fs/btrfs/extent_io.c

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