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