fs/btrfs/extent_io.c

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