = 0;
4802         int all_uptodate = 1;
4803         unsigned long num_pages;
4804         unsigned long num_reads = 0;
4805         struct bio *bio = NULL;
4806         unsigned long bio_flags = 0;
4807
4808         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4809                 return 0;
4810
4811         if (start) {
4812                 WARN_ON(start < eb->start);
4813                 start_i = (start >> PAGE_CACHE_SHIFT) -
4814                         (eb->start >> PAGE_CACHE_SHIFT);
4815         } else {
4816                 start_i = 0;
4817         }
4818
4819         num_pages = num_extent_pages(eb->start, eb->len);
4820         for (i = start_i; i < num_pages; i++) {
4821                 page = extent_buffer_page(eb, i);
4822                 if (wait == WAIT_NONE) {
4823                         if (!trylock_page(page))
4824                                 goto unlock_exit;
4825                 } else {
4826                         lock_page(page);
4827                 }
4828                 locked_pages++;
4829                 if (!PageUptodate(page)) {
4830                         num_reads++;
4831                         all_uptodate = 0;
4832                 }
4833         }
4834         if (all_uptodate) {
4835                 if (start_i == 0)
4836                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4837                 goto unlock_exit;
4838         }
4839
4840         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4841         eb->read_mirror = 0;
4842         atomic_set(&eb->io_pages, num_reads);
4843         for (i = start_i; i < num_pages; i++) {
4844                 page = extent_buffer_page(eb, i);
4845                 if (!PageUptodate(page)) {
4846                         ClearPageError(page);
4847                         err = __extent_read_full_page(tree, page,
4848                                                       get_extent, &bio,
4849                                                       mirror_num, &bio_flags,
4850                                                       READ | REQ_META);
4851                         if (err)
4852                                 ret = err;
4853                 } else {
4854                         unlock_page(page);
4855                 }
4856         }
4857
4858         if (bio) {
4859                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
4860                                      bio_flags);
4861                 if (err)
4862                         return err;
4863         }
4864
4865         if (ret || wait != WAIT_COMPLETE)
4866                 return ret;
4867
4868         for (i = start_i; i < num_pages; i++) {
4869                 page = extent_buffer_page(eb, i);
4870                 wait_on_page_locked(page);
4871                 if (!PageUptodate(page))
4872                         ret = -EIO;
4873         }
4874
4875         return ret;
4876
4877 unlock_exit:
4878         i = start_i;
4879         while (locked_pages > 0) {
4880                 page = extent_buffer_page(eb, i);
4881                 i++;
4882                 unlock_page(page);
4883                 locked_pages--;
4884         }
4885         return ret;
4886 }
4887
4888 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4889                         unsigned long start,
4890                         unsigned long len)
4891 {
4892         size_t cur;
4893         size_t offset;
4894         struct page *page;
4895         char *kaddr;
4896         char *dst = (char *)dstv;
4897         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4898         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4899
4900         WARN_ON(start > eb->len);
4901         WARN_ON(start + len > eb->start + eb->len);
4902
4903         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
4904
4905         while (len > 0) {
4906                 page = extent_buffer_page(eb, i);
4907
4908                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4909                 kaddr = page_address(page);
4910                 memcpy(dst, kaddr + offset, cur);
4911
4912                 dst += cur;
4913                 len -= cur;
4914                 offset = 0;
4915                 i++;
4916         }
4917 }
4918
4919 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4920                                unsigned long min_len, char **map,
4921                                unsigned long *map_start,
4922                                unsigned long *map_len)
4923 {
4924         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4925         char *kaddr;
4926         struct page *p;
4927         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4928         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4929         unsigned long end_i = (start_offset + start + min_len - 1) >>
4930                 PAGE_CACHE_SHIFT;
4931
4932         if (i != end_i)
4933                 return -EINVAL;
4934
4935         if (i == 0) {
4936                 offset = start_offset;
4937                 *map_start = 0;
4938         } else {
4939                 offset = 0;
4940                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4941         }
4942
4943         if (start + min_len > eb->len) {
4944                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4945                        "wanted %lu %lu\n",
4946                        eb->start, eb->len, start, min_len);
4947                 return -EINVAL;
4948         }
4949
4950         p = extent_buffer_page(eb, i);
4951         kaddr = page_address(p);
4952         *map = kaddr + offset;
4953         *map_len = PAGE_CACHE_SIZE - offset;
4954         return 0;
4955 }
4956
4957 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4958                           unsigned long start,
4959                           unsigned long len)
4960 {
4961         size_t cur;
4962         size_t offset;
4963         struct page *page;
4964         char *kaddr;
4965         char *ptr = (char *)ptrv;
4966         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4967         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4968         int ret = 0;
4969
4970         WARN_ON(start > eb->len);
4971         WARN_ON(start + len > eb->start + eb->len);
4972
4973         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
4974
4975         while (len > 0) {
4976                 page = extent_buffer_page(eb, i);
4977
4978                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4979
4980                 kaddr = page_address(page);
4981                 ret = memcmp(ptr, kaddr + offset, cur);
4982                 if (ret)
4983                         break;
4984
4985                 ptr += cur;
4986                 len -= cur;
4987                 offset = 0;
4988                 i++;
4989         }
4990         return ret;
4991 }
4992
4993 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4994                          unsigned long start, unsigned long len)
4995 {
4996         size_t cur;
4997         size_t offset;
4998         struct page *page;
4999         char *kaddr;
5000         char *src = (char *)srcv;
5001         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5002         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5003
5004         WARN_ON(start > eb->len);
5005         WARN_ON(start + len > eb->start + eb->len);
5006
5007         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5008
5009         while (len > 0) {
5010                 page = extent_buffer_page(eb, i);
5011                 WARN_ON(!PageUptodate(page));
5012
5013                 cur = min(len, PAGE_CACHE_SIZE - offset);
5014                 kaddr = page_address(page);
5015                 memcpy(kaddr + offset, src, cur);
5016
5017                 src += cur;
5018                 len -= cur;
5019                 offset = 0;
5020                 i++;
5021         }
5022 }
5023
5024 void memset_extent_buffer(struct extent_buffer *eb, char c,
5025                           unsigned long start, unsigned long len)
5026 {
5027         size_t cur;
5028         size_t offset;
5029         struct page *page;
5030         char *kaddr;
5031         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5032         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5033
5034         WARN_ON(start > eb->len);
5035         WARN_ON(start + len > eb->start + eb->len);
5036
5037         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5038
5039         while (len > 0) {
5040                 page = extent_buffer_page(eb, i);
5041                 WARN_ON(!PageUptodate(page));
5042
5043                 cur = min(len, PAGE_CACHE_SIZE - offset);
5044                 kaddr = page_address(page);
5045                 memset(kaddr + offset, c, cur);
5046
5047                 len -= cur;
5048                 offset = 0;
5049                 i++;
5050         }
5051 }
5052
5053 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5054                         unsigned long dst_offset, unsigned long src_offset,
5055                         unsigned long len)
5056 {
5057         u64 dst_len = dst->len;
5058         size_t cur;
5059         size_t offset;
5060         struct page *page;
5061         char *kaddr;
5062         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5063         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5064
5065         WARN_ON(src->len != dst_len);
5066
5067         offset = (start_offset + dst_offset) &
5068                 (PAGE_CACHE_SIZE - 1);
5069
5070         while (len > 0) {
5071                 page = extent_buffer_page(dst, i);
5072                 WARN_ON(!PageUptodate(page));
5073
5074                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5075
5076                 kaddr = page_address(page);
5077                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5078
5079                 src_offset += cur;
5080                 len -= cur;
5081                 offset = 0;
5082                 i++;
5083         }
5084 }
5085
5086 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5087 {
5088         unsigned long distance = (src > dst) ? src - dst : dst - src;
5089         return distance < len;
5090 }
5091
5092 static void copy_pages(struct page *dst_page, struct page *src_page,
5093                        unsigned long dst_off, unsigned long src_off,
5094                        unsigned long len)
5095 {
5096         char *dst_kaddr = page_address(dst_page);
5097         char *src_kaddr;
5098         int must_memmove = 0;
5099
5100         if (dst_page != src_page) {
5101                 src_kaddr = page_address(src_page);
5102         } else {
5103                 src_kaddr = dst_kaddr;
5104                 if (areas_overlap(src_off, dst_off, len))
5105                         must_memmove = 1;
5106         }
5107
5108         if (must_memmove)
5109                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5110         else
5111                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5112 }
5113
5114 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5115                            unsigned long src_offset, unsigned long len)
5116 {
5117         size_t cur;
5118         size_t dst_off_in_page;
5119         size_t src_off_in_page;
5120         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5121         unsigned long dst_i;
5122         unsigned long src_i;
5123
5124         if (src_offset + len > dst->len) {
5125                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
5126                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5127                 BUG_ON(1);
5128         }
5129         if (dst_offset + len > dst->len) {
5130                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
5131                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5132                 BUG_ON(1);
5133         }
5134
5135         while (len > 0) {
5136                 dst_off_in_page = (start_offset + dst_offset) &
5137                         (PAGE_CACHE_SIZE - 1);
5138                 src_off_in_page = (start_offset + src_offset) &
5139                         (PAGE_CACHE_SIZE - 1);
5140
5141                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5142                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5143
5144                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5145                                                src_off_in_page));
5146                 cur = min_t(unsigned long, cur,
5147                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5148
5149                 copy_pages(extent_buffer_page(dst, dst_i),
5150                            extent_buffer_page(dst, src_i),
5151                            dst_off_in_page, src_off_in_page, cur);
5152
5153                 src_offset += cur;
5154                 dst_offset += cur;
5155                 len -= cur;
5156         }
5157 }
5158
5159 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5160                            unsigned long src_offset, unsigned long len)
5161 {
5162         size_t cur;
5163         size_t dst_off_in_page;
5164         size_t src_off_in_page;
5165         unsigned long dst_end = dst_offset + len - 1;
5166         unsigned long src_end = src_offset + len - 1;
5167         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5168         unsigned long dst_i;
5169         unsigned long src_i;
5170
5171         if (src_offset + len > dst->len) {
5172                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
5173                        "len %lu len %lu\n", src_offset, len, dst->len);
5174                 BUG_ON(1);
5175         }
5176         if (dst_offset + len > dst->len) {
5177                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
5178                        "len %lu len %lu\n", dst_offset, len, dst->len);
5179                 BUG_ON(1);
5180         }
5181         if (dst_offset < src_offset) {
5182                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5183                 return;
5184         }
5185         while (len > 0) {
5186                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5187                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5188
5189                 dst_off_in_page = (start_offset + dst_end) &
5190                         (PAGE_CACHE_SIZE - 1);
5191                 src_off_in_page = (start_offset + src_end) &
5192                         (PAGE_CACHE_SIZE - 1);
5193
5194                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5195                 cur = min(cur, dst_off_in_page + 1);
5196                 copy_pages(extent_buffer_page(dst, dst_i),
5197                            extent_buffer_page(dst, src_i),
5198                            dst_off_in_page - cur + 1,
5199                            src_off_in_page - cur + 1, cur);
5200
5201                 dst_end -= cur;
5202                 src_end -= cur;
5203                 len -= cur;
5204         }
5205 }
5206
5207 int try_release_extent_buffer(struct page *page)
5208 {
5209         struct extent_buffer *eb;
5210
5211         /*
5212          * We need to make sure noboody is attaching this page to an eb right
5213          * now.
5214          */
5215         spin_lock(&page->mapping->private_lock);
5216         if (!PagePrivate(page)) {
5217                 spin_unlock(&page->mapping->private_lock);
5218                 return 1;
5219         }
5220
5221         eb = (struct extent_buffer *)page->private;
5222         BUG_ON(!eb);
5223
5224         /*
5225          * This is a little awful but should be ok, we need to make sure that
5226          * the eb doesn't disappear out from under us while we're looking at
5227          * this page.
5228          */
5229         spin_lock(&eb->refs_lock);
5230         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5231                 spin_unlock(&eb->refs_lock);
5232                 spin_unlock(&page->mapping->private_lock);
5233                 return 0;
5234         }
5235         spin_unlock(&page->mapping->private_lock);
5236
5237         /*
5238          * If tree ref isn't set then we know the ref on this eb is a real ref,
5239          * so just return, this page will likely be freed soon anyway.
5240          */
5241         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5242                 spin_unlock(&eb->refs_lock);
5243                 return 0;
5244         }
5245
5246         return release_extent_buffer(eb);
5247 }