fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26 static struct bio_set *btrfs_bioset;
  27
  28 static inline bool extent_state_in_tree(const struct extent_state *state)
  29 {
  30         return !RB_EMPTY_NODE(&state->rb_node);
  31 }
  32
  33 #ifdef CONFIG_BTRFS_DEBUG
  34 static LIST_HEAD(buffers);
  35 static LIST_HEAD(states);
  36
  37 static DEFINE_SPINLOCK(leak_lock);
  38
  39 static inline
  40 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  41 {
  42         unsigned long flags;
  43
  44         spin_lock_irqsave(&leak_lock, flags);
  45         list_add(new, head);
  46         spin_unlock_irqrestore(&leak_lock, flags);
  47 }
  48
  49 static inline
  50 void btrfs_leak_debug_del(struct list_head *entry)
  51 {
  52         unsigned long flags;
  53
  54         spin_lock_irqsave(&leak_lock, flags);
  55         list_del(entry);
  56         spin_unlock_irqrestore(&leak_lock, flags);
  57 }
  58
  59 static inline
  60 void btrfs_leak_debug_check(void)
  61 {
  62         struct extent_state *state;
  63         struct extent_buffer *eb;
  64
  65         while (!list_empty(&states)) {
  66                 state = list_entry(states.next, struct extent_state, leak_list);
  67                 pr_err("BTRFS: state leak: start %llu end %llu state %lu in tree %d refs %d\n",
  68                        state->start, state->end, state->state,
  69                        extent_state_in_tree(state),
  70                        atomic_read(&state->refs));
  71                 list_del(&state->leak_list);
  72                 kmem_cache_free(extent_state_cache, state);
  73         }
  74
  75         while (!list_empty(&buffers)) {
  76                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  77                 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu "
  78                        "refs %d\n",
  79                        eb->start, eb->len, atomic_read(&eb->refs));
  80                 list_del(&eb->leak_list);
  81                 kmem_cache_free(extent_buffer_cache, eb);
  82         }
  83 }
  84
  85 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  86         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  87 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  88                 struct extent_io_tree *tree, u64 start, u64 end)
  89 {
  90         struct inode *inode;
  91         u64 isize;
  92
  93         if (!tree->mapping)
  94                 return;
  95
  96         inode = tree->mapping->host;
  97         isize = i_size_read(inode);
  98         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  99                 printk_ratelimited(KERN_DEBUG
 100                     "BTRFS: %s: ino %llu isize %llu odd range [%llu,%llu]\n",
 101                                 caller, btrfs_ino(inode), isize, start, end);
 102         }
 103 }
 104 #else
 105 #define btrfs_leak_debug_add(new, head) do {} while (0)
 106 #define btrfs_leak_debug_del(entry)     do {} while (0)
 107 #define btrfs_leak_debug_check()        do {} while (0)
 108 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 109 #endif
 110
 111 #define BUFFER_LRU_MAX 64
 112
 113 struct tree_entry {
 114         u64 start;
 115         u64 end;
 116         struct rb_node rb_node;
 117 };
 118
 119 struct extent_page_data {
 120         struct bio *bio;
 121         struct extent_io_tree *tree;
 122         get_extent_t *get_extent;
 123         unsigned long bio_flags;
 124
 125         /* tells writepage not to lock the state bits for this range
 126          * it still does the unlocking
 127          */
 128         unsigned int extent_locked:1;
 129
 130         /* tells the submit_bio code to use a WRITE_SYNC */
 131         unsigned int sync_io:1;
 132 };
 133
 134 static noinline void flush_write_bio(void *data);
 135 static inline struct btrfs_fs_info *
 136 tree_fs_info(struct extent_io_tree *tree)
 137 {
 138         if (!tree->mapping)
 139                 return NULL;
 140         return btrfs_sb(tree->mapping->host->i_sb);
 141 }
 142
 143 int __init extent_io_init(void)
 144 {
 145         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 146                         sizeof(struct extent_state), 0,
 147                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 148         if (!extent_state_cache)
 149                 return -ENOMEM;
 150
 151         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 152                         sizeof(struct extent_buffer), 0,
 153                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 154         if (!extent_buffer_cache)
 155                 goto free_state_cache;
 156
 157         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 158                                      offsetof(struct btrfs_io_bio, bio));
 159         if (!btrfs_bioset)
 160                 goto free_buffer_cache;
 161
 162         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 163                 goto free_bioset;
 164
 165         return 0;
 166
 167 free_bioset:
 168         bioset_free(btrfs_bioset);
 169         btrfs_bioset = NULL;
 170
 171 free_buffer_cache:
 172         kmem_cache_destroy(extent_buffer_cache);
 173         extent_buffer_cache = NULL;
 174
 175 free_state_cache:
 176         kmem_cache_destroy(extent_state_cache);
 177         extent_state_cache = NULL;
 178         return -ENOMEM;
 179 }
 180
 181 void extent_io_exit(void)
 182 {
 183         btrfs_leak_debug_check();
 184
 185         /*
 186          * Make sure all delayed rcu free are flushed before we
 187          * destroy caches.
 188          */
 189         rcu_barrier();
 190         if (extent_state_cache)
 191                 kmem_cache_destroy(extent_state_cache);
 192         if (extent_buffer_cache)
 193                 kmem_cache_destroy(extent_buffer_cache);
 194         if (btrfs_bioset)
 195                 bioset_free(btrfs_bioset);
 196 }
 197
 198 void extent_io_tree_init(struct extent_io_tree *tree,
 199                          struct address_space *mapping)
 200 {
 201         tree->state = RB_ROOT;
 202         tree->ops = NULL;
 203         tree->dirty_bytes = 0;
 204         spin_lock_init(&tree->lock);
 205         tree->mapping = mapping;
 206 }
 207
 208 static struct extent_state *alloc_extent_state(gfp_t mask)
 209 {
 210         struct extent_state *state;
 211
 212         state = kmem_cache_alloc(extent_state_cache, mask);
 213         if (!state)
 214                 return state;
 215         state->state = 0;
 216         state->private = 0;
 217         RB_CLEAR_NODE(&state->rb_node);
 218         btrfs_leak_debug_add(&state->leak_list, &states);
 219         atomic_set(&state->refs, 1);
 220         init_waitqueue_head(&state->wq);
 221         trace_alloc_extent_state(state, mask, _RET_IP_);
 222         return state;
 223 }
 224
 225 void free_extent_state(struct extent_state *state)
 226 {
 227         if (!state)
 228                 return;
 229         if (atomic_dec_and_test(&state->refs)) {
 230                 WARN_ON(extent_state_in_tree(state));
 231                 btrfs_leak_debug_del(&state->leak_list);
 232                 trace_free_extent_state(state, _RET_IP_);
 233                 kmem_cache_free(extent_state_cache, state);
 234         }
 235 }
 236
 237 static struct rb_node *tree_insert(struct rb_root *root,
 238                                    struct rb_node *search_start,
 239                                    u64 offset,
 240                                    struct rb_node *node,
 241                                    struct rb_node ***p_in,
 242                                    struct rb_node **parent_in)
 243 {
 244         struct rb_node **p;
 245         struct rb_node *parent = NULL;
 246         struct tree_entry *entry;
 247
 248         if (p_in && parent_in) {
 249                 p = *p_in;
 250                 parent = *parent_in;
 251                 goto do_insert;
 252         }
 253
 254         p = search_start ? &search_start : &root->rb_node;
 255         while (*p) {
 256                 parent = *p;
 257                 entry = rb_entry(parent, struct tree_entry, rb_node);
 258
 259                 if (offset < entry->start)
 260                         p = &(*p)->rb_left;
 261                 else if (offset > entry->end)
 262                         p = &(*p)->rb_right;
 263                 else
 264                         return parent;
 265         }
 266
 267 do_insert:
 268         rb_link_node(node, parent, p);
 269         rb_insert_color(node, root);
 270         return NULL;
 271 }
 272
 273 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 274                                       struct rb_node **prev_ret,
 275                                       struct rb_node **next_ret,
 276                                       struct rb_node ***p_ret,
 277                                       struct rb_node **parent_ret)
 278 {
 279         struct rb_root *root = &tree->state;
 280         struct rb_node **n = &root->rb_node;
 281         struct rb_node *prev = NULL;
 282         struct rb_node *orig_prev = NULL;
 283         struct tree_entry *entry;
 284         struct tree_entry *prev_entry = NULL;
 285
 286         while (*n) {
 287                 prev = *n;
 288                 entry = rb_entry(prev, struct tree_entry, rb_node);
 289                 prev_entry = entry;
 290
 291                 if (offset < entry->start)
 292                         n = &(*n)->rb_left;
 293                 else if (offset > entry->end)
 294                         n = &(*n)->rb_right;
 295                 else
 296                         return *n;
 297         }
 298
 299         if (p_ret)
 300                 *p_ret = n;
 301         if (parent_ret)
 302                 *parent_ret = prev;
 303
 304         if (prev_ret) {
 305                 orig_prev = prev;
 306                 while (prev && offset > prev_entry->end) {
 307                         prev = rb_next(prev);
 308                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 309                 }
 310                 *prev_ret = prev;
 311                 prev = orig_prev;
 312         }
 313
 314         if (next_ret) {
 315                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 316                 while (prev && offset < prev_entry->start) {
 317                         prev = rb_prev(prev);
 318                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 319                 }
 320                 *next_ret = prev;
 321         }
 322         return NULL;
 323 }
 324
 325 static inline struct rb_node *
 326 tree_search_for_insert(struct extent_io_tree *tree,
 327                        u64 offset,
 328                        struct rb_node ***p_ret,
 329                        struct rb_node **parent_ret)
 330 {
 331         struct rb_node *prev = NULL;
 332         struct rb_node *ret;
 333
 334         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 335         if (!ret)
 336                 return prev;
 337         return ret;
 338 }
 339
 340 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 341                                           u64 offset)
 342 {
 343         return tree_search_for_insert(tree, offset, NULL, NULL);
 344 }
 345
 346 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 347                      struct extent_state *other)
 348 {
 349         if (tree->ops && tree->ops->merge_extent_hook)
 350                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 351                                              other);
 352 }
 353
 354 /*
 355  * utility function to look for merge candidates inside a given range.
 356  * Any extents with matching state are merged together into a single
 357  * extent in the tree.  Extents with EXTENT_IO in their state field
 358  * are not merged because the end_io handlers need to be able to do
 359  * operations on them without sleeping (or doing allocations/splits).
 360  *
 361  * This should be called with the tree lock held.
 362  */
 363 static void merge_state(struct extent_io_tree *tree,
 364                         struct extent_state *state)
 365 {
 366         struct extent_state *other;
 367         struct rb_node *other_node;
 368
 369         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 370                 return;
 371
 372         other_node = rb_prev(&state->rb_node);
 373         if (other_node) {
 374                 other = rb_entry(other_node, struct extent_state, rb_node);
 375                 if (other->end == state->start - 1 &&
 376                     other->state == state->state) {
 377                         merge_cb(tree, state, other);
 378                         state->start = other->start;
 379                         rb_erase(&other->rb_node, &tree->state);
 380                         RB_CLEAR_NODE(&other->rb_node);
 381                         free_extent_state(other);
 382                 }
 383         }
 384         other_node = rb_next(&state->rb_node);
 385         if (other_node) {
 386                 other = rb_entry(other_node, struct extent_state, rb_node);
 387                 if (other->start == state->end + 1 &&
 388                     other->state == state->state) {
 389                         merge_cb(tree, state, other);
 390                         state->end = other->end;
 391                         rb_erase(&other->rb_node, &tree->state);
 392                         RB_CLEAR_NODE(&other->rb_node);
 393                         free_extent_state(other);
 394                 }
 395         }
 396 }
 397
 398 static void set_state_cb(struct extent_io_tree *tree,
 399                          struct extent_state *state, unsigned long *bits)
 400 {
 401         if (tree->ops && tree->ops->set_bit_hook)
 402                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 403 }
 404
 405 static void clear_state_cb(struct extent_io_tree *tree,
 406                            struct extent_state *state, unsigned long *bits)
 407 {
 408         if (tree->ops && tree->ops->clear_bit_hook)
 409                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 410 }
 411
 412 static void set_state_bits(struct extent_io_tree *tree,
 413                            struct extent_state *state, unsigned long *bits);
 414
 415 /*
 416  * insert an extent_state struct into the tree.  'bits' are set on the
 417  * struct before it is inserted.
 418  *
 419  * This may return -EEXIST if the extent is already there, in which case the
 420  * state struct is freed.
 421  *
 422  * The tree lock is not taken internally.  This is a utility function and
 423  * probably isn't what you want to call (see set/clear_extent_bit).
 424  */
 425 static int insert_state(struct extent_io_tree *tree,
 426                         struct extent_state *state, u64 start, u64 end,
 427                         struct rb_node ***p,
 428                         struct rb_node **parent,
 429                         unsigned long *bits)
 430 {
 431         struct rb_node *node;
 432
 433         if (end < start)
 434                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 435                        end, start);
 436         state->start = start;
 437         state->end = end;
 438
 439         set_state_bits(tree, state, bits);
 440
 441         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 442         if (node) {
 443                 struct extent_state *found;
 444                 found = rb_entry(node, struct extent_state, rb_node);
 445                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
 446                        "%llu %llu\n",
 447                        found->start, found->end, start, end);
 448                 return -EEXIST;
 449         }
 450         merge_state(tree, state);
 451         return 0;
 452 }
 453
 454 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 455                      u64 split)
 456 {
 457         if (tree->ops && tree->ops->split_extent_hook)
 458                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 459 }
 460
 461 /*
 462  * split a given extent state struct in two, inserting the preallocated
 463  * struct 'prealloc' as the newly created second half.  'split' indicates an
 464  * offset inside 'orig' where it should be split.
 465  *
 466  * Before calling,
 467  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 468  * are two extent state structs in the tree:
 469  * prealloc: [orig->start, split - 1]
 470  * orig: [ split, orig->end ]
 471  *
 472  * The tree locks are not taken by this function. They need to be held
 473  * by the caller.
 474  */
 475 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 476                        struct extent_state *prealloc, u64 split)
 477 {
 478         struct rb_node *node;
 479
 480         split_cb(tree, orig, split);
 481
 482         prealloc->start = orig->start;
 483         prealloc->end = split - 1;
 484         prealloc->state = orig->state;
 485         orig->start = split;
 486
 487         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 488                            &prealloc->rb_node, NULL, NULL);
 489         if (node) {
 490                 free_extent_state(prealloc);
 491                 return -EEXIST;
 492         }
 493         return 0;
 494 }
 495
 496 static struct extent_state *next_state(struct extent_state *state)
 497 {
 498         struct rb_node *next = rb_next(&state->rb_node);
 499         if (next)
 500                 return rb_entry(next, struct extent_state, rb_node);
 501         else
 502                 return NULL;
 503 }
 504
 505 /*
 506  * utility function to clear some bits in an extent state struct.
 507  * it will optionally wake up any one waiting on this state (wake == 1).
 508  *
 509  * If no bits are set on the state struct after clearing things, the
 510  * struct is freed and removed from the tree
 511  */
 512 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 513                                             struct extent_state *state,
 514                                             unsigned long *bits, int wake)
 515 {
 516         struct extent_state *next;
 517         unsigned long bits_to_clear = *bits & ~EXTENT_CTLBITS;
 518
 519         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 520                 u64 range = state->end - state->start + 1;
 521                 WARN_ON(range > tree->dirty_bytes);
 522                 tree->dirty_bytes -= range;
 523         }
 524         clear_state_cb(tree, state, bits);
 525         state->state &= ~bits_to_clear;
 526         if (wake)
 527                 wake_up(&state->wq);
 528         if (state->state == 0) {
 529                 next = next_state(state);
 530                 if (extent_state_in_tree(state)) {
 531                         rb_erase(&state->rb_node, &tree->state);
 532                         RB_CLEAR_NODE(&state->rb_node);
 533                         free_extent_state(state);
 534                 } else {
 535                         WARN_ON(1);
 536                 }
 537         } else {
 538                 merge_state(tree, state);
 539                 next = next_state(state);
 540         }
 541         return next;
 542 }
 543
 544 static struct extent_state *
 545 alloc_extent_state_atomic(struct extent_state *prealloc)
 546 {
 547         if (!prealloc)
 548                 prealloc = alloc_extent_state(GFP_ATOMIC);
 549
 550         return prealloc;
 551 }
 552
 553 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 554 {
 555         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 556                     "Extent tree was modified by another "
 557                     "thread while locked.");
 558 }
 559
 560 /*
 561  * clear some bits on a range in the tree.  This may require splitting
 562  * or inserting elements in the tree, so the gfp mask is used to
 563  * indicate which allocations or sleeping are allowed.
 564  *
 565  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 566  * the given range from the tree regardless of state (ie for truncate).
 567  *
 568  * the range [start, end] is inclusive.
 569  *
 570  * This takes the tree lock, and returns 0 on success and < 0 on error.
 571  */
 572 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 573                      unsigned long bits, int wake, int delete,
 574                      struct extent_state **cached_state,
 575                      gfp_t mask)
 576 {
 577         struct extent_state *state;
 578         struct extent_state *cached;
 579         struct extent_state *prealloc = NULL;
 580         struct rb_node *node;
 581         u64 last_end;
 582         int err;
 583         int clear = 0;
 584
 585         btrfs_debug_check_extent_io_range(tree, start, end);
 586
 587         if (bits & EXTENT_DELALLOC)
 588                 bits |= EXTENT_NORESERVE;
 589
 590         if (delete)
 591                 bits |= ~EXTENT_CTLBITS;
 592         bits |= EXTENT_FIRST_DELALLOC;
 593
 594         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 595                 clear = 1;
 596 again:
 597         if (!prealloc && (mask & __GFP_WAIT)) {
 598                 prealloc = alloc_extent_state(mask);
 599                 if (!prealloc)
 600                         return -ENOMEM;
 601         }
 602
 603         spin_lock(&tree->lock);
 604         if (cached_state) {
 605                 cached = *cached_state;
 606
 607                 if (clear) {
 608                         *cached_state = NULL;
 609                         cached_state = NULL;
 610                 }
 611
 612                 if (cached && extent_state_in_tree(cached) &&
 613                     cached->start <= start && cached->end > start) {
 614                         if (clear)
 615                                 atomic_dec(&cached->refs);
 616                         state = cached;
 617                         goto hit_next;
 618                 }
 619                 if (clear)
 620                         free_extent_state(cached);
 621         }
 622         /*
 623          * this search will find the extents that end after
 624          * our range starts
 625          */
 626         node = tree_search(tree, start);
 627         if (!node)
 628                 goto out;
 629         state = rb_entry(node, struct extent_state, rb_node);
 630 hit_next:
 631         if (state->start > end)
 632                 goto out;
 633         WARN_ON(state->end < start);
 634         last_end = state->end;
 635
 636         /* the state doesn't have the wanted bits, go ahead */
 637         if (!(state->state & bits)) {
 638                 state = next_state(state);
 639                 goto next;
 640         }
 641
 642         /*
 643          *     | ---- desired range ---- |
 644          *  | state | or
 645          *  | ------------- state -------------- |
 646          *
 647          * We need to split the extent we found, and may flip
 648          * bits on second half.
 649          *
 650          * If the extent we found extends past our range, we
 651          * just split and search again.  It'll get split again
 652          * the next time though.
 653          *
 654          * If the extent we found is inside our range, we clear
 655          * the desired bit on it.
 656          */
 657
 658         if (state->start < start) {
 659                 prealloc = alloc_extent_state_atomic(prealloc);
 660                 BUG_ON(!prealloc);
 661                 err = split_state(tree, state, prealloc, start);
 662                 if (err)
 663                         extent_io_tree_panic(tree, err);
 664
 665                 prealloc = NULL;
 666                 if (err)
 667                         goto out;
 668                 if (state->end <= end) {
 669                         state = clear_state_bit(tree, state, &bits, wake);
 670                         goto next;
 671                 }
 672                 goto search_again;
 673         }
 674         /*
 675          * | ---- desired range ---- |
 676          *                        | state |
 677          * We need to split the extent, and clear the bit
 678          * on the first half
 679          */
 680         if (state->start <= end && state->end > end) {
 681                 prealloc = alloc_extent_state_atomic(prealloc);
 682                 BUG_ON(!prealloc);
 683                 err = split_state(tree, state, prealloc, end + 1);
 684                 if (err)
 685                         extent_io_tree_panic(tree, err);
 686
 687                 if (wake)
 688                         wake_up(&state->wq);
 689
 690                 clear_state_bit(tree, prealloc, &bits, wake);
 691
 692                 prealloc = NULL;
 693                 goto out;
 694         }
 695
 696         state = clear_state_bit(tree, state, &bits, wake);
 697 next:
 698         if (last_end == (u64)-1)
 699                 goto out;
 700         start = last_end + 1;
 701         if (start <= end && state && !need_resched())
 702                 goto hit_next;
 703         goto search_again;
 704
 705 out:
 706         spin_unlock(&tree->lock);
 707         if (prealloc)
 708                 free_extent_state(prealloc);
 709
 710         return 0;
 711
 712 search_again:
 713         if (start > end)
 714                 goto out;
 715         spin_unlock(&tree->lock);
 716         if (mask & __GFP_WAIT)
 717                 cond_resched();
 718         goto again;
 719 }
 720
 721 static void wait_on_state(struct extent_io_tree *tree,
 722                           struct extent_state *state)
 723                 __releases(tree->lock)
 724                 __acquires(tree->lock)
 725 {
 726         DEFINE_WAIT(wait);
 727         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 728         spin_unlock(&tree->lock);
 729         schedule();
 730         spin_lock(&tree->lock);
 731         finish_wait(&state->wq, &wait);
 732 }
 733
 734 /*
 735  * waits for one or more bits to clear on a range in the state tree.
 736  * The range [start, end] is inclusive.
 737  * The tree lock is taken by this function
 738  */
 739 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 740                             unsigned long bits)
 741 {
 742         struct extent_state *state;
 743         struct rb_node *node;
 744
 745         btrfs_debug_check_extent_io_range(tree, start, end);
 746
 747         spin_lock(&tree->lock);
 748 again:
 749         while (1) {
 750                 /*
 751                  * this search will find all the extents that end after
 752                  * our range starts
 753                  */
 754                 node = tree_search(tree, start);
 755 process_node:
 756                 if (!node)
 757                         break;
 758
 759                 state = rb_entry(node, struct extent_state, rb_node);
 760
 761                 if (state->start > end)
 762                         goto out;
 763
 764                 if (state->state & bits) {
 765                         start = state->start;
 766                         atomic_inc(&state->refs);
 767                         wait_on_state(tree, state);
 768                         free_extent_state(state);
 769                         goto again;
 770                 }
 771                 start = state->end + 1;
 772
 773                 if (start > end)
 774                         break;
 775
 776                 if (!cond_resched_lock(&tree->lock)) {
 777                         node = rb_next(node);
 778                         goto process_node;
 779                 }
 780         }
 781 out:
 782         spin_unlock(&tree->lock);
 783 }
 784
 785 static void set_state_bits(struct extent_io_tree *tree,
 786                            struct extent_state *state,
 787                            unsigned long *bits)
 788 {
 789         unsigned long bits_to_set = *bits & ~EXTENT_CTLBITS;
 790
 791         set_state_cb(tree, state, bits);
 792         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 793                 u64 range = state->end - state->start + 1;
 794                 tree->dirty_bytes += range;
 795         }
 796         state->state |= bits_to_set;
 797 }
 798
 799 static void cache_state(struct extent_state *state,
 800                         struct extent_state **cached_ptr)
 801 {
 802         if (cached_ptr && !(*cached_ptr)) {
 803                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 804                         *cached_ptr = state;
 805                         atomic_inc(&state->refs);
 806                 }
 807         }
 808 }
 809
 810 /*
 811  * set some bits on a range in the tree.  This may require allocations or
 812  * sleeping, so the gfp mask is used to indicate what is allowed.
 813  *
 814  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 815  * part of the range already has the desired bits set.  The start of the
 816  * existing range is returned in failed_start in this case.
 817  *
 818  * [start, end] is inclusive This takes the tree lock.
 819  */
 820
 821 static int __must_check
 822 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 823                  unsigned long bits, unsigned long exclusive_bits,
 824                  u64 *failed_start, struct extent_state **cached_state,
 825                  gfp_t mask)
 826 {
 827         struct extent_state *state;
 828         struct extent_state *prealloc = NULL;
 829         struct rb_node *node;
 830         struct rb_node **p;
 831         struct rb_node *parent;
 832         int err = 0;
 833         u64 last_start;
 834         u64 last_end;
 835
 836         btrfs_debug_check_extent_io_range(tree, start, end);
 837
 838         bits |= EXTENT_FIRST_DELALLOC;
 839 again:
 840         if (!prealloc && (mask & __GFP_WAIT)) {
 841                 prealloc = alloc_extent_state(mask);
 842                 BUG_ON(!prealloc);
 843         }
 844
 845         spin_lock(&tree->lock);
 846         if (cached_state && *cached_state) {
 847                 state = *cached_state;
 848                 if (state->start <= start && state->end > start &&
 849                     extent_state_in_tree(state)) {
 850                         node = &state->rb_node;
 851                         goto hit_next;
 852                 }
 853         }
 854         /*
 855          * this search will find all the extents that end after
 856          * our range starts.
 857          */
 858         node = tree_search_for_insert(tree, start, &p, &parent);
 859         if (!node) {
 860                 prealloc = alloc_extent_state_atomic(prealloc);
 861                 BUG_ON(!prealloc);
 862                 err = insert_state(tree, prealloc, start, end,
 863                                    &p, &parent, &bits);
 864                 if (err)
 865                         extent_io_tree_panic(tree, err);
 866
 867                 cache_state(prealloc, cached_state);
 868                 prealloc = NULL;
 869                 goto out;
 870         }
 871         state = rb_entry(node, struct extent_state, rb_node);
 872 hit_next:
 873         last_start = state->start;
 874         last_end = state->end;
 875
 876         /*
 877          * | ---- desired range ---- |
 878          * | state |
 879          *
 880          * Just lock what we found and keep going
 881          */
 882         if (state->start == start && state->end <= end) {
 883                 if (state->state & exclusive_bits) {
 884                         *failed_start = state->start;
 885                         err = -EEXIST;
 886                         goto out;
 887                 }
 888
 889                 set_state_bits(tree, state, &bits);
 890                 cache_state(state, cached_state);
 891                 merge_state(tree, state);
 892                 if (last_end == (u64)-1)
 893                         goto out;
 894                 start = last_end + 1;
 895                 state = next_state(state);
 896                 if (start < end && state && state->start == start &&
 897                     !need_resched())
 898                         goto hit_next;
 899                 goto search_again;
 900         }
 901
 902         /*
 903          *     | ---- desired range ---- |
 904          * | state |
 905          *   or
 906          * | ------------- state -------------- |
 907          *
 908          * We need to split the extent we found, and may flip bits on
 909          * second half.
 910          *
 911          * If the extent we found extends past our
 912          * range, we just split and search again.  It'll get split
 913          * again the next time though.
 914          *
 915          * If the extent we found is inside our range, we set the
 916          * desired bit on it.
 917          */
 918         if (state->start < start) {
 919                 if (state->state & exclusive_bits) {
 920                         *failed_start = start;
 921                         err = -EEXIST;
 922                         goto out;
 923                 }
 924
 925                 prealloc = alloc_extent_state_atomic(prealloc);
 926                 BUG_ON(!prealloc);
 927                 err = split_state(tree, state, prealloc, start);
 928                 if (err)
 929                         extent_io_tree_panic(tree, err);
 930
 931                 prealloc = NULL;
 932                 if (err)
 933                         goto out;
 934                 if (state->end <= end) {
 935                         set_state_bits(tree, state, &bits);
 936                         cache_state(state, cached_state);
 937                         merge_state(tree, state);
 938                         if (last_end == (u64)-1)
 939                                 goto out;
 940                         start = last_end + 1;
 941                         state = next_state(state);
 942                         if (start < end && state && state->start == start &&
 943                             !need_resched())
 944                                 goto hit_next;
 945                 }
 946                 goto search_again;
 947         }
 948         /*
 949          * | ---- desired range ---- |
 950          *     | state | or               | state |
 951          *
 952          * There's a hole, we need to insert something in it and
 953          * ignore the extent we found.
 954          */
 955         if (state->start > start) {
 956                 u64 this_end;
 957                 if (end < last_start)
 958                         this_end = end;
 959                 else
 960                         this_end = last_start - 1;
 961
 962                 prealloc = alloc_extent_state_atomic(prealloc);
 963                 BUG_ON(!prealloc);
 964
 965                 /*
 966                  * Avoid to free 'prealloc' if it can be merged with
 967                  * the later extent.
 968                  */
 969                 err = insert_state(tree, prealloc, start, this_end,
 970                                    NULL, NULL, &bits);
 971                 if (err)
 972                         extent_io_tree_panic(tree, err);
 973
 974                 cache_state(prealloc, cached_state);
 975                 prealloc = NULL;
 976                 start = this_end + 1;
 977                 goto search_again;
 978         }
 979         /*
 980          * | ---- desired range ---- |
 981          *                        | state |
 982          * We need to split the extent, and set the bit
 983          * on the first half
 984          */
 985         if (state->start <= end && state->end > end) {
 986                 if (state->state & exclusive_bits) {
 987                         *failed_start = start;
 988                         err = -EEXIST;
 989                         goto out;
 990                 }
 991
 992                 prealloc = alloc_extent_state_atomic(prealloc);
 993                 BUG_ON(!prealloc);
 994                 err = split_state(tree, state, prealloc, end + 1);
 995                 if (err)
 996                         extent_io_tree_panic(tree, err);
 997
 998                 set_state_bits(tree, prealloc, &bits);
 999                 cache_state(prealloc, cached_state);
1000                 merge_state(tree, prealloc);
1001                 prealloc = NULL;
1002                 goto out;
1003         }
1004
1005         goto search_again;
1006
1007 out:
1008         spin_unlock(&tree->lock);
1009         if (prealloc)
1010                 free_extent_state(prealloc);
1011
1012         return err;
1013
1014 search_again:
1015         if (start > end)
1016                 goto out;
1017         spin_unlock(&tree->lock);
1018         if (mask & __GFP_WAIT)
1019                 cond_resched();
1020         goto again;
1021 }
1022
1023 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1024                    unsigned long bits, u64 * failed_start,
1025                    struct extent_state **cached_state, gfp_t mask)
1026 {
1027         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1028                                 cached_state, mask);
1029 }
1030
1031
1032 /**
1033  * convert_extent_bit - convert all bits in a given range from one bit to
1034  *                      another
1035  * @tree:       the io tree to search
1036  * @start:      the start offset in bytes
1037  * @end:        the end offset in bytes (inclusive)
1038  * @bits:       the bits to set in this range
1039  * @clear_bits: the bits to clear in this range
1040  * @cached_state:       state that we're going to cache
1041  * @mask:       the allocation mask
1042  *
1043  * This will go through and set bits for the given range.  If any states exist
1044  * already in this range they are set with the given bit and cleared of the
1045  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1046  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1047  * boundary bits like LOCK.
1048  */
1049 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1050                        unsigned long bits, unsigned long clear_bits,
1051                        struct extent_state **cached_state, gfp_t mask)
1052 {
1053         struct extent_state *state;
1054         struct extent_state *prealloc = NULL;
1055         struct rb_node *node;
1056         struct rb_node **p;
1057         struct rb_node *parent;
1058         int err = 0;
1059         u64 last_start;
1060         u64 last_end;
1061
1062         btrfs_debug_check_extent_io_range(tree, start, end);
1063
1064 again:
1065         if (!prealloc && (mask & __GFP_WAIT)) {
1066                 prealloc = alloc_extent_state(mask);
1067                 if (!prealloc)
1068                         return -ENOMEM;
1069         }
1070
1071         spin_lock(&tree->lock);
1072         if (cached_state && *cached_state) {
1073                 state = *cached_state;
1074                 if (state->start <= start && state->end > start &&
1075                     extent_state_in_tree(state)) {
1076                         node = &state->rb_node;
1077                         goto hit_next;
1078                 }
1079         }
1080
1081         /*
1082          * this search will find all the extents that end after
1083          * our range starts.
1084          */
1085         node = tree_search_for_insert(tree, start, &p, &parent);
1086         if (!node) {
1087                 prealloc = alloc_extent_state_atomic(prealloc);
1088                 if (!prealloc) {
1089                         err = -ENOMEM;
1090                         goto out;
1091                 }
1092                 err = insert_state(tree, prealloc, start, end,
1093                                    &p, &parent, &bits);
1094                 if (err)
1095                         extent_io_tree_panic(tree, err);
1096                 cache_state(prealloc, cached_state);
1097                 prealloc = NULL;
1098                 goto out;
1099         }
1100         state = rb_entry(node, struct extent_state, rb_node);
1101 hit_next:
1102         last_start = state->start;
1103         last_end = state->end;
1104
1105         /*
1106          * | ---- desired range ---- |
1107          * | state |
1108          *
1109          * Just lock what we found and keep going
1110          */
1111         if (state->start == start && state->end <= end) {
1112                 set_state_bits(tree, state, &bits);
1113                 cache_state(state, cached_state);
1114                 state = clear_state_bit(tree, state, &clear_bits, 0);
1115                 if (last_end == (u64)-1)
1116                         goto out;
1117                 start = last_end + 1;
1118                 if (start < end && state && state->start == start &&
1119                     !need_resched())
1120                         goto hit_next;
1121                 goto search_again;
1122         }
1123
1124         /*
1125          *     | ---- desired range ---- |
1126          * | state |
1127          *   or
1128          * | ------------- state -------------- |
1129          *
1130          * We need to split the extent we found, and may flip bits on
1131          * second half.
1132          *
1133          * If the extent we found extends past our
1134          * range, we just split and search again.  It'll get split
1135          * again the next time though.
1136          *
1137          * If the extent we found is inside our range, we set the
1138          * desired bit on it.
1139          */
1140         if (state->start < start) {
1141                 prealloc = alloc_extent_state_atomic(prealloc);
1142                 if (!prealloc) {
1143                         err = -ENOMEM;
1144                         goto out;
1145                 }
1146                 err = split_state(tree, state, prealloc, start);
1147                 if (err)
1148                         extent_io_tree_panic(tree, err);
1149                 prealloc = NULL;
1150                 if (err)
1151                         goto out;
1152                 if (state->end <= end) {
1153                         set_state_bits(tree, state, &bits);
1154                         cache_state(state, cached_state);
1155                         state = clear_state_bit(tree, state, &clear_bits, 0);
1156                         if (last_end == (u64)-1)
1157                                 goto out;
1158                         start = last_end + 1;
1159                         if (start < end && state && state->start == start &&
1160                             !need_resched())
1161                                 goto hit_next;
1162                 }
1163                 goto search_again;
1164         }
1165         /*
1166          * | ---- desired range ---- |
1167          *     | state | or               | state |
1168          *
1169          * There's a hole, we need to insert something in it and
1170          * ignore the extent we found.
1171          */
1172         if (state->start > start) {
1173                 u64 this_end;
1174                 if (end < last_start)
1175                         this_end = end;
1176                 else
1177                         this_end = last_start - 1;
1178
1179                 prealloc = alloc_extent_state_atomic(prealloc);
1180                 if (!prealloc) {
1181                         err = -ENOMEM;
1182                         goto out;
1183                 }
1184
1185                 /*
1186                  * Avoid to free 'prealloc' if it can be merged with
1187                  * the later extent.
1188                  */
1189                 err = insert_state(tree, prealloc, start, this_end,
1190                                    NULL, NULL, &bits);
1191                 if (err)
1192                         extent_io_tree_panic(tree, err);
1193                 cache_state(prealloc, cached_state);
1194                 prealloc = NULL;
1195                 start = this_end + 1;
1196                 goto search_again;
1197         }
1198         /*
1199          * | ---- desired range ---- |
1200          *                        | state |
1201          * We need to split the extent, and set the bit
1202          * on the first half
1203          */
1204         if (state->start <= end && state->end > end) {
1205                 prealloc = alloc_extent_state_atomic(prealloc);
1206                 if (!prealloc) {
1207                         err = -ENOMEM;
1208                         goto out;
1209                 }
1210
1211                 err = split_state(tree, state, prealloc, end + 1);
1212                 if (err)
1213                         extent_io_tree_panic(tree, err);
1214
1215                 set_state_bits(tree, prealloc, &bits);
1216                 cache_state(prealloc, cached_state);
1217                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1218                 prealloc = NULL;
1219                 goto out;
1220         }
1221
1222         goto search_again;
1223
1224 out:
1225         spin_unlock(&tree->lock);
1226         if (prealloc)
1227                 free_extent_state(prealloc);
1228
1229         return err;
1230
1231 search_again:
1232         if (start > end)
1233                 goto out;
1234         spin_unlock(&tree->lock);
1235         if (mask & __GFP_WAIT)
1236                 cond_resched();
1237         goto again;
1238 }
1239
1240 /* wrappers around set/clear extent bit */
1241 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1242                      gfp_t mask)
1243 {
1244         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1245                               NULL, mask);
1246 }
1247
1248 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1249                     unsigned long bits, gfp_t mask)
1250 {
1251         return set_extent_bit(tree, start, end, bits, NULL,
1252                               NULL, mask);
1253 }
1254
1255 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1256                       unsigned long bits, gfp_t mask)
1257 {
1258         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1259 }
1260
1261 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1262                         struct extent_state **cached_state, gfp_t mask)
1263 {
1264         return set_extent_bit(tree, start, end,
1265                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1266                               NULL, cached_state, mask);
1267 }
1268
1269 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1270                       struct extent_state **cached_state, gfp_t mask)
1271 {
1272         return set_extent_bit(tree, start, end,
1273                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1274                               NULL, cached_state, mask);
1275 }
1276
1277 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1278                        gfp_t mask)
1279 {
1280         return clear_extent_bit(tree, start, end,
1281                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1282                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1283 }
1284
1285 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1286                      gfp_t mask)
1287 {
1288         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1289                               NULL, mask);
1290 }
1291
1292 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1293                         struct extent_state **cached_state, gfp_t mask)
1294 {
1295         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1296                               cached_state, mask);
1297 }
1298
1299 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1300                           struct extent_state **cached_state, gfp_t mask)
1301 {
1302         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1303                                 cached_state, mask);
1304 }
1305
1306 /*
1307  * either insert or lock state struct between start and end use mask to tell
1308  * us if waiting is desired.
1309  */
1310 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1311                      unsigned long bits, struct extent_state **cached_state)
1312 {
1313         int err;
1314         u64 failed_start;
1315         while (1) {
1316                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1317                                        EXTENT_LOCKED, &failed_start,
1318                                        cached_state, GFP_NOFS);
1319                 if (err == -EEXIST) {
1320                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1321                         start = failed_start;
1322                 } else
1323                         break;
1324                 WARN_ON(start > end);
1325         }
1326         return err;
1327 }
1328
1329 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1330 {
1331         return lock_extent_bits(tree, start, end, 0, NULL);
1332 }
1333
1334 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1335 {
1336         int err;
1337         u64 failed_start;
1338
1339         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1340                                &failed_start, NULL, GFP_NOFS);
1341         if (err == -EEXIST) {
1342                 if (failed_start > start)
1343                         clear_extent_bit(tree, start, failed_start - 1,
1344                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1345                 return 0;
1346         }
1347         return 1;
1348 }
1349
1350 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1351                          struct extent_state **cached, gfp_t mask)
1352 {
1353         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1354                                 mask);
1355 }
1356
1357 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1358 {
1359         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1360                                 GFP_NOFS);
1361 }
1362
1363 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1364 {
1365         unsigned long index = start >> PAGE_CACHE_SHIFT;
1366         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1367         struct page *page;
1368
1369         while (index <= end_index) {
1370                 page = find_get_page(inode->i_mapping, index);
1371                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1372                 clear_page_dirty_for_io(page);
1373                 page_cache_release(page);
1374                 index++;
1375         }
1376         return 0;
1377 }
1378
1379 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1380 {
1381         unsigned long index = start >> PAGE_CACHE_SHIFT;
1382         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1383         struct page *page;
1384
1385         while (index <= end_index) {
1386                 page = find_get_page(inode->i_mapping, index);
1387                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1388                 account_page_redirty(page);
1389                 __set_page_dirty_nobuffers(page);
1390                 page_cache_release(page);
1391                 index++;
1392         }
1393         return 0;
1394 }
1395
1396 /*
1397  * helper function to set both pages and extents in the tree writeback
1398  */
1399 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1400 {
1401         unsigned long index = start >> PAGE_CACHE_SHIFT;
1402         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1403         struct page *page;
1404
1405         while (index <= end_index) {
1406                 page = find_get_page(tree->mapping, index);
1407                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1408                 set_page_writeback(page);
1409                 page_cache_release(page);
1410                 index++;
1411         }
1412         return 0;
1413 }
1414
1415 /* find the first state struct with 'bits' set after 'start', and
1416  * return it.  tree->lock must be held.  NULL will returned if
1417  * nothing was found after 'start'
1418  */
1419 static struct extent_state *
1420 find_first_extent_bit_state(struct extent_io_tree *tree,
1421                             u64 start, unsigned long bits)
1422 {
1423         struct rb_node *node;
1424         struct extent_state *state;
1425
1426         /*
1427          * this search will find all the extents that end after
1428          * our range starts.
1429          */
1430         node = tree_search(tree, start);
1431         if (!node)
1432                 goto out;
1433
1434         while (1) {
1435                 state = rb_entry(node, struct extent_state, rb_node);
1436                 if (state->end >= start && (state->state & bits))
1437                         return state;
1438
1439                 node = rb_next(node);
1440                 if (!node)
1441                         break;
1442         }
1443 out:
1444         return NULL;
1445 }
1446
1447 /*
1448  * find the first offset in the io tree with 'bits' set. zero is
1449  * returned if we find something, and *start_ret and *end_ret are
1450  * set to reflect the state struct that was found.
1451  *
1452  * If nothing was found, 1 is returned. If found something, return 0.
1453  */
1454 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1455                           u64 *start_ret, u64 *end_ret, unsigned long bits,
1456                           struct extent_state **cached_state)
1457 {
1458         struct extent_state *state;
1459         struct rb_node *n;
1460         int ret = 1;
1461
1462         spin_lock(&tree->lock);
1463         if (cached_state && *cached_state) {
1464                 state = *cached_state;
1465                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1466                         n = rb_next(&state->rb_node);
1467                         while (n) {
1468                                 state = rb_entry(n, struct extent_state,
1469                                                  rb_node);
1470                                 if (state->state & bits)
1471                                         goto got_it;
1472                                 n = rb_next(n);
1473                         }
1474                         free_extent_state(*cached_state);
1475                         *cached_state = NULL;
1476                         goto out;
1477                 }
1478                 free_extent_state(*cached_state);
1479                 *cached_state = NULL;
1480         }
1481
1482         state = find_first_extent_bit_state(tree, start, bits);
1483 got_it:
1484         if (state) {
1485                 cache_state(state, cached_state);
1486                 *start_ret = state->start;
1487                 *end_ret = state->end;
1488                 ret = 0;
1489         }
1490 out:
1491         spin_unlock(&tree->lock);
1492         return ret;
1493 }
1494
1495 /*
1496  * find a contiguous range of bytes in the file marked as delalloc, not
1497  * more than 'max_bytes'.  start and end are used to return the range,
1498  *
1499  * 1 is returned if we find something, 0 if nothing was in the tree
1500  */
1501 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1502                                         u64 *start, u64 *end, u64 max_bytes,
1503                                         struct extent_state **cached_state)
1504 {
1505         struct rb_node *node;
1506         struct extent_state *state;
1507         u64 cur_start = *start;
1508         u64 found = 0;
1509         u64 total_bytes = 0;
1510
1511         spin_lock(&tree->lock);
1512
1513         /*
1514          * this search will find all the extents that end after
1515          * our range starts.
1516          */
1517         node = tree_search(tree, cur_start);
1518         if (!node) {
1519                 if (!found)
1520                         *end = (u64)-1;
1521                 goto out;
1522         }
1523
1524         while (1) {
1525                 state = rb_entry(node, struct extent_state, rb_node);
1526                 if (found && (state->start != cur_start ||
1527                               (state->state & EXTENT_BOUNDARY))) {
1528                         goto out;
1529                 }
1530                 if (!(state->state & EXTENT_DELALLOC)) {
1531                         if (!found)
1532                                 *end = state->end;
1533                         goto out;
1534                 }
1535                 if (!found) {
1536                         *start = state->start;
1537                         *cached_state = state;
1538                         atomic_inc(&state->refs);
1539                 }
1540                 found++;
1541                 *end = state->end;
1542                 cur_start = state->end + 1;
1543                 node = rb_next(node);
1544                 total_bytes += state->end - state->start + 1;
1545                 if (total_bytes >= max_bytes)
1546                         break;
1547                 if (!node)
1548                         break;
1549         }
1550 out:
1551         spin_unlock(&tree->lock);
1552         return found;
1553 }
1554
1555 static noinline void __unlock_for_delalloc(struct inode *inode,
1556                                            struct page *locked_page,
1557                                            u64 start, u64 end)
1558 {
1559         int ret;
1560         struct page *pages[16];
1561         unsigned long index = start >> PAGE_CACHE_SHIFT;
1562         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1563         unsigned long nr_pages = end_index - index + 1;
1564         int i;
1565
1566         if (index == locked_page->index && end_index == index)
1567                 return;
1568
1569         while (nr_pages > 0) {
1570                 ret = find_get_pages_contig(inode->i_mapping, index,
1571                                      min_t(unsigned long, nr_pages,
1572                                      ARRAY_SIZE(pages)), pages);
1573                 for (i = 0; i < ret; i++) {
1574                         if (pages[i] != locked_page)
1575                                 unlock_page(pages[i]);
1576                         page_cache_release(pages[i]);
1577                 }
1578                 nr_pages -= ret;
1579                 index += ret;
1580                 cond_resched();
1581         }
1582 }
1583
1584 static noinline int lock_delalloc_pages(struct inode *inode,
1585                                         struct page *locked_page,
1586                                         u64 delalloc_start,
1587                                         u64 delalloc_end)
1588 {
1589         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1590         unsigned long start_index = index;
1591         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1592         unsigned long pages_locked = 0;
1593         struct page *pages[16];
1594         unsigned long nrpages;
1595         int ret;
1596         int i;
1597
1598         /* the caller is responsible for locking the start index */
1599         if (index == locked_page->index && index == end_index)
1600                 return 0;
1601
1602         /* skip the page at the start index */
1603         nrpages = end_index - index + 1;
1604         while (nrpages > 0) {
1605                 ret = find_get_pages_contig(inode->i_mapping, index,
1606                                      min_t(unsigned long,
1607                                      nrpages, ARRAY_SIZE(pages)), pages);
1608                 if (ret == 0) {
1609                         ret = -EAGAIN;
1610                         goto done;
1611                 }
1612                 /* now we have an array of pages, lock them all */
1613                 for (i = 0; i < ret; i++) {
1614                         /*
1615                          * the caller is taking responsibility for
1616                          * locked_page
1617                          */
1618                         if (pages[i] != locked_page) {
1619                                 lock_page(pages[i]);
1620                                 if (!PageDirty(pages[i]) ||
1621                                     pages[i]->mapping != inode->i_mapping) {
1622                                         ret = -EAGAIN;
1623                                         unlock_page(pages[i]);
1624                                         page_cache_release(pages[i]);
1625                                         goto done;
1626                                 }
1627                         }
1628                         page_cache_release(pages[i]);
1629                         pages_locked++;
1630                 }
1631                 nrpages -= ret;
1632                 index += ret;
1633                 cond_resched();
1634         }
1635         ret = 0;
1636 done:
1637         if (ret && pages_locked) {
1638                 __unlock_for_delalloc(inode, locked_page,
1639                               delalloc_start,
1640                               ((u64)(start_index + pages_locked - 1)) <<
1641                               PAGE_CACHE_SHIFT);
1642         }
1643         return ret;
1644 }
1645
1646 /*
1647  * find a contiguous range of bytes in the file marked as delalloc, not
1648  * more than 'max_bytes'.  start and end are used to return the range,
1649  *
1650  * 1 is returned if we find something, 0 if nothing was in the tree
1651  */
1652 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1653                                     struct extent_io_tree *tree,
1654                                     struct page *locked_page, u64 *start,
1655                                     u64 *end, u64 max_bytes)
1656 {
1657         u64 delalloc_start;
1658         u64 delalloc_end;
1659         u64 found;
1660         struct extent_state *cached_state = NULL;
1661         int ret;
1662         int loops = 0;
1663
1664 again:
1665         /* step one, find a bunch of delalloc bytes starting at start */
1666         delalloc_start = *start;
1667         delalloc_end = 0;
1668         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1669                                     max_bytes, &cached_state);
1670         if (!found || delalloc_end <= *start) {
1671                 *start = delalloc_start;
1672                 *end = delalloc_end;
1673                 free_extent_state(cached_state);
1674                 return 0;
1675         }
1676
1677         /*
1678          * start comes from the offset of locked_page.  We have to lock
1679          * pages in order, so we can't process delalloc bytes before
1680          * locked_page
1681          */
1682         if (delalloc_start < *start)
1683                 delalloc_start = *start;
1684
1685         /*
1686          * make sure to limit the number of pages we try to lock down
1687          */
1688         if (delalloc_end + 1 - delalloc_start > max_bytes)
1689                 delalloc_end = delalloc_start + max_bytes - 1;
1690
1691         /* step two, lock all the pages after the page that has start */
1692         ret = lock_delalloc_pages(inode, locked_page,
1693                                   delalloc_start, delalloc_end);
1694         if (ret == -EAGAIN) {
1695                 /* some of the pages are gone, lets avoid looping by
1696                  * shortening the size of the delalloc range we're searching
1697                  */
1698                 free_extent_state(cached_state);
1699                 cached_state = NULL;
1700                 if (!loops) {
1701                         max_bytes = PAGE_CACHE_SIZE;
1702                         loops = 1;
1703                         goto again;
1704                 } else {
1705                         found = 0;
1706                         goto out_failed;
1707                 }
1708         }
1709         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1710
1711         /* step three, lock the state bits for the whole range */
1712         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1713
1714         /* then test to make sure it is all still delalloc */
1715         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1716                              EXTENT_DELALLOC, 1, cached_state);
1717         if (!ret) {
1718                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1719                                      &cached_state, GFP_NOFS);
1720                 __unlock_for_delalloc(inode, locked_page,
1721                               delalloc_start, delalloc_end);
1722                 cond_resched();
1723                 goto again;
1724         }
1725         free_extent_state(cached_state);
1726         *start = delalloc_start;
1727         *end = delalloc_end;
1728 out_failed:
1729         return found;
1730 }
1731
1732 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1733                                  struct page *locked_page,
1734                                  unsigned long clear_bits,
1735                                  unsigned long page_ops)
1736 {
1737         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1738         int ret;
1739         struct page *pages[16];
1740         unsigned long index = start >> PAGE_CACHE_SHIFT;
1741         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1742         unsigned long nr_pages = end_index - index + 1;
1743         int i;
1744
1745         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1746         if (page_ops == 0)
1747                 return 0;
1748
1749         while (nr_pages > 0) {
1750                 ret = find_get_pages_contig(inode->i_mapping, index,
1751                                      min_t(unsigned long,
1752                                      nr_pages, ARRAY_SIZE(pages)), pages);
1753                 for (i = 0; i < ret; i++) {
1754
1755                         if (page_ops & PAGE_SET_PRIVATE2)
1756                                 SetPagePrivate2(pages[i]);
1757
1758                         if (pages[i] == locked_page) {
1759                                 page_cache_release(pages[i]);
1760                                 continue;
1761                         }
1762                         if (page_ops & PAGE_CLEAR_DIRTY)
1763                                 clear_page_dirty_for_io(pages[i]);
1764                         if (page_ops & PAGE_SET_WRITEBACK)
1765                                 set_page_writeback(pages[i]);
1766                         if (page_ops & PAGE_END_WRITEBACK)
1767                                 end_page_writeback(pages[i]);
1768                         if (page_ops & PAGE_UNLOCK)
1769                                 unlock_page(pages[i]);
1770                         page_cache_release(pages[i]);
1771                 }
1772                 nr_pages -= ret;
1773                 index += ret;
1774                 cond_resched();
1775         }
1776         return 0;
1777 }
1778
1779 /*
1780  * count the number of bytes in the tree that have a given bit(s)
1781  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1782  * cached.  The total number found is returned.
1783  */
1784 u64 count_range_bits(struct extent_io_tree *tree,
1785                      u64 *start, u64 search_end, u64 max_bytes,
1786                      unsigned long bits, int contig)
1787 {
1788         struct rb_node *node;
1789         struct extent_state *state;
1790         u64 cur_start = *start;
1791         u64 total_bytes = 0;
1792         u64 last = 0;
1793         int found = 0;
1794
1795         if (WARN_ON(search_end <= cur_start))
1796                 return 0;
1797
1798         spin_lock(&tree->lock);
1799         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1800                 total_bytes = tree->dirty_bytes;
1801                 goto out;
1802         }
1803         /*
1804          * this search will find all the extents that end after
1805          * our range starts.
1806          */
1807         node = tree_search(tree, cur_start);
1808         if (!node)
1809                 goto out;
1810
1811         while (1) {
1812                 state = rb_entry(node, struct extent_state, rb_node);
1813                 if (state->start > search_end)
1814                         break;
1815                 if (contig && found && state->start > last + 1)
1816                         break;
1817                 if (state->end >= cur_start && (state->state & bits) == bits) {
1818                         total_bytes += min(search_end, state->end) + 1 -
1819                                        max(cur_start, state->start);
1820                         if (total_bytes >= max_bytes)
1821                                 break;
1822                         if (!found) {
1823                                 *start = max(cur_start, state->start);
1824                                 found = 1;
1825                         }
1826                         last = state->end;
1827                 } else if (contig && found) {
1828                         break;
1829                 }
1830                 node = rb_next(node);
1831                 if (!node)
1832                         break;
1833         }
1834 out:
1835         spin_unlock(&tree->lock);
1836         return total_bytes;
1837 }
1838
1839 /*
1840  * set the private field for a given byte offset in the tree.  If there isn't
1841  * an extent_state there already, this does nothing.
1842  */
1843 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1844 {
1845         struct rb_node *node;
1846         struct extent_state *state;
1847         int ret = 0;
1848
1849         spin_lock(&tree->lock);
1850         /*
1851          * this search will find all the extents that end after
1852          * our range starts.
1853          */
1854         node = tree_search(tree, start);
1855         if (!node) {
1856                 ret = -ENOENT;
1857                 goto out;
1858         }
1859         state = rb_entry(node, struct extent_state, rb_node);
1860         if (state->start != start) {
1861                 ret = -ENOENT;
1862                 goto out;
1863         }
1864         state->private = private;
1865 out:
1866         spin_unlock(&tree->lock);
1867         return ret;
1868 }
1869
1870 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1871 {
1872         struct rb_node *node;
1873         struct extent_state *state;
1874         int ret = 0;
1875
1876         spin_lock(&tree->lock);
1877         /*
1878          * this search will find all the extents that end after
1879          * our range starts.
1880          */
1881         node = tree_search(tree, start);
1882         if (!node) {
1883                 ret = -ENOENT;
1884                 goto out;
1885         }
1886         state = rb_entry(node, struct extent_state, rb_node);
1887         if (state->start != start) {
1888                 ret = -ENOENT;
1889                 goto out;
1890         }
1891         *private = state->private;
1892 out:
1893         spin_unlock(&tree->lock);
1894         return ret;
1895 }
1896
1897 /*
1898  * searches a range in the state tree for a given mask.
1899  * If 'filled' == 1, this returns 1 only if every extent in the tree
1900  * has the bits set.  Otherwise, 1 is returned if any bit in the
1901  * range is found set.
1902  */
1903 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1904                    unsigned long bits, int filled, struct extent_state *cached)
1905 {
1906         struct extent_state *state = NULL;
1907         struct rb_node *node;
1908         int bitset = 0;
1909
1910         spin_lock(&tree->lock);
1911         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1912             cached->end > start)
1913                 node = &cached->rb_node;
1914         else
1915                 node = tree_search(tree, start);
1916         while (node && start <= end) {
1917                 state = rb_entry(node, struct extent_state, rb_node);
1918
1919                 if (filled && state->start > start) {
1920                         bitset = 0;
1921                         break;
1922                 }
1923
1924                 if (state->start > end)
1925                         break;
1926
1927                 if (state->state & bits) {
1928                         bitset = 1;
1929                         if (!filled)
1930                                 break;
1931                 } else if (filled) {
1932                         bitset = 0;
1933                         break;
1934                 }
1935
1936                 if (state->end == (u64)-1)
1937                         break;
1938
1939                 start = state->end + 1;
1940                 if (start > end)
1941                         break;
1942                 node = rb_next(node);
1943                 if (!node) {
1944                         if (filled)
1945                                 bitset = 0;
1946                         break;
1947                 }
1948         }
1949         spin_unlock(&tree->lock);
1950         return bitset;
1951 }
1952
1953 /*
1954  * helper function to set a given page up to date if all the
1955  * extents in the tree for that page are up to date
1956  */
1957 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1958 {
1959         u64 start = page_offset(page);
1960         u64 end = start + PAGE_CACHE_SIZE - 1;
1961         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1962                 SetPageUptodate(page);
1963 }
1964
1965 static int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1966 {
1967         int ret;
1968         int err = 0;
1969         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1970
1971         set_state_private(failure_tree, rec->start, 0);
1972         ret = clear_extent_bits(failure_tree, rec->start,
1973                                 rec->start + rec->len - 1,
1974                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1975         if (ret)
1976                 err = ret;
1977
1978         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1979                                 rec->start + rec->len - 1,
1980                                 EXTENT_DAMAGED, GFP_NOFS);
1981         if (ret && !err)
1982                 err = ret;
1983
1984         kfree(rec);
1985         return err;
1986 }
1987
1988 /*
1989  * this bypasses the standard btrfs submit functions deliberately, as
1990  * the standard behavior is to write all copies in a raid setup. here we only
1991  * want to write the one bad copy. so we do the mapping for ourselves and issue
1992  * submit_bio directly.
1993  * to avoid any synchronization issues, wait for the data after writing, which
1994  * actually prevents the read that triggered the error from finishing.
1995  * currently, there can be no more than two copies of every data bit. thus,
1996  * exactly one rewrite is required.
1997  */
1998 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 start,
1999                         u64 length, u64 logical, struct page *page,
2000                         int mirror_num)
2001 {
2002         struct bio *bio;
2003         struct btrfs_device *dev;
2004         u64 map_length = 0;
2005         u64 sector;
2006         struct btrfs_bio *bbio = NULL;
2007         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2008         int ret;
2009
2010         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2011         BUG_ON(!mirror_num);
2012
2013         /* we can't repair anything in raid56 yet */
2014         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2015                 return 0;
2016
2017         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2018         if (!bio)
2019                 return -EIO;
2020         bio->bi_iter.bi_size = 0;
2021         map_length = length;
2022
2023         ret = btrfs_map_block(fs_info, WRITE, logical,
2024                               &map_length, &bbio, mirror_num);
2025         if (ret) {
2026                 bio_put(bio);
2027                 return -EIO;
2028         }
2029         BUG_ON(mirror_num != bbio->mirror_num);
2030         sector = bbio->stripes[mirror_num-1].physical >> 9;
2031         bio->bi_iter.bi_sector = sector;
2032         dev = bbio->stripes[mirror_num-1].dev;
2033         kfree(bbio);
2034         if (!dev || !dev->bdev || !dev->writeable) {
2035                 bio_put(bio);
2036                 return -EIO;
2037         }
2038         bio->bi_bdev = dev->bdev;
2039         bio_add_page(bio, page, length, start - page_offset(page));
2040
2041         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2042                 /* try to remap that extent elsewhere? */
2043                 bio_put(bio);
2044                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2045                 return -EIO;
2046         }
2047
2048         printk_ratelimited_in_rcu(KERN_INFO
2049                         "BTRFS: read error corrected: ino %lu off %llu "
2050                     "(dev %s sector %llu)\n", page->mapping->host->i_ino,
2051                     start, rcu_str_deref(dev->name), sector);
2052
2053         bio_put(bio);
2054         return 0;
2055 }
2056
2057 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2058                          int mirror_num)
2059 {
2060         u64 start = eb->start;
2061         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2062         int ret = 0;
2063
2064         if (root->fs_info->sb->s_flags & MS_RDONLY)
2065                 return -EROFS;
2066
2067         for (i = 0; i < num_pages; i++) {
2068                 struct page *p = extent_buffer_page(eb, i);
2069                 ret = repair_io_failure(root->fs_info, start, PAGE_CACHE_SIZE,
2070                                         start, p, mirror_num);
2071                 if (ret)
2072                         break;
2073                 start += PAGE_CACHE_SIZE;
2074         }
2075
2076         return ret;
2077 }
2078
2079 /*
2080  * each time an IO finishes, we do a fast check in the IO failure tree
2081  * to see if we need to process or clean up an io_failure_record
2082  */
2083 static int clean_io_failure(u64 start, struct page *page)
2084 {
2085         u64 private;
2086         u64 private_failure;
2087         struct io_failure_record *failrec;
2088         struct inode *inode = page->mapping->host;
2089         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2090         struct extent_state *state;
2091         int num_copies;
2092         int ret;
2093
2094         private = 0;
2095         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2096                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2097         if (!ret)
2098                 return 0;
2099
2100         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2101                                 &private_failure);
2102         if (ret)
2103                 return 0;
2104
2105         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2106         BUG_ON(!failrec->this_mirror);
2107
2108         if (failrec->in_validation) {
2109                 /* there was no real error, just free the record */
2110                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2111                          failrec->start);
2112                 goto out;
2113         }
2114         if (fs_info->sb->s_flags & MS_RDONLY)
2115                 goto out;
2116
2117         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2118         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2119                                             failrec->start,
2120                                             EXTENT_LOCKED);
2121         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2122
2123         if (state && state->start <= failrec->start &&
2124             state->end >= failrec->start + failrec->len - 1) {
2125                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2126                                               failrec->len);
2127                 if (num_copies > 1)  {
2128                         repair_io_failure(fs_info, start, failrec->len,
2129                                           failrec->logical, page,
2130                                           failrec->failed_mirror);
2131                 }
2132         }
2133
2134 out:
2135         free_io_failure(inode, failrec);
2136
2137         return 0;
2138 }
2139
2140 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2141                                 struct io_failure_record **failrec_ret)
2142 {
2143         struct io_failure_record *failrec;
2144         u64 private;
2145         struct extent_map *em;
2146         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2147         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2148         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2149         int ret;
2150         u64 logical;
2151
2152         ret = get_state_private(failure_tree, start, &private);
2153         if (ret) {
2154                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2155                 if (!failrec)
2156                         return -ENOMEM;
2157
2158                 failrec->start = start;
2159                 failrec->len = end - start + 1;
2160                 failrec->this_mirror = 0;
2161                 failrec->bio_flags = 0;
2162                 failrec->in_validation = 0;
2163
2164                 read_lock(&em_tree->lock);
2165                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2166                 if (!em) {
2167                         read_unlock(&em_tree->lock);
2168                         kfree(failrec);
2169                         return -EIO;
2170                 }
2171
2172                 if (em->start > start || em->start + em->len <= start) {
2173                         free_extent_map(em);
2174                         em = NULL;
2175                 }
2176                 read_unlock(&em_tree->lock);
2177                 if (!em) {
2178                         kfree(failrec);
2179                         return -EIO;
2180                 }
2181
2182                 logical = start - em->start;
2183                 logical = em->block_start + logical;
2184                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2185                         logical = em->block_start;
2186                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2187                         extent_set_compress_type(&failrec->bio_flags,
2188                                                  em->compress_type);
2189                 }
2190
2191                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2192                          logical, start, failrec->len);
2193
2194                 failrec->logical = logical;
2195                 free_extent_map(em);
2196
2197                 /* set the bits in the private failure tree */
2198                 ret = set_extent_bits(failure_tree, start, end,
2199                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2200                 if (ret >= 0)
2201                         ret = set_state_private(failure_tree, start,
2202                                                 (u64)(unsigned long)failrec);
2203                 /* set the bits in the inode's tree */
2204                 if (ret >= 0)
2205                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2206                                                 GFP_NOFS);
2207                 if (ret < 0) {
2208                         kfree(failrec);
2209                         return ret;
2210                 }
2211         } else {
2212                 failrec = (struct io_failure_record *)(unsigned long)private;
2213                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2214                          failrec->logical, failrec->start, failrec->len,
2215                          failrec->in_validation);
2216                 /*
2217                  * when data can be on disk more than twice, add to failrec here
2218                  * (e.g. with a list for failed_mirror) to make
2219                  * clean_io_failure() clean all those errors at once.
2220                  */
2221         }
2222
2223         *failrec_ret = failrec;
2224
2225         return 0;
2226 }
2227
2228 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2229                            struct io_failure_record *failrec, int failed_mirror)
2230 {
2231         int num_copies;
2232
2233         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2234                                       failrec->logical, failrec->len);
2235         if (num_copies == 1) {
2236                 /*
2237                  * we only have a single copy of the data, so don't bother with
2238                  * all the retry and error correction code that follows. no
2239                  * matter what the error is, it is very likely to persist.
2240                  */
2241                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2242                          num_copies, failrec->this_mirror, failed_mirror);
2243                 return 0;
2244         }
2245
2246         /*
2247          * there are two premises:
2248          *      a) deliver good data to the caller
2249          *      b) correct the bad sectors on disk
2250          */
2251         if (failed_bio->bi_vcnt > 1) {
2252                 /*
2253                  * to fulfill b), we need to know the exact failing sectors, as
2254                  * we don't want to rewrite any more than the failed ones. thus,
2255                  * we need separate read requests for the failed bio
2256                  *
2257                  * if the following BUG_ON triggers, our validation request got
2258                  * merged. we need separate requests for our algorithm to work.
2259                  */
2260                 BUG_ON(failrec->in_validation);
2261                 failrec->in_validation = 1;
2262                 failrec->this_mirror = failed_mirror;
2263         } else {
2264                 /*
2265                  * we're ready to fulfill a) and b) alongside. get a good copy
2266                  * of the failed sector and if we succeed, we have setup
2267                  * everything for repair_io_failure to do the rest for us.
2268                  */
2269                 if (failrec->in_validation) {
2270                         BUG_ON(failrec->this_mirror != failed_mirror);
2271                         failrec->in_validation = 0;
2272                         failrec->this_mirror = 0;
2273                 }
2274                 failrec->failed_mirror = failed_mirror;
2275                 failrec->this_mirror++;
2276                 if (failrec->this_mirror == failed_mirror)
2277                         failrec->this_mirror++;
2278         }
2279
2280         if (failrec->this_mirror > num_copies) {
2281                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2282                          num_copies, failrec->this_mirror, failed_mirror);
2283                 return 0;
2284         }
2285
2286         return 1;
2287 }
2288
2289
2290 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2291                                     struct io_failure_record *failrec,
2292                                     struct page *page, int pg_offset, int icsum,
2293                                     bio_end_io_t *endio_func)
2294 {
2295         struct bio *bio;
2296         struct btrfs_io_bio *btrfs_failed_bio;
2297         struct btrfs_io_bio *btrfs_bio;
2298
2299         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2300         if (!bio)
2301                 return NULL;
2302
2303         bio->bi_end_io = endio_func;
2304         bio->bi_iter.bi_sector = failrec->logical >> 9;
2305         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2306         bio->bi_iter.bi_size = 0;
2307
2308         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2309         if (btrfs_failed_bio->csum) {
2310                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2311                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2312
2313                 btrfs_bio = btrfs_io_bio(bio);
2314                 btrfs_bio->csum = btrfs_bio->csum_inline;
2315                 icsum *= csum_size;
2316                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2317                        csum_size);
2318         }
2319
2320         bio_add_page(bio, page, failrec->len, pg_offset);
2321
2322         return bio;
2323 }
2324
2325 /*
2326  * this is a generic handler for readpage errors (default
2327  * readpage_io_failed_hook). if other copies exist, read those and write back
2328  * good data to the failed position. does not investigate in remapping the
2329  * failed extent elsewhere, hoping the device will be smart enough to do this as
2330  * needed
2331  */
2332
2333 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2334                               struct page *page, u64 start, u64 end,
2335                               int failed_mirror)
2336 {
2337         struct io_failure_record *failrec;
2338         struct inode *inode = page->mapping->host;
2339         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2340         struct bio *bio;
2341         int read_mode;
2342         int ret;
2343
2344         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2345
2346         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2347         if (ret)
2348                 return ret;
2349
2350         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2351         if (!ret) {
2352                 free_io_failure(inode, failrec);
2353                 return -EIO;
2354         }
2355
2356         if (failed_bio->bi_vcnt > 1)
2357                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2358         else
2359                 read_mode = READ_SYNC;
2360
2361         phy_offset >>= inode->i_sb->s_blocksize_bits;
2362         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2363                                       start - page_offset(page),
2364                                       (int)phy_offset, failed_bio->bi_end_io);
2365         if (!bio) {
2366                 free_io_failure(inode, failrec);
2367                 return -EIO;
2368         }
2369
2370         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2371                  read_mode, failrec->this_mirror, failrec->in_validation);
2372
2373         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2374                                          failrec->this_mirror,
2375                                          failrec->bio_flags, 0);
2376         if (ret) {
2377                 free_io_failure(inode, failrec);
2378                 bio_put(bio);
2379         }
2380
2381         return ret;
2382 }
2383
2384 /* lots and lots of room for performance fixes in the end_bio funcs */
2385
2386 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2387 {
2388         int uptodate = (err == 0);
2389         struct extent_io_tree *tree;
2390         int ret = 0;
2391
2392         tree = &BTRFS_I(page->mapping->host)->io_tree;
2393
2394         if (tree->ops && tree->ops->writepage_end_io_hook) {
2395                 ret = tree->ops->writepage_end_io_hook(page, start,
2396                                                end, NULL, uptodate);
2397                 if (ret)
2398                         uptodate = 0;
2399         }
2400
2401         if (!uptodate) {
2402                 ClearPageUptodate(page);
2403                 SetPageError(page);
2404                 ret = ret < 0 ? ret : -EIO;
2405                 mapping_set_error(page->mapping, ret);
2406         }
2407         return 0;
2408 }
2409
2410 /*
2411  * after a writepage IO is done, we need to:
2412  * clear the uptodate bits on error
2413  * clear the writeback bits in the extent tree for this IO
2414  * end_page_writeback if the page has no more pending IO
2415  *
2416  * Scheduling is not allowed, so the extent state tree is expected
2417  * to have one and only one object corresponding to this IO.
2418  */
2419 static void end_bio_extent_writepage(struct bio *bio, int err)
2420 {
2421         struct bio_vec *bvec;
2422         u64 start;
2423         u64 end;
2424         int i;
2425
2426         bio_for_each_segment_all(bvec, bio, i) {
2427                 struct page *page = bvec->bv_page;
2428
2429                 /* We always issue full-page reads, but if some block
2430                  * in a page fails to read, blk_update_request() will
2431                  * advance bv_offset and adjust bv_len to compensate.
2432                  * Print a warning for nonzero offsets, and an error
2433                  * if they don't add up to a full page.  */
2434                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2435                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2436                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2437                                    "partial page write in btrfs with offset %u and length %u",
2438                                         bvec->bv_offset, bvec->bv_len);
2439                         else
2440                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2441                                    "incomplete page write in btrfs with offset %u and "
2442                                    "length %u",
2443                                         bvec->bv_offset, bvec->bv_len);
2444                 }
2445
2446                 start = page_offset(page);
2447                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2448
2449                 if (end_extent_writepage(page, err, start, end))
2450                         continue;
2451
2452                 end_page_writeback(page);
2453         }
2454
2455         bio_put(bio);
2456 }
2457
2458 static void
2459 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2460                               int uptodate)
2461 {
2462         struct extent_state *cached = NULL;
2463         u64 end = start + len - 1;
2464
2465         if (uptodate && tree->track_uptodate)
2466                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2467         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2468 }
2469
2470 /*
2471  * after a readpage IO is done, we need to:
2472  * clear the uptodate bits on error
2473  * set the uptodate bits if things worked
2474  * set the page up to date if all extents in the tree are uptodate
2475  * clear the lock bit in the extent tree
2476  * unlock the page if there are no other extents locked for it
2477  *
2478  * Scheduling is not allowed, so the extent state tree is expected
2479  * to have one and only one object corresponding to this IO.
2480  */
2481 static void end_bio_extent_readpage(struct bio *bio, int err)
2482 {
2483         struct bio_vec *bvec;
2484         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2485         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2486         struct extent_io_tree *tree;
2487         u64 offset = 0;
2488         u64 start;
2489         u64 end;
2490         u64 len;
2491         u64 extent_start = 0;
2492         u64 extent_len = 0;
2493         int mirror;
2494         int ret;
2495         int i;
2496
2497         if (err)
2498                 uptodate = 0;
2499
2500         bio_for_each_segment_all(bvec, bio, i) {
2501                 struct page *page = bvec->bv_page;
2502                 struct inode *inode = page->mapping->host;
2503
2504                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2505                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector, err,
2506                          io_bio->mirror_num);
2507                 tree = &BTRFS_I(inode)->io_tree;
2508
2509                 /* We always issue full-page reads, but if some block
2510                  * in a page fails to read, blk_update_request() will
2511                  * advance bv_offset and adjust bv_len to compensate.
2512                  * Print a warning for nonzero offsets, and an error
2513                  * if they don't add up to a full page.  */
2514                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2515                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2516                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2517                                    "partial page read in btrfs with offset %u and length %u",
2518                                         bvec->bv_offset, bvec->bv_len);
2519                         else
2520                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2521                                    "incomplete page read in btrfs with offset %u and "
2522                                    "length %u",
2523                                         bvec->bv_offset, bvec->bv_len);
2524                 }
2525
2526                 start = page_offset(page);
2527                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2528                 len = bvec->bv_len;
2529
2530                 mirror = io_bio->mirror_num;
2531                 if (likely(uptodate && tree->ops &&
2532                            tree->ops->readpage_end_io_hook)) {
2533                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2534                                                               page, start, end,
2535                                                               mirror);
2536                         if (ret)
2537                                 uptodate = 0;
2538                         else
2539                                 clean_io_failure(start, page);
2540                 }
2541
2542                 if (likely(uptodate))
2543                         goto readpage_ok;
2544
2545                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2546                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2547                         if (!ret && !err &&
2548                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2549                                 uptodate = 1;
2550                 } else {
2551                         /*
2552                          * The generic bio_readpage_error handles errors the
2553                          * following way: If possible, new read requests are
2554                          * created and submitted and will end up in
2555                          * end_bio_extent_readpage as well (if we're lucky, not
2556                          * in the !uptodate case). In that case it returns 0 and
2557                          * we just go on with the next page in our bio. If it
2558                          * can't handle the error it will return -EIO and we
2559                          * remain responsible for that page.
2560                          */
2561                         ret = bio_readpage_error(bio, offset, page, start, end,
2562                                                  mirror);
2563                         if (ret == 0) {
2564                                 uptodate =
2565                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2566                                 if (err)
2567                                         uptodate = 0;
2568                                 offset += len;
2569                                 continue;
2570                         }
2571                 }
2572 readpage_ok:
2573                 if (likely(uptodate)) {
2574                         loff_t i_size = i_size_read(inode);
2575                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2576                         unsigned off;
2577
2578                         /* Zero out the end if this page straddles i_size */
2579                         off = i_size & (PAGE_CACHE_SIZE-1);
2580                         if (page->index == end_index && off)
2581                                 zero_user_segment(page, off, PAGE_CACHE_SIZE);
2582                         SetPageUptodate(page);
2583                 } else {
2584                         ClearPageUptodate(page);
2585                         SetPageError(page);
2586                 }
2587                 unlock_page(page);
2588                 offset += len;
2589
2590                 if (unlikely(!uptodate)) {
2591                         if (extent_len) {
2592                                 endio_readpage_release_extent(tree,
2593                                                               extent_start,
2594                                                               extent_len, 1);
2595                                 extent_start = 0;
2596                                 extent_len = 0;
2597                         }
2598                         endio_readpage_release_extent(tree, start,
2599                                                       end - start + 1, 0);
2600                 } else if (!extent_len) {
2601                         extent_start = start;
2602                         extent_len = end + 1 - start;
2603                 } else if (extent_start + extent_len == start) {
2604                         extent_len += end + 1 - start;
2605                 } else {
2606                         endio_readpage_release_extent(tree, extent_start,
2607                                                       extent_len, uptodate);
2608                         extent_start = start;
2609                         extent_len = end + 1 - start;
2610                 }
2611         }
2612
2613         if (extent_len)
2614                 endio_readpage_release_extent(tree, extent_start, extent_len,
2615                                               uptodate);
2616         if (io_bio->end_io)
2617                 io_bio->end_io(io_bio, err);
2618         bio_put(bio);
2619 }
2620
2621 /*
2622  * this allocates from the btrfs_bioset.  We're returning a bio right now
2623  * but you can call btrfs_io_bio for the appropriate container_of magic
2624  */
2625 struct bio *
2626 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2627                 gfp_t gfp_flags)
2628 {
2629         struct btrfs_io_bio *btrfs_bio;
2630         struct bio *bio;
2631
2632         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2633
2634         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2635                 while (!bio && (nr_vecs /= 2)) {
2636                         bio = bio_alloc_bioset(gfp_flags,
2637                                                nr_vecs, btrfs_bioset);
2638                 }
2639         }
2640
2641         if (bio) {
2642                 bio->bi_bdev = bdev;
2643                 bio->bi_iter.bi_sector = first_sector;
2644                 btrfs_bio = btrfs_io_bio(bio);
2645                 btrfs_bio->csum = NULL;
2646                 btrfs_bio->csum_allocated = NULL;
2647                 btrfs_bio->end_io = NULL;
2648         }
2649         return bio;
2650 }
2651
2652 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2653 {
2654         struct btrfs_io_bio *btrfs_bio;
2655         struct bio *new;
2656
2657         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2658         if (new) {
2659                 btrfs_bio = btrfs_io_bio(new);
2660                 btrfs_bio->csum = NULL;
2661                 btrfs_bio->csum_allocated = NULL;
2662                 btrfs_bio->end_io = NULL;
2663         }
2664         return new;
2665 }
2666
2667 /* this also allocates from the btrfs_bioset */
2668 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2669 {
2670         struct btrfs_io_bio *btrfs_bio;
2671         struct bio *bio;
2672
2673         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2674         if (bio) {
2675                 btrfs_bio = btrfs_io_bio(bio);
2676                 btrfs_bio->csum = NULL;
2677                 btrfs_bio->csum_allocated = NULL;
2678                 btrfs_bio->end_io = NULL;
2679         }
2680         return bio;
2681 }
2682
2683
2684 static int __must_check submit_one_bio(int rw, struct bio *bio,
2685                                        int mirror_num, unsigned long bio_flags)
2686 {
2687         int ret = 0;
2688         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2689         struct page *page = bvec->bv_page;
2690         struct extent_io_tree *tree = bio->bi_private;
2691         u64 start;
2692
2693         start = page_offset(page) + bvec->bv_offset;
2694
2695         bio->bi_private = NULL;
2696
2697         bio_get(bio);
2698
2699         if (tree->ops && tree->ops->submit_bio_hook)
2700                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2701                                            mirror_num, bio_flags, start);
2702         else
2703                 btrfsic_submit_bio(rw, bio);
2704
2705         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2706                 ret = -EOPNOTSUPP;
2707         bio_put(bio);
2708         return ret;
2709 }
2710
2711 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2712                      unsigned long offset, size_t size, struct bio *bio,
2713                      unsigned long bio_flags)
2714 {
2715         int ret = 0;
2716         if (tree->ops && tree->ops->merge_bio_hook)
2717                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2718                                                 bio_flags);
2719         BUG_ON(ret < 0);
2720         return ret;
2721
2722 }
2723
2724 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2725                               struct page *page, sector_t sector,
2726                               size_t size, unsigned long offset,
2727                               struct block_device *bdev,
2728                               struct bio **bio_ret,
2729                               unsigned long max_pages,
2730                               bio_end_io_t end_io_func,
2731                               int mirror_num,
2732                               unsigned long prev_bio_flags,
2733                               unsigned long bio_flags)
2734 {
2735         int ret = 0;
2736         struct bio *bio;
2737         int nr;
2738         int contig = 0;
2739         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2740         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2741         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2742
2743         if (bio_ret && *bio_ret) {
2744                 bio = *bio_ret;
2745                 if (old_compressed)
2746                         contig = bio->bi_iter.bi_sector == sector;
2747                 else
2748                         contig = bio_end_sector(bio) == sector;
2749
2750                 if (prev_bio_flags != bio_flags || !contig ||
2751                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2752                     bio_add_page(bio, page, page_size, offset) < page_size) {
2753                         ret = submit_one_bio(rw, bio, mirror_num,
2754                                              prev_bio_flags);
2755                         if (ret < 0)
2756                                 return ret;
2757                         bio = NULL;
2758                 } else {
2759                         return 0;
2760                 }
2761         }
2762         if (this_compressed)
2763                 nr = BIO_MAX_PAGES;
2764         else
2765                 nr = bio_get_nr_vecs(bdev);
2766
2767         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2768         if (!bio)
2769                 return -ENOMEM;
2770
2771         bio_add_page(bio, page, page_size, offset);
2772         bio->bi_end_io = end_io_func;
2773         bio->bi_private = tree;
2774
2775         if (bio_ret)
2776                 *bio_ret = bio;
2777         else
2778                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2779
2780         return ret;
2781 }
2782
2783 static void attach_extent_buffer_page(struct extent_buffer *eb,
2784                                       struct page *page)
2785 {
2786         if (!PagePrivate(page)) {
2787                 SetPagePrivate(page);
2788                 page_cache_get(page);
2789                 set_page_private(page, (unsigned long)eb);
2790         } else {
2791                 WARN_ON(page->private != (unsigned long)eb);
2792         }
2793 }
2794
2795 void set_page_extent_mapped(struct page *page)
2796 {
2797         if (!PagePrivate(page)) {
2798                 SetPagePrivate(page);
2799                 page_cache_get(page);
2800                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2801         }
2802 }
2803
2804 static struct extent_map *
2805 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2806                  u64 start, u64 len, get_extent_t *get_extent,
2807                  struct extent_map **em_cached)
2808 {
2809         struct extent_map *em;
2810
2811         if (em_cached && *em_cached) {
2812                 em = *em_cached;
2813                 if (extent_map_in_tree(em) && start >= em->start &&
2814                     start < extent_map_end(em)) {
2815                         atomic_inc(&em->refs);
2816                         return em;
2817                 }
2818
2819                 free_extent_map(em);
2820                 *em_cached = NULL;
2821         }
2822
2823         em = get_extent(inode, page, pg_offset, start, len, 0);
2824         if (em_cached && !IS_ERR_OR_NULL(em)) {
2825                 BUG_ON(*em_cached);
2826                 atomic_inc(&em->refs);
2827                 *em_cached = em;
2828         }
2829         return em;
2830 }
2831 /*
2832  * basic readpage implementation.  Locked extent state structs are inserted
2833  * into the tree that are removed when the IO is done (by the end_io
2834  * handlers)
2835  * XXX JDM: This needs looking at to ensure proper page locking
2836  */
2837 static int __do_readpage(struct extent_io_tree *tree,
2838                          struct page *page,
2839                          get_extent_t *get_extent,
2840                          struct extent_map **em_cached,
2841                          struct bio **bio, int mirror_num,
2842                          unsigned long *bio_flags, int rw)
2843 {
2844         struct inode *inode = page->mapping->host;
2845         u64 start = page_offset(page);
2846         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2847         u64 end;
2848         u64 cur = start;
2849         u64 extent_offset;
2850         u64 last_byte = i_size_read(inode);
2851         u64 block_start;
2852         u64 cur_end;
2853         sector_t sector;
2854         struct extent_map *em;
2855         struct block_device *bdev;
2856         int ret;
2857         int nr = 0;
2858         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2859         size_t pg_offset = 0;
2860         size_t iosize;
2861         size_t disk_io_size;
2862         size_t blocksize = inode->i_sb->s_blocksize;
2863         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2864
2865         set_page_extent_mapped(page);
2866
2867         end = page_end;
2868         if (!PageUptodate(page)) {
2869                 if (cleancache_get_page(page) == 0) {
2870                         BUG_ON(blocksize != PAGE_SIZE);
2871                         unlock_extent(tree, start, end);
2872                         goto out;
2873                 }
2874         }
2875
2876         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2877                 char *userpage;
2878                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2879
2880                 if (zero_offset) {
2881                         iosize = PAGE_CACHE_SIZE - zero_offset;
2882                         userpage = kmap_atomic(page);
2883                         memset(userpage + zero_offset, 0, iosize);
2884                         flush_dcache_page(page);
2885                         kunmap_atomic(userpage);
2886                 }
2887         }
2888         while (cur <= end) {
2889                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2890
2891                 if (cur >= last_byte) {
2892                         char *userpage;
2893                         struct extent_state *cached = NULL;
2894
2895                         iosize = PAGE_CACHE_SIZE - pg_offset;
2896                         userpage = kmap_atomic(page);
2897                         memset(userpage + pg_offset, 0, iosize);
2898                         flush_dcache_page(page);
2899                         kunmap_atomic(userpage);
2900                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2901                                             &cached, GFP_NOFS);
2902                         if (!parent_locked)
2903                                 unlock_extent_cached(tree, cur,
2904                                                      cur + iosize - 1,
2905                                                      &cached, GFP_NOFS);
2906                         break;
2907                 }
2908                 em = __get_extent_map(inode, page, pg_offset, cur,
2909                                       end - cur + 1, get_extent, em_cached);
2910                 if (IS_ERR_OR_NULL(em)) {
2911                         SetPageError(page);
2912                         if (!parent_locked)
2913                                 unlock_extent(tree, cur, end);
2914                         break;
2915                 }
2916                 extent_offset = cur - em->start;
2917                 BUG_ON(extent_map_end(em) <= cur);
2918                 BUG_ON(end < cur);
2919
2920                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2921                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2922                         extent_set_compress_type(&this_bio_flag,
2923                                                  em->compress_type);
2924                 }
2925
2926                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2927                 cur_end = min(extent_map_end(em) - 1, end);
2928                 iosize = ALIGN(iosize, blocksize);
2929                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2930                         disk_io_size = em->block_len;
2931                         sector = em->block_start >> 9;
2932                 } else {
2933                         sector = (em->block_start + extent_offset) >> 9;
2934                         disk_io_size = iosize;
2935                 }
2936                 bdev = em->bdev;
2937                 block_start = em->block_start;
2938                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2939                         block_start = EXTENT_MAP_HOLE;
2940                 free_extent_map(em);
2941                 em = NULL;
2942
2943                 /* we've found a hole, just zero and go on */
2944                 if (block_start == EXTENT_MAP_HOLE) {
2945                         char *userpage;
2946                         struct extent_state *cached = NULL;
2947
2948                         userpage = kmap_atomic(page);
2949                         memset(userpage + pg_offset, 0, iosize);
2950                         flush_dcache_page(page);
2951                         kunmap_atomic(userpage);
2952
2953                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2954                                             &cached, GFP_NOFS);
2955                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2956                                              &cached, GFP_NOFS);
2957                         cur = cur + iosize;
2958                         pg_offset += iosize;
2959                         continue;
2960                 }
2961                 /* the get_extent function already copied into the page */
2962                 if (test_range_bit(tree, cur, cur_end,
2963                                    EXTENT_UPTODATE, 1, NULL)) {
2964                         check_page_uptodate(tree, page);
2965                         if (!parent_locked)
2966                                 unlock_extent(tree, cur, cur + iosize - 1);
2967                         cur = cur + iosize;
2968                         pg_offset += iosize;
2969                         continue;
2970                 }
2971                 /* we have an inline extent but it didn't get marked up
2972                  * to date.  Error out
2973                  */
2974                 if (block_start == EXTENT_MAP_INLINE) {
2975                         SetPageError(page);
2976                         if (!parent_locked)
2977                                 unlock_extent(tree, cur, cur + iosize - 1);
2978                         cur = cur + iosize;
2979                         pg_offset += iosize;
2980                         continue;
2981                 }
2982
2983                 pnr -= page->index;
2984                 ret = submit_extent_page(rw, tree, page,
2985                                          sector, disk_io_size, pg_offset,
2986                                          bdev, bio, pnr,
2987                                          end_bio_extent_readpage, mirror_num,
2988                                          *bio_flags,
2989                                          this_bio_flag);
2990                 if (!ret) {
2991                         nr++;
2992                         *bio_flags = this_bio_flag;
2993                 } else {
2994                         SetPageError(page);
2995                         if (!parent_locked)
2996                                 unlock_extent(tree, cur, cur + iosize - 1);
2997                 }
2998                 cur = cur + iosize;
2999                 pg_offset += iosize;
3000         }
3001 out:
3002         if (!nr) {
3003                 if (!PageError(page))
3004                         SetPageUptodate(page);
3005                 unlock_page(page);
3006         }
3007         return 0;
3008 }
3009
3010 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3011                                              struct page *pages[], int nr_pages,
3012                                              u64 start, u64 end,
3013                                              get_extent_t *get_extent,
3014                                              struct extent_map **em_cached,
3015                                              struct bio **bio, int mirror_num,
3016                                              unsigned long *bio_flags, int rw)
3017 {
3018         struct inode *inode;
3019         struct btrfs_ordered_extent *ordered;
3020         int index;
3021
3022         inode = pages[0]->mapping->host;
3023         while (1) {
3024                 lock_extent(tree, start, end);
3025                 ordered = btrfs_lookup_ordered_range(inode, start,
3026                                                      end - start + 1);
3027                 if (!ordered)
3028                         break;
3029                 unlock_extent(tree, start, end);
3030                 btrfs_start_ordered_extent(inode, ordered, 1);
3031                 btrfs_put_ordered_extent(ordered);
3032         }
3033
3034         for (index = 0; index < nr_pages; index++) {
3035                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3036                               mirror_num, bio_flags, rw);
3037                 page_cache_release(pages[index]);
3038         }
3039 }
3040
3041 static void __extent_readpages(struct extent_io_tree *tree,
3042                                struct page *pages[],
3043                                int nr_pages, get_extent_t *get_extent,
3044                                struct extent_map **em_cached,
3045                                struct bio **bio, int mirror_num,
3046                                unsigned long *bio_flags, int rw)
3047 {
3048         u64 start = 0;
3049         u64 end = 0;
3050         u64 page_start;
3051         int index;
3052         int first_index = 0;
3053
3054         for (index = 0; index < nr_pages; index++) {
3055                 page_start = page_offset(pages[index]);
3056                 if (!end) {
3057                         start = page_start;
3058                         end = start + PAGE_CACHE_SIZE - 1;
3059                         first_index = index;
3060                 } else if (end + 1 == page_start) {
3061                         end += PAGE_CACHE_SIZE;
3062                 } else {
3063                         __do_contiguous_readpages(tree, &pages[first_index],
3064                                                   index - first_index, start,
3065                                                   end, get_extent, em_cached,
3066                                                   bio, mirror_num, bio_flags,
3067                                                   rw);
3068                         start = page_start;
3069                         end = start + PAGE_CACHE_SIZE - 1;
3070                         first_index = index;
3071                 }
3072         }
3073
3074         if (end)
3075                 __do_contiguous_readpages(tree, &pages[first_index],
3076                                           index - first_index, start,
3077                                           end, get_extent, em_cached, bio,
3078                                           mirror_num, bio_flags, rw);
3079 }
3080
3081 static int __extent_read_full_page(struct extent_io_tree *tree,
3082                                    struct page *page,
3083                                    get_extent_t *get_extent,
3084                                    struct bio **bio, int mirror_num,
3085                                    unsigned long *bio_flags, int rw)
3086 {
3087         struct inode *inode = page->mapping->host;
3088         struct btrfs_ordered_extent *ordered;
3089         u64 start = page_offset(page);
3090         u64 end = start + PAGE_CACHE_SIZE - 1;
3091         int ret;
3092
3093         while (1) {
3094                 lock_extent(tree, start, end);
3095                 ordered = btrfs_lookup_ordered_extent(inode, start);
3096                 if (!ordered)
3097                         break;
3098                 unlock_extent(tree, start, end);
3099                 btrfs_start_ordered_extent(inode, ordered, 1);
3100                 btrfs_put_ordered_extent(ordered);
3101         }
3102
3103         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3104                             bio_flags, rw);
3105         return ret;
3106 }
3107
3108 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3109                             get_extent_t *get_extent, int mirror_num)
3110 {
3111         struct bio *bio = NULL;
3112         unsigned long bio_flags = 0;
3113         int ret;
3114
3115         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3116                                       &bio_flags, READ);
3117         if (bio)
3118                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3119         return ret;
3120 }
3121
3122 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3123                                  get_extent_t *get_extent, int mirror_num)
3124 {
3125         struct bio *bio = NULL;
3126         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3127         int ret;
3128
3129         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3130                                       &bio_flags, READ);
3131         if (bio)
3132                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3133         return ret;
3134 }
3135
3136 static noinline void update_nr_written(struct page *page,
3137                                       struct writeback_control *wbc,
3138                                       unsigned long nr_written)
3139 {
3140         wbc->nr_to_write -= nr_written;
3141         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3142             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3143                 page->mapping->writeback_index = page->index + nr_written;
3144 }
3145
3146 /*
3147  * helper for __extent_writepage, doing all of the delayed allocation setup.
3148  *
3149  * This returns 1 if our fill_delalloc function did all the work required
3150  * to write the page (copy into inline extent).  In this case the IO has
3151  * been started and the page is already unlocked.
3152  *
3153  * This returns 0 if all went well (page still locked)
3154  * This returns < 0 if there were errors (page still locked)
3155  */
3156 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3157                               struct page *page, struct writeback_control *wbc,
3158                               struct extent_page_data *epd,
3159                               u64 delalloc_start,
3160                               unsigned long *nr_written)
3161 {
3162         struct extent_io_tree *tree = epd->tree;
3163         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3164         u64 nr_delalloc;
3165         u64 delalloc_to_write = 0;
3166         u64 delalloc_end = 0;
3167         int ret;
3168         int page_started = 0;
3169
3170         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3171                 return 0;
3172
3173         while (delalloc_end < page_end) {
3174                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3175                                                page,
3176                                                &delalloc_start,
3177                                                &delalloc_end,
3178                                                128 * 1024 * 1024);
3179                 if (nr_delalloc == 0) {
3180                         delalloc_start = delalloc_end + 1;
3181                         continue;
3182                 }
3183                 ret = tree->ops->fill_delalloc(inode, page,
3184                                                delalloc_start,
3185                                                delalloc_end,
3186                                                &page_started,
3187                                                nr_written);
3188                 /* File system has been set read-only */
3189                 if (ret) {
3190                         SetPageError(page);
3191                         /* fill_delalloc should be return < 0 for error
3192                          * but just in case, we use > 0 here meaning the
3193                          * IO is started, so we don't want to return > 0
3194                          * unless things are going well.
3195                          */
3196                         ret = ret < 0 ? ret : -EIO;
3197                         goto done;
3198                 }
3199                 /*
3200                  * delalloc_end is already one less than the total
3201                  * length, so we don't subtract one from
3202                  * PAGE_CACHE_SIZE
3203                  */
3204                 delalloc_to_write += (delalloc_end - delalloc_start +
3205                                       PAGE_CACHE_SIZE) >>
3206                                       PAGE_CACHE_SHIFT;
3207                 delalloc_start = delalloc_end + 1;
3208         }
3209         if (wbc->nr_to_write < delalloc_to_write) {
3210                 int thresh = 8192;
3211
3212                 if (delalloc_to_write < thresh * 2)
3213                         thresh = delalloc_to_write;
3214                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3215                                          thresh);
3216         }
3217
3218         /* did the fill delalloc function already unlock and start
3219          * the IO?
3220          */
3221         if (page_started) {
3222                 /*
3223                  * we've unlocked the page, so we can't update
3224                  * the mapping's writeback index, just update
3225                  * nr_to_write.
3226                  */
3227                 wbc->nr_to_write -= *nr_written;
3228                 return 1;
3229         }
3230
3231         ret = 0;
3232
3233 done:
3234         return ret;
3235 }
3236
3237 /*
3238  * helper for __extent_writepage.  This calls the writepage start hooks,
3239  * and does the loop to map the page into extents and bios.
3240  *
3241  * We return 1 if the IO is started and the page is unlocked,
3242  * 0 if all went well (page still locked)
3243  * < 0 if there were errors (page still locked)
3244  */
3245 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3246                                  struct page *page,
3247                                  struct writeback_control *wbc,
3248                                  struct extent_page_data *epd,
3249                                  loff_t i_size,
3250                                  unsigned long nr_written,
3251                                  int write_flags, int *nr_ret)
3252 {
3253         struct extent_io_tree *tree = epd->tree;
3254         u64 start = page_offset(page);
3255         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3256         u64 end;
3257         u64 cur = start;
3258         u64 extent_offset;
3259         u64 block_start;
3260         u64 iosize;
3261         sector_t sector;
3262         struct extent_state *cached_state = NULL;
3263         struct extent_map *em;
3264         struct block_device *bdev;
3265         size_t pg_offset = 0;
3266         size_t blocksize;
3267         int ret = 0;
3268         int nr = 0;
3269         bool compressed;
3270
3271         if (tree->ops && tree->ops->writepage_start_hook) {
3272                 ret = tree->ops->writepage_start_hook(page, start,
3273                                                       page_end);
3274                 if (ret) {
3275                         /* Fixup worker will requeue */
3276                         if (ret == -EBUSY)
3277                                 wbc->pages_skipped++;
3278                         else
3279                                 redirty_page_for_writepage(wbc, page);
3280
3281                         update_nr_written(page, wbc, nr_written);
3282                         unlock_page(page);
3283                         ret = 1;
3284                         goto done_unlocked;
3285                 }
3286         }
3287
3288         /*
3289          * we don't want to touch the inode after unlocking the page,
3290          * so we update the mapping writeback index now
3291          */
3292         update_nr_written(page, wbc, nr_written + 1);
3293
3294         end = page_end;
3295         if (i_size <= start) {
3296                 if (tree->ops && tree->ops->writepage_end_io_hook)
3297                         tree->ops->writepage_end_io_hook(page, start,
3298                                                          page_end, NULL, 1);
3299                 goto done;
3300         }
3301
3302         blocksize = inode->i_sb->s_blocksize;
3303
3304         while (cur <= end) {
3305                 u64 em_end;
3306                 if (cur >= i_size) {
3307                         if (tree->ops && tree->ops->writepage_end_io_hook)
3308                                 tree->ops->writepage_end_io_hook(page, cur,
3309                                                          page_end, NULL, 1);
3310                         break;
3311                 }
3312                 em = epd->get_extent(inode, page, pg_offset, cur,
3313                                      end - cur + 1, 1);
3314                 if (IS_ERR_OR_NULL(em)) {
3315                         SetPageError(page);
3316                         ret = PTR_ERR_OR_ZERO(em);
3317                         break;
3318                 }
3319
3320                 extent_offset = cur - em->start;
3321                 em_end = extent_map_end(em);
3322                 BUG_ON(em_end <= cur);
3323                 BUG_ON(end < cur);
3324                 iosize = min(em_end - cur, end - cur + 1);
3325                 iosize = ALIGN(iosize, blocksize);
3326                 sector = (em->block_start + extent_offset) >> 9;
3327                 bdev = em->bdev;
3328                 block_start = em->block_start;
3329                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3330                 free_extent_map(em);
3331                 em = NULL;
3332
3333                 /*
3334                  * compressed and inline extents are written through other
3335                  * paths in the FS
3336                  */
3337                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3338                     block_start == EXTENT_MAP_INLINE) {
3339                         /*
3340                          * end_io notification does not happen here for
3341                          * compressed extents
3342                          */
3343                         if (!compressed && tree->ops &&
3344                             tree->ops->writepage_end_io_hook)
3345                                 tree->ops->writepage_end_io_hook(page, cur,
3346                                                          cur + iosize - 1,
3347                                                          NULL, 1);
3348                         else if (compressed) {
3349                                 /* we don't want to end_page_writeback on
3350                                  * a compressed extent.  this happens
3351                                  * elsewhere
3352                                  */
3353                                 nr++;
3354                         }
3355
3356                         cur += iosize;
3357                         pg_offset += iosize;
3358                         continue;
3359                 }
3360
3361                 if (tree->ops && tree->ops->writepage_io_hook) {
3362                         ret = tree->ops->writepage_io_hook(page, cur,
3363                                                 cur + iosize - 1);
3364                 } else {
3365                         ret = 0;
3366                 }
3367                 if (ret) {
3368                         SetPageError(page);
3369                 } else {
3370                         unsigned long max_nr = (i_size >> PAGE_CACHE_SHIFT) + 1;
3371
3372                         set_range_writeback(tree, cur, cur + iosize - 1);
3373                         if (!PageWriteback(page)) {
3374                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3375                                            "page %lu not writeback, cur %llu end %llu",
3376                                        page->index, cur, end);
3377                         }
3378
3379                         ret = submit_extent_page(write_flags, tree, page,
3380                                                  sector, iosize, pg_offset,
3381                                                  bdev, &epd->bio, max_nr,
3382                                                  end_bio_extent_writepage,
3383                                                  0, 0, 0);
3384                         if (ret)
3385                                 SetPageError(page);
3386                 }
3387                 cur = cur + iosize;
3388                 pg_offset += iosize;
3389                 nr++;
3390         }
3391 done:
3392         *nr_ret = nr;
3393
3394 done_unlocked:
3395
3396         /* drop our reference on any cached states */
3397         free_extent_state(cached_state);
3398         return ret;
3399 }
3400
3401 /*
3402  * the writepage semantics are similar to regular writepage.  extent
3403  * records are inserted to lock ranges in the tree, and as dirty areas
3404  * are found, they are marked writeback.  Then the lock bits are removed
3405  * and the end_io handler clears the writeback ranges
3406  */
3407 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3408                               void *data)
3409 {
3410         struct inode *inode = page->mapping->host;
3411         struct extent_page_data *epd = data;
3412         u64 start = page_offset(page);
3413         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3414         int ret;
3415         int nr = 0;
3416         size_t pg_offset = 0;
3417         loff_t i_size = i_size_read(inode);
3418         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3419         int write_flags;
3420         unsigned long nr_written = 0;
3421
3422         if (wbc->sync_mode == WB_SYNC_ALL)
3423                 write_flags = WRITE_SYNC;
3424         else
3425                 write_flags = WRITE;
3426
3427         trace___extent_writepage(page, inode, wbc);
3428
3429         WARN_ON(!PageLocked(page));
3430
3431         ClearPageError(page);
3432
3433         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3434         if (page->index > end_index ||
3435            (page->index == end_index && !pg_offset)) {
3436                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3437                 unlock_page(page);
3438                 return 0;
3439         }
3440
3441         if (page->index == end_index) {
3442                 char *userpage;
3443
3444                 userpage = kmap_atomic(page);
3445                 memset(userpage + pg_offset, 0,
3446                        PAGE_CACHE_SIZE - pg_offset);
3447                 kunmap_atomic(userpage);
3448                 flush_dcache_page(page);
3449         }
3450
3451         pg_offset = 0;
3452
3453         set_page_extent_mapped(page);
3454
3455         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3456         if (ret == 1)
3457                 goto done_unlocked;
3458         if (ret)
3459                 goto done;
3460
3461         ret = __extent_writepage_io(inode, page, wbc, epd,
3462                                     i_size, nr_written, write_flags, &nr);
3463         if (ret == 1)
3464                 goto done_unlocked;
3465
3466 done:
3467         if (nr == 0) {
3468                 /* make sure the mapping tag for page dirty gets cleared */
3469                 set_page_writeback(page);
3470                 end_page_writeback(page);
3471         }
3472         if (PageError(page)) {
3473                 ret = ret < 0 ? ret : -EIO;
3474                 end_extent_writepage(page, ret, start, page_end);
3475         }
3476         unlock_page(page);
3477         return ret;
3478
3479 done_unlocked:
3480         return 0;
3481 }
3482
3483 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3484 {
3485         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3486                        TASK_UNINTERRUPTIBLE);
3487 }
3488
3489 static noinline_for_stack int
3490 lock_extent_buffer_for_io(struct extent_buffer *eb,
3491                           struct btrfs_fs_info *fs_info,
3492                           struct extent_page_data *epd)
3493 {
3494         unsigned long i, num_pages;
3495         int flush = 0;
3496         int ret = 0;
3497
3498         if (!btrfs_try_tree_write_lock(eb)) {
3499                 flush = 1;
3500                 flush_write_bio(epd);
3501                 btrfs_tree_lock(eb);
3502         }
3503
3504         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3505                 btrfs_tree_unlock(eb);
3506                 if (!epd->sync_io)
3507                         return 0;
3508                 if (!flush) {
3509                         flush_write_bio(epd);
3510                         flush = 1;
3511                 }
3512                 while (1) {
3513                         wait_on_extent_buffer_writeback(eb);
3514                         btrfs_tree_lock(eb);
3515                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3516                                 break;
3517                         btrfs_tree_unlock(eb);
3518                 }
3519         }
3520
3521         /*
3522          * We need to do this to prevent races in people who check if the eb is
3523          * under IO since we can end up having no IO bits set for a short period
3524          * of time.
3525          */
3526         spin_lock(&eb->refs_lock);
3527         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3528                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3529                 spin_unlock(&eb->refs_lock);
3530                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3531                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3532                                      -eb->len,
3533                                      fs_info->dirty_metadata_batch);
3534                 ret = 1;
3535         } else {
3536                 spin_unlock(&eb->refs_lock);
3537         }
3538
3539         btrfs_tree_unlock(eb);
3540
3541         if (!ret)
3542                 return ret;
3543
3544         num_pages = num_extent_pages(eb->start, eb->len);
3545         for (i = 0; i < num_pages; i++) {
3546                 struct page *p = extent_buffer_page(eb, i);
3547
3548                 if (!trylock_page(p)) {
3549                         if (!flush) {
3550                                 flush_write_bio(epd);
3551                                 flush = 1;
3552                         }
3553                         lock_page(p);
3554                 }
3555         }
3556
3557         return ret;
3558 }
3559
3560 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3561 {
3562         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3563         smp_mb__after_atomic();
3564         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3565 }
3566
3567 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3568 {
3569         struct bio_vec *bvec;
3570         struct extent_buffer *eb;
3571         int i, done;
3572
3573         bio_for_each_segment_all(bvec, bio, i) {
3574                 struct page *page = bvec->bv_page;
3575
3576                 eb = (struct extent_buffer *)page->private;
3577                 BUG_ON(!eb);
3578                 done = atomic_dec_and_test(&eb->io_pages);
3579
3580                 if (err || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3581                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3582                         ClearPageUptodate(page);
3583                         SetPageError(page);
3584                 }
3585
3586                 end_page_writeback(page);
3587
3588                 if (!done)
3589                         continue;
3590
3591                 end_extent_buffer_writeback(eb);
3592         }
3593
3594         bio_put(bio);
3595 }
3596
3597 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3598                         struct btrfs_fs_info *fs_info,
3599                         struct writeback_control *wbc,
3600                         struct extent_page_data *epd)
3601 {
3602         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3603         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3604         u64 offset = eb->start;
3605         unsigned long i, num_pages;
3606         unsigned long bio_flags = 0;
3607         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3608         int ret = 0;
3609
3610         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3611         num_pages = num_extent_pages(eb->start, eb->len);
3612         atomic_set(&eb->io_pages, num_pages);
3613         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3614                 bio_flags = EXTENT_BIO_TREE_LOG;
3615
3616         for (i = 0; i < num_pages; i++) {
3617                 struct page *p = extent_buffer_page(eb, i);
3618
3619                 clear_page_dirty_for_io(p);
3620                 set_page_writeback(p);
3621                 ret = submit_extent_page(rw, tree, p, offset >> 9,
3622                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3623                                          -1, end_bio_extent_buffer_writepage,
3624                                          0, epd->bio_flags, bio_flags);
3625                 epd->bio_flags = bio_flags;
3626                 if (ret) {
3627                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3628                         SetPageError(p);
3629                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3630                                 end_extent_buffer_writeback(eb);
3631                         ret = -EIO;
3632                         break;
3633                 }
3634                 offset += PAGE_CACHE_SIZE;
3635                 update_nr_written(p, wbc, 1);
3636                 unlock_page(p);
3637         }
3638
3639         if (unlikely(ret)) {
3640                 for (; i < num_pages; i++) {
3641                         struct page *p = extent_buffer_page(eb, i);
3642                         unlock_page(p);
3643                 }
3644         }
3645
3646         return ret;
3647 }
3648
3649 int btree_write_cache_pages(struct address_space *mapping,
3650                                    struct writeback_control *wbc)
3651 {
3652         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3653         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3654         struct extent_buffer *eb, *prev_eb = NULL;
3655         struct extent_page_data epd = {
3656                 .bio = NULL,
3657                 .tree = tree,
3658                 .extent_locked = 0,
3659                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3660                 .bio_flags = 0,
3661         };
3662         int ret = 0;
3663         int done = 0;
3664         int nr_to_write_done = 0;
3665         struct pagevec pvec;
3666         int nr_pages;
3667         pgoff_t index;
3668         pgoff_t end;            /* Inclusive */
3669         int scanned = 0;
3670         int tag;
3671
3672         pagevec_init(&pvec, 0);
3673         if (wbc->range_cyclic) {
3674                 index = mapping->writeback_index; /* Start from prev offset */
3675                 end = -1;
3676         } else {
3677                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3678                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3679                 scanned = 1;
3680         }
3681         if (wbc->sync_mode == WB_SYNC_ALL)
3682                 tag = PAGECACHE_TAG_TOWRITE;
3683         else
3684                 tag = PAGECACHE_TAG_DIRTY;
3685 retry:
3686         if (wbc->sync_mode == WB_SYNC_ALL)
3687                 tag_pages_for_writeback(mapping, index, end);
3688         while (!done && !nr_to_write_done && (index <= end) &&
3689                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3690                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3691                 unsigned i;
3692
3693                 scanned = 1;
3694                 for (i = 0; i < nr_pages; i++) {
3695                         struct page *page = pvec.pages[i];
3696
3697                         if (!PagePrivate(page))
3698                                 continue;
3699
3700                         if (!wbc->range_cyclic && page->index > end) {
3701                                 done = 1;
3702                                 break;
3703                         }
3704
3705                         spin_lock(&mapping->private_lock);
3706                         if (!PagePrivate(page)) {
3707                                 spin_unlock(&mapping->private_lock);
3708                                 continue;
3709                         }
3710
3711                         eb = (struct extent_buffer *)page->private;
3712
3713                         /*
3714                          * Shouldn't happen and normally this would be a BUG_ON
3715                          * but no sense in crashing the users box for something
3716                          * we can survive anyway.
3717                          */
3718                         if (WARN_ON(!eb)) {
3719                                 spin_unlock(&mapping->private_lock);
3720                                 continue;
3721                         }
3722
3723                         if (eb == prev_eb) {
3724                                 spin_unlock(&mapping->private_lock);
3725                                 continue;
3726                         }
3727
3728                         ret = atomic_inc_not_zero(&eb->refs);
3729                         spin_unlock(&mapping->private_lock);
3730                         if (!ret)
3731                                 continue;
3732
3733                         prev_eb = eb;
3734                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3735                         if (!ret) {
3736                                 free_extent_buffer(eb);
3737                                 continue;
3738                         }
3739
3740                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3741                         if (ret) {
3742                                 done = 1;
3743                                 free_extent_buffer(eb);
3744                                 break;
3745                         }
3746                         free_extent_buffer(eb);
3747
3748                         /*
3749                          * the filesystem may choose to bump up nr_to_write.
3750                          * We have to make sure to honor the new nr_to_write
3751                          * at any time
3752                          */
3753                         nr_to_write_done = wbc->nr_to_write <= 0;
3754                 }
3755                 pagevec_release(&pvec);
3756                 cond_resched();
3757         }
3758         if (!scanned && !done) {
3759                 /*
3760                  * We hit the last page and there is more work to be done: wrap
3761                  * back to the start of the file
3762                  */
3763                 scanned = 1;
3764                 index = 0;
3765                 goto retry;
3766         }
3767         flush_write_bio(&epd);
3768         return ret;
3769 }
3770
3771 /**
3772  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3773  * @mapping: address space structure to write
3774  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3775  * @writepage: function called for each page
3776  * @data: data passed to writepage function
3777  *
3778  * If a page is already under I/O, write_cache_pages() skips it, even
3779  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3780  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3781  * and msync() need to guarantee that all the data which was dirty at the time
3782  * the call was made get new I/O started against them.  If wbc->sync_mode is
3783  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3784  * existing IO to complete.
3785  */
3786 static int extent_write_cache_pages(struct extent_io_tree *tree,
3787                              struct address_space *mapping,
3788                              struct writeback_control *wbc,
3789                              writepage_t writepage, void *data,
3790                              void (*flush_fn)(void *))
3791 {
3792         struct inode *inode = mapping->host;
3793         int ret = 0;
3794         int done = 0;
3795         int err = 0;
3796         int nr_to_write_done = 0;
3797         struct pagevec pvec;
3798         int nr_pages;
3799         pgoff_t index;
3800         pgoff_t end;            /* Inclusive */
3801         int scanned = 0;
3802         int tag;
3803
3804         /*
3805          * We have to hold onto the inode so that ordered extents can do their
3806          * work when the IO finishes.  The alternative to this is failing to add
3807          * an ordered extent if the igrab() fails there and that is a huge pain
3808          * to deal with, so instead just hold onto the inode throughout the
3809          * writepages operation.  If it fails here we are freeing up the inode
3810          * anyway and we'd rather not waste our time writing out stuff that is
3811          * going to be truncated anyway.
3812          */
3813         if (!igrab(inode))
3814                 return 0;
3815
3816         pagevec_init(&pvec, 0);
3817         if (wbc->range_cyclic) {
3818                 index = mapping->writeback_index; /* Start from prev offset */
3819                 end = -1;
3820         } else {
3821                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3822                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3823                 scanned = 1;
3824         }
3825         if (wbc->sync_mode == WB_SYNC_ALL)
3826                 tag = PAGECACHE_TAG_TOWRITE;
3827         else
3828                 tag = PAGECACHE_TAG_DIRTY;
3829 retry:
3830         if (wbc->sync_mode == WB_SYNC_ALL)
3831                 tag_pages_for_writeback(mapping, index, end);
3832         while (!done && !nr_to_write_done && (index <= end) &&
3833                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3834                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3835                 unsigned i;
3836
3837                 scanned = 1;
3838                 for (i = 0; i < nr_pages; i++) {
3839                         struct page *page = pvec.pages[i];
3840
3841                         /*
3842                          * At this point we hold neither mapping->tree_lock nor
3843                          * lock on the page itself: the page may be truncated or
3844                          * invalidated (changing page->mapping to NULL), or even
3845                          * swizzled back from swapper_space to tmpfs file
3846                          * mapping
3847                          */
3848                         if (!trylock_page(page)) {
3849                                 flush_fn(data);
3850                                 lock_page(page);
3851                         }
3852
3853                         if (unlikely(page->mapping != mapping)) {
3854                                 unlock_page(page);
3855                                 continue;
3856                         }
3857
3858                         if (!wbc->range_cyclic && page->index > end) {
3859                                 done = 1;
3860                                 unlock_page(page);
3861                                 continue;
3862                         }
3863
3864                         if (wbc->sync_mode != WB_SYNC_NONE) {
3865                                 if (PageWriteback(page))
3866                                         flush_fn(data);
3867                                 wait_on_page_writeback(page);
3868                         }
3869
3870                         if (PageWriteback(page) ||
3871                             !clear_page_dirty_for_io(page)) {
3872                                 unlock_page(page);
3873                                 continue;
3874                         }
3875
3876                         ret = (*writepage)(page, wbc, data);
3877
3878                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3879                                 unlock_page(page);
3880                                 ret = 0;
3881                         }
3882                         if (!err && ret < 0)
3883                                 err = ret;
3884
3885                         /*
3886                          * the filesystem may choose to bump up nr_to_write.
3887                          * We have to make sure to honor the new nr_to_write
3888                          * at any time
3889                          */
3890                         nr_to_write_done = wbc->nr_to_write <= 0;
3891                 }
3892                 pagevec_release(&pvec);
3893                 cond_resched();
3894         }
3895         if (!scanned && !done && !err) {
3896                 /*
3897                  * We hit the last page and there is more work to be done: wrap
3898                  * back to the start of the file
3899                  */
3900                 scanned = 1;
3901                 index = 0;
3902                 goto retry;
3903         }
3904         btrfs_add_delayed_iput(inode);
3905         return err;
3906 }
3907
3908 static void flush_epd_write_bio(struct extent_page_data *epd)
3909 {
3910         if (epd->bio) {
3911                 int rw = WRITE;
3912                 int ret;
3913
3914                 if (epd->sync_io)
3915                         rw = WRITE_SYNC;
3916
3917                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3918                 BUG_ON(ret < 0); /* -ENOMEM */
3919                 epd->bio = NULL;
3920         }
3921 }
3922
3923 static noinline void flush_write_bio(void *data)
3924 {
3925         struct extent_page_data *epd = data;
3926         flush_epd_write_bio(epd);
3927 }
3928
3929 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3930                           get_extent_t *get_extent,
3931                           struct writeback_control *wbc)
3932 {
3933         int ret;
3934         struct extent_page_data epd = {
3935                 .bio = NULL,
3936                 .tree = tree,
3937                 .get_extent = get_extent,
3938                 .extent_locked = 0,
3939                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3940                 .bio_flags = 0,
3941         };
3942
3943         ret = __extent_writepage(page, wbc, &epd);
3944
3945         flush_epd_write_bio(&epd);
3946         return ret;
3947 }
3948
3949 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3950                               u64 start, u64 end, get_extent_t *get_extent,
3951                               int mode)
3952 {
3953         int ret = 0;
3954         struct address_space *mapping = inode->i_mapping;
3955         struct page *page;
3956         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3957                 PAGE_CACHE_SHIFT;
3958
3959         struct extent_page_data epd = {
3960                 .bio = NULL,
3961                 .tree = tree,
3962                 .get_extent = get_extent,
3963                 .extent_locked = 1,
3964                 .sync_io = mode == WB_SYNC_ALL,
3965                 .bio_flags = 0,
3966         };
3967         struct writeback_control wbc_writepages = {
3968                 .sync_mode      = mode,
3969                 .nr_to_write    = nr_pages * 2,
3970                 .range_start    = start,
3971                 .range_end      = end + 1,
3972         };
3973
3974         while (start <= end) {
3975                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3976                 if (clear_page_dirty_for_io(page))
3977                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3978                 else {
3979                         if (tree->ops && tree->ops->writepage_end_io_hook)
3980                                 tree->ops->writepage_end_io_hook(page, start,
3981                                                  start + PAGE_CACHE_SIZE - 1,
3982                                                  NULL, 1);
3983                         unlock_page(page);
3984                 }
3985                 page_cache_release(page);
3986                 start += PAGE_CACHE_SIZE;
3987         }
3988
3989         flush_epd_write_bio(&epd);
3990         return ret;
3991 }
3992
3993 int extent_writepages(struct extent_io_tree *tree,
3994                       struct address_space *mapping,
3995                       get_extent_t *get_extent,
3996                       struct writeback_control *wbc)
3997 {
3998         int ret = 0;
3999         struct extent_page_data epd = {
4000                 .bio = NULL,
4001                 .tree = tree,
4002                 .get_extent = get_extent,
4003                 .extent_locked = 0,
4004                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4005                 .bio_flags = 0,
4006         };
4007
4008         ret = extent_write_cache_pages(tree, mapping, wbc,
4009                                        __extent_writepage, &epd,
4010                                        flush_write_bio);
4011         flush_epd_write_bio(&epd);
4012         return ret;
4013 }
4014
4015 int extent_readpages(struct extent_io_tree *tree,
4016                      struct address_space *mapping,
4017                      struct list_head *pages, unsigned nr_pages,
4018                      get_extent_t get_extent)
4019 {
4020         struct bio *bio = NULL;
4021         unsigned page_idx;
4022         unsigned long bio_flags = 0;
4023         struct page *pagepool[16];
4024         struct page *page;
4025         struct extent_map *em_cached = NULL;
4026         int nr = 0;
4027
4028         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4029                 page = list_entry(pages->prev, struct page, lru);
4030
4031                 prefetchw(&page->flags);
4032                 list_del(&page->lru);
4033                 if (add_to_page_cache_lru(page, mapping,
4034                                         page->index, GFP_NOFS)) {
4035                         page_cache_release(page);
4036                         continue;
4037                 }
4038
4039                 pagepool[nr++] = page;
4040                 if (nr < ARRAY_SIZE(pagepool))
4041                         continue;
4042                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4043                                    &bio, 0, &bio_flags, READ);
4044                 nr = 0;
4045         }
4046         if (nr)
4047                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4048                                    &bio, 0, &bio_flags, READ);
4049
4050         if (em_cached)
4051                 free_extent_map(em_cached);
4052
4053         BUG_ON(!list_empty(pages));
4054         if (bio)
4055                 return submit_one_bio(READ, bio, 0, bio_flags);
4056         return 0;
4057 }
4058
4059 /*
4060  * basic invalidatepage code, this waits on any locked or writeback
4061  * ranges corresponding to the page, and then deletes any extent state
4062  * records from the tree
4063  */
4064 int extent_invalidatepage(struct extent_io_tree *tree,
4065                           struct page *page, unsigned long offset)
4066 {
4067         struct extent_state *cached_state = NULL;
4068         u64 start = page_offset(page);
4069         u64 end = start + PAGE_CACHE_SIZE - 1;
4070         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4071
4072         start += ALIGN(offset, blocksize);
4073         if (start > end)
4074                 return 0;
4075
4076         lock_extent_bits(tree, start, end, 0, &cached_state);
4077         wait_on_page_writeback(page);
4078         clear_extent_bit(tree, start, end,
4079                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4080                          EXTENT_DO_ACCOUNTING,
4081                          1, 1, &cached_state, GFP_NOFS);
4082         return 0;
4083 }
4084
4085 /*
4086  * a helper for releasepage, this tests for areas of the page that
4087  * are locked or under IO and drops the related state bits if it is safe
4088  * to drop the page.
4089  */
4090 static int try_release_extent_state(struct extent_map_tree *map,
4091                                     struct extent_io_tree *tree,
4092                                     struct page *page, gfp_t mask)
4093 {
4094         u64 start = page_offset(page);
4095         u64 end = start + PAGE_CACHE_SIZE - 1;
4096         int ret = 1;
4097
4098         if (test_range_bit(tree, start, end,
4099                            EXTENT_IOBITS, 0, NULL))
4100                 ret = 0;
4101         else {
4102                 if ((mask & GFP_NOFS) == GFP_NOFS)
4103                         mask = GFP_NOFS;
4104                 /*
4105                  * at this point we can safely clear everything except the
4106                  * locked bit and the nodatasum bit
4107                  */
4108                 ret = clear_extent_bit(tree, start, end,
4109                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4110                                  0, 0, NULL, mask);
4111
4112                 /* if clear_extent_bit failed for enomem reasons,
4113                  * we can't allow the release to continue.
4114                  */
4115                 if (ret < 0)
4116                         ret = 0;
4117                 else
4118                         ret = 1;
4119         }
4120         return ret;
4121 }
4122
4123 /*
4124  * a helper for releasepage.  As long as there are no locked extents
4125  * in the range corresponding to the page, both state records and extent
4126  * map records are removed
4127  */
4128 int try_release_extent_mapping(struct extent_map_tree *map,
4129                                struct extent_io_tree *tree, struct page *page,
4130                                gfp_t mask)
4131 {
4132         struct extent_map *em;
4133         u64 start = page_offset(page);
4134         u64 end = start + PAGE_CACHE_SIZE - 1;
4135
4136         if ((mask & __GFP_WAIT) &&
4137             page->mapping->host->i_size > 16 * 1024 * 1024) {
4138                 u64 len;
4139                 while (start <= end) {
4140                         len = end - start + 1;
4141                         write_lock(&map->lock);
4142                         em = lookup_extent_mapping(map, start, len);
4143                         if (!em) {
4144                                 write_unlock(&map->lock);
4145                                 break;
4146                         }
4147                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4148                             em->start != start) {
4149                                 write_unlock(&map->lock);
4150                                 free_extent_map(em);
4151                                 break;
4152                         }
4153                         if (!test_range_bit(tree, em->start,
4154                                             extent_map_end(em) - 1,
4155                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4156                                             0, NULL)) {
4157                                 remove_extent_mapping(map, em);
4158                                 /* once for the rb tree */
4159                                 free_extent_map(em);
4160                         }
4161                         start = extent_map_end(em);
4162                         write_unlock(&map->lock);
4163
4164                         /* once for us */
4165                         free_extent_map(em);
4166                 }
4167         }
4168         return try_release_extent_state(map, tree, page, mask);
4169 }
4170
4171 /*
4172  * helper function for fiemap, which doesn't want to see any holes.
4173  * This maps until we find something past 'last'
4174  */
4175 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4176                                                 u64 offset,
4177                                                 u64 last,
4178                                                 get_extent_t *get_extent)
4179 {
4180         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4181         struct extent_map *em;
4182         u64 len;
4183
4184         if (offset >= last)
4185                 return NULL;
4186
4187         while (1) {
4188                 len = last - offset;
4189                 if (len == 0)
4190                         break;
4191                 len = ALIGN(len, sectorsize);
4192                 em = get_extent(inode, NULL, 0, offset, len, 0);
4193                 if (IS_ERR_OR_NULL(em))
4194                         return em;
4195
4196                 /* if this isn't a hole return it */
4197                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4198                     em->block_start != EXTENT_MAP_HOLE) {
4199                         return em;
4200                 }
4201
4202                 /* this is a hole, advance to the next extent */
4203                 offset = extent_map_end(em);
4204                 free_extent_map(em);
4205                 if (offset >= last)
4206                         break;
4207         }
4208         return NULL;
4209 }
4210
4211 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4212                 __u64 start, __u64 len, get_extent_t *get_extent)
4213 {
4214         int ret = 0;
4215         u64 off = start;
4216         u64 max = start + len;
4217         u32 flags = 0;
4218         u32 found_type;
4219         u64 last;
4220         u64 last_for_get_extent = 0;
4221         u64 disko = 0;
4222         u64 isize = i_size_read(inode);
4223         struct btrfs_key found_key;
4224         struct extent_map *em = NULL;
4225         struct extent_state *cached_state = NULL;
4226         struct btrfs_path *path;
4227         struct btrfs_root *root = BTRFS_I(inode)->root;
4228         int end = 0;
4229         u64 em_start = 0;
4230         u64 em_len = 0;
4231         u64 em_end = 0;
4232
4233         if (len == 0)
4234                 return -EINVAL;
4235
4236         path = btrfs_alloc_path();
4237         if (!path)
4238                 return -ENOMEM;
4239         path->leave_spinning = 1;
4240
4241         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4242         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4243
4244         /*
4245          * lookup the last file extent.  We're not using i_size here
4246          * because there might be preallocation past i_size
4247          */
4248         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
4249                                        0);
4250         if (ret < 0) {
4251                 btrfs_free_path(path);
4252                 return ret;
4253         }
4254         WARN_ON(!ret);
4255         path->slots[0]--;
4256         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4257         found_type = found_key.type;
4258
4259         /* No extents, but there might be delalloc bits */
4260         if (found_key.objectid != btrfs_ino(inode) ||
4261             found_type != BTRFS_EXTENT_DATA_KEY) {
4262                 /* have to trust i_size as the end */
4263                 last = (u64)-1;
4264                 last_for_get_extent = isize;
4265         } else {
4266                 /*
4267                  * remember the start of the last extent.  There are a
4268                  * bunch of different factors that go into the length of the
4269                  * extent, so its much less complex to remember where it started
4270                  */
4271                 last = found_key.offset;
4272                 last_for_get_extent = last + 1;
4273         }
4274         btrfs_release_path(path);
4275
4276         /*
4277          * we might have some extents allocated but more delalloc past those
4278          * extents.  so, we trust isize unless the start of the last extent is
4279          * beyond isize
4280          */
4281         if (last < isize) {
4282                 last = (u64)-1;
4283                 last_for_get_extent = isize;
4284         }
4285
4286         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4287                          &cached_state);
4288
4289         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4290                                    get_extent);
4291         if (!em)
4292                 goto out;
4293         if (IS_ERR(em)) {
4294                 ret = PTR_ERR(em);
4295                 goto out;
4296         }
4297
4298         while (!end) {
4299                 u64 offset_in_extent = 0;
4300
4301                 /* break if the extent we found is outside the range */
4302                 if (em->start >= max || extent_map_end(em) < off)
4303                         break;
4304
4305                 /*
4306                  * get_extent may return an extent that starts before our
4307                  * requested range.  We have to make sure the ranges
4308                  * we return to fiemap always move forward and don't
4309                  * overlap, so adjust the offsets here
4310                  */
4311                 em_start = max(em->start, off);
4312
4313                 /*
4314                  * record the offset from the start of the extent
4315                  * for adjusting the disk offset below.  Only do this if the
4316                  * extent isn't compressed since our in ram offset may be past
4317                  * what we have actually allocated on disk.
4318                  */
4319                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4320                         offset_in_extent = em_start - em->start;
4321                 em_end = extent_map_end(em);
4322                 em_len = em_end - em_start;
4323                 disko = 0;
4324                 flags = 0;
4325
4326                 /*
4327                  * bump off for our next call to get_extent
4328                  */
4329                 off = extent_map_end(em);
4330                 if (off >= max)
4331                         end = 1;
4332
4333                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4334                         end = 1;
4335                         flags |= FIEMAP_EXTENT_LAST;
4336                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4337                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4338                                   FIEMAP_EXTENT_NOT_ALIGNED);
4339                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4340                         flags |= (FIEMAP_EXTENT_DELALLOC |
4341                                   FIEMAP_EXTENT_UNKNOWN);
4342                 } else if (fieinfo->fi_extents_max) {
4343                         u64 bytenr = em->block_start -
4344                                 (em->start - em->orig_start);
4345
4346                         disko = em->block_start + offset_in_extent;
4347
4348                         /*
4349                          * As btrfs supports shared space, this information
4350                          * can be exported to userspace tools via
4351                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4352                          * then we're just getting a count and we can skip the
4353                          * lookup stuff.
4354                          */
4355                         ret = btrfs_check_shared(NULL, root->fs_info,
4356                                                  root->objectid,
4357                                                  btrfs_ino(inode), bytenr);
4358                         if (ret < 0)
4359                                 goto out_free;
4360                         if (ret)
4361                                 flags |= FIEMAP_EXTENT_SHARED;
4362                         ret = 0;
4363                 }
4364                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4365                         flags |= FIEMAP_EXTENT_ENCODED;
4366
4367                 free_extent_map(em);
4368                 em = NULL;
4369                 if ((em_start >= last) || em_len == (u64)-1 ||
4370                    (last == (u64)-1 && isize <= em_end)) {
4371                         flags |= FIEMAP_EXTENT_LAST;
4372                         end = 1;
4373                 }
4374
4375                 /* now scan forward to see if this is really the last extent. */
4376                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4377                                            get_extent);
4378                 if (IS_ERR(em)) {
4379                         ret = PTR_ERR(em);
4380                         goto out;
4381                 }
4382                 if (!em) {
4383                         flags |= FIEMAP_EXTENT_LAST;
4384                         end = 1;
4385                 }
4386                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4387                                               em_len, flags);
4388                 if (ret)
4389                         goto out_free;
4390         }
4391 out_free:
4392         free_extent_map(em);
4393 out:
4394         btrfs_free_path(path);
4395         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4396                              &cached_state, GFP_NOFS);
4397         return ret;
4398 }
4399
4400 static void __free_extent_buffer(struct extent_buffer *eb)
4401 {
4402         btrfs_leak_debug_del(&eb->leak_list);
4403         kmem_cache_free(extent_buffer_cache, eb);
4404 }
4405
4406 int extent_buffer_under_io(struct extent_buffer *eb)
4407 {
4408         return (atomic_read(&eb->io_pages) ||
4409                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4410                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4411 }
4412
4413 /*
4414  * Helper for releasing extent buffer page.
4415  */
4416 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4417                                                 unsigned long start_idx)
4418 {
4419         unsigned long index;
4420         unsigned long num_pages;
4421         struct page *page;
4422         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4423
4424         BUG_ON(extent_buffer_under_io(eb));
4425
4426         num_pages = num_extent_pages(eb->start, eb->len);
4427         index = start_idx + num_pages;
4428         if (start_idx >= index)
4429                 return;
4430
4431         do {
4432                 index--;
4433                 page = extent_buffer_page(eb, index);
4434                 if (page && mapped) {
4435                         spin_lock(&page->mapping->private_lock);
4436                         /*
4437                          * We do this since we'll remove the pages after we've
4438                          * removed the eb from the radix tree, so we could race
4439                          * and have this page now attached to the new eb.  So
4440                          * only clear page_private if it's still connected to
4441                          * this eb.
4442                          */
4443                         if (PagePrivate(page) &&
4444                             page->private == (unsigned long)eb) {
4445                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4446                                 BUG_ON(PageDirty(page));
4447                                 BUG_ON(PageWriteback(page));
4448                                 /*
4449                                  * We need to make sure we haven't be attached
4450                                  * to a new eb.
4451                                  */
4452                                 ClearPagePrivate(page);
4453                                 set_page_private(page, 0);
4454                                 /* One for the page private */
4455                                 page_cache_release(page);
4456                         }
4457                         spin_unlock(&page->mapping->private_lock);
4458
4459                 }
4460                 if (page) {
4461                         /* One for when we alloced the page */
4462                         page_cache_release(page);
4463                 }
4464         } while (index != start_idx);
4465 }
4466
4467 /*
4468  * Helper for releasing the extent buffer.
4469  */
4470 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4471 {
4472         btrfs_release_extent_buffer_page(eb, 0);
4473         __free_extent_buffer(eb);
4474 }
4475
4476 static struct extent_buffer *
4477 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4478                       unsigned long len, gfp_t mask)
4479 {
4480         struct extent_buffer *eb = NULL;
4481
4482         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4483         if (eb == NULL)
4484                 return NULL;
4485         eb->start = start;
4486         eb->len = len;
4487         eb->fs_info = fs_info;
4488         eb->bflags = 0;
4489         rwlock_init(&eb->lock);
4490         atomic_set(&eb->write_locks, 0);
4491         atomic_set(&eb->read_locks, 0);
4492         atomic_set(&eb->blocking_readers, 0);
4493         atomic_set(&eb->blocking_writers, 0);
4494         atomic_set(&eb->spinning_readers, 0);
4495         atomic_set(&eb->spinning_writers, 0);
4496         eb->lock_nested = 0;
4497         init_waitqueue_head(&eb->write_lock_wq);
4498         init_waitqueue_head(&eb->read_lock_wq);
4499
4500         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4501
4502         spin_lock_init(&eb->refs_lock);
4503         atomic_set(&eb->refs, 1);
4504         atomic_set(&eb->io_pages, 0);
4505
4506         /*
4507          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4508          */
4509         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4510                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4511         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4512
4513         return eb;
4514 }
4515
4516 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4517 {
4518         unsigned long i;
4519         struct page *p;
4520         struct extent_buffer *new;
4521         unsigned long num_pages = num_extent_pages(src->start, src->len);
4522
4523         new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_NOFS);
4524         if (new == NULL)
4525                 return NULL;
4526
4527         for (i = 0; i < num_pages; i++) {
4528                 p = alloc_page(GFP_NOFS);
4529                 if (!p) {
4530                         btrfs_release_extent_buffer(new);
4531                         return NULL;
4532                 }
4533                 attach_extent_buffer_page(new, p);
4534                 WARN_ON(PageDirty(p));
4535                 SetPageUptodate(p);
4536                 new->pages[i] = p;
4537         }
4538
4539         copy_extent_buffer(new, src, 0, 0, src->len);
4540         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4541         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4542
4543         return new;
4544 }
4545
4546 struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4547 {
4548         struct extent_buffer *eb;
4549         unsigned long num_pages = num_extent_pages(0, len);
4550         unsigned long i;
4551
4552         eb = __alloc_extent_buffer(NULL, start, len, GFP_NOFS);
4553         if (!eb)
4554                 return NULL;
4555
4556         for (i = 0; i < num_pages; i++) {
4557                 eb->pages[i] = alloc_page(GFP_NOFS);
4558                 if (!eb->pages[i])
4559                         goto err;
4560         }
4561         set_extent_buffer_uptodate(eb);
4562         btrfs_set_header_nritems(eb, 0);
4563         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4564
4565         return eb;
4566 err:
4567         for (; i > 0; i--)
4568                 __free_page(eb->pages[i - 1]);
4569         __free_extent_buffer(eb);
4570         return NULL;
4571 }
4572
4573 static void check_buffer_tree_ref(struct extent_buffer *eb)
4574 {
4575         int refs;
4576         /* the ref bit is tricky.  We have to make sure it is set
4577          * if we have the buffer dirty.   Otherwise the
4578          * code to free a buffer can end up dropping a dirty
4579          * page
4580          *
4581          * Once the ref bit is set, it won't go away while the
4582          * buffer is dirty or in writeback, and it also won't
4583          * go away while we have the reference count on the
4584          * eb bumped.
4585          *
4586          * We can't just set the ref bit without bumping the
4587          * ref on the eb because free_extent_buffer might
4588          * see the ref bit and try to clear it.  If this happens
4589          * free_extent_buffer might end up dropping our original
4590          * ref by mistake and freeing the page before we are able
4591          * to add one more ref.
4592          *
4593          * So bump the ref count first, then set the bit.  If someone
4594          * beat us to it, drop the ref we added.
4595          */
4596         refs = atomic_read(&eb->refs);
4597         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4598                 return;
4599
4600         spin_lock(&eb->refs_lock);
4601         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4602                 atomic_inc(&eb->refs);
4603         spin_unlock(&eb->refs_lock);
4604 }
4605
4606 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4607                 struct page *accessed)
4608 {
4609         unsigned long num_pages, i;
4610
4611         check_buffer_tree_ref(eb);
4612
4613         num_pages = num_extent_pages(eb->start, eb->len);
4614         for (i = 0; i < num_pages; i++) {
4615                 struct page *p = extent_buffer_page(eb, i);
4616                 if (p != accessed)
4617                         mark_page_accessed(p);
4618         }
4619 }
4620
4621 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4622                                          u64 start)
4623 {
4624         struct extent_buffer *eb;
4625
4626         rcu_read_lock();
4627         eb = radix_tree_lookup(&fs_info->buffer_radix,
4628                                start >> PAGE_CACHE_SHIFT);
4629         if (eb && atomic_inc_not_zero(&eb->refs)) {
4630                 rcu_read_unlock();
4631                 mark_extent_buffer_accessed(eb, NULL);
4632                 return eb;
4633         }
4634         rcu_read_unlock();
4635
4636         return NULL;
4637 }
4638
4639 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4640 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4641                                                u64 start, unsigned long len)
4642 {
4643         struct extent_buffer *eb, *exists = NULL;
4644         int ret;
4645
4646         eb = find_extent_buffer(fs_info, start);
4647         if (eb)
4648                 return eb;
4649         eb = alloc_dummy_extent_buffer(start, len);
4650         if (!eb)
4651                 return NULL;
4652         eb->fs_info = fs_info;
4653 again:
4654         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4655         if (ret)
4656                 goto free_eb;
4657         spin_lock(&fs_info->buffer_lock);
4658         ret = radix_tree_insert(&fs_info->buffer_radix,
4659                                 start >> PAGE_CACHE_SHIFT, eb);
4660         spin_unlock(&fs_info->buffer_lock);
4661         radix_tree_preload_end();
4662         if (ret == -EEXIST) {
4663                 exists = find_extent_buffer(fs_info, start);
4664                 if (exists)
4665                         goto free_eb;
4666                 else
4667                         goto again;
4668         }
4669         check_buffer_tree_ref(eb);
4670         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4671
4672         /*
4673          * We will free dummy extent buffer's if they come into
4674          * free_extent_buffer with a ref count of 2, but if we are using this we
4675          * want the buffers to stay in memory until we're done with them, so
4676          * bump the ref count again.
4677          */
4678         atomic_inc(&eb->refs);
4679         return eb;
4680 free_eb:
4681         btrfs_release_extent_buffer(eb);
4682         return exists;
4683 }
4684 #endif
4685
4686 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4687                                           u64 start, unsigned long len)
4688 {
4689         unsigned long num_pages = num_extent_pages(start, len);
4690         unsigned long i;
4691         unsigned long index = start >> PAGE_CACHE_SHIFT;
4692         struct extent_buffer *eb;
4693         struct extent_buffer *exists = NULL;
4694         struct page *p;
4695         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4696         int uptodate = 1;
4697         int ret;
4698
4699         eb = find_extent_buffer(fs_info, start);
4700         if (eb)
4701                 return eb;
4702
4703         eb = __alloc_extent_buffer(fs_info, start, len, GFP_NOFS);
4704         if (!eb)
4705                 return NULL;
4706
4707         for (i = 0; i < num_pages; i++, index++) {
4708                 p = find_or_create_page(mapping, index, GFP_NOFS);
4709                 if (!p)
4710                         goto free_eb;
4711
4712                 spin_lock(&mapping->private_lock);
4713                 if (PagePrivate(p)) {
4714                         /*
4715                          * We could have already allocated an eb for this page
4716                          * and attached one so lets see if we can get a ref on
4717                          * the existing eb, and if we can we know it's good and
4718                          * we can just return that one, else we know we can just
4719                          * overwrite page->private.
4720                          */
4721                         exists = (struct extent_buffer *)p->private;
4722                         if (atomic_inc_not_zero(&exists->refs)) {
4723                                 spin_unlock(&mapping->private_lock);
4724                                 unlock_page(p);
4725                                 page_cache_release(p);
4726                                 mark_extent_buffer_accessed(exists, p);
4727                                 goto free_eb;
4728                         }
4729
4730                         /*
4731                          * Do this so attach doesn't complain and we need to
4732                          * drop the ref the old guy had.
4733                          */
4734                         ClearPagePrivate(p);
4735                         WARN_ON(PageDirty(p));
4736                         page_cache_release(p);
4737                 }
4738                 attach_extent_buffer_page(eb, p);
4739                 spin_unlock(&mapping->private_lock);
4740                 WARN_ON(PageDirty(p));
4741                 eb->pages[i] = p;
4742                 if (!PageUptodate(p))
4743                         uptodate = 0;
4744
4745                 /*
4746                  * see below about how we avoid a nasty race with release page
4747                  * and why we unlock later
4748                  */
4749         }
4750         if (uptodate)
4751                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4752 again:
4753         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4754         if (ret)
4755                 goto free_eb;
4756
4757         spin_lock(&fs_info->buffer_lock);
4758         ret = radix_tree_insert(&fs_info->buffer_radix,
4759                                 start >> PAGE_CACHE_SHIFT, eb);
4760         spin_unlock(&fs_info->buffer_lock);
4761         radix_tree_preload_end();
4762         if (ret == -EEXIST) {
4763                 exists = find_extent_buffer(fs_info, start);
4764                 if (exists)
4765                         goto free_eb;
4766                 else
4767                         goto again;
4768         }
4769         /* add one reference for the tree */
4770         check_buffer_tree_ref(eb);
4771         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4772
4773         /*
4774          * there is a race where release page may have
4775          * tried to find this extent buffer in the radix
4776          * but failed.  It will tell the VM it is safe to
4777          * reclaim the, and it will clear the page private bit.
4778          * We must make sure to set the page private bit properly
4779          * after the extent buffer is in the radix tree so
4780          * it doesn't get lost
4781          */
4782         SetPageChecked(eb->pages[0]);
4783         for (i = 1; i < num_pages; i++) {
4784                 p = extent_buffer_page(eb, i);
4785                 ClearPageChecked(p);
4786                 unlock_page(p);
4787         }
4788         unlock_page(eb->pages[0]);
4789         return eb;
4790
4791 free_eb:
4792         for (i = 0; i < num_pages; i++) {
4793                 if (eb->pages[i])
4794                         unlock_page(eb->pages[i]);
4795         }
4796
4797         WARN_ON(!atomic_dec_and_test(&eb->refs));
4798         btrfs_release_extent_buffer(eb);
4799         return exists;
4800 }
4801
4802 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4803 {
4804         struct extent_buffer *eb =
4805                         container_of(head, struct extent_buffer, rcu_head);
4806
4807         __free_extent_buffer(eb);
4808 }
4809
4810 /* Expects to have eb->eb_lock already held */
4811 static int release_extent_buffer(struct extent_buffer *eb)
4812 {
4813         WARN_ON(atomic_read(&eb->refs) == 0);
4814         if (atomic_dec_and_test(&eb->refs)) {
4815                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4816                         struct btrfs_fs_info *fs_info = eb->fs_info;
4817
4818                         spin_unlock(&eb->refs_lock);
4819
4820                         spin_lock(&fs_info->buffer_lock);
4821                         radix_tree_delete(&fs_info->buffer_radix,
4822                                           eb->start >> PAGE_CACHE_SHIFT);
4823                         spin_unlock(&fs_info->buffer_lock);
4824                 } else {
4825                         spin_unlock(&eb->refs_lock);
4826                 }
4827
4828                 /* Should be safe to release our pages at this point */
4829                 btrfs_release_extent_buffer_page(eb, 0);
4830                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4831                 return 1;
4832         }
4833         spin_unlock(&eb->refs_lock);
4834
4835         return 0;
4836 }
4837
4838 void free_extent_buffer(struct extent_buffer *eb)
4839 {
4840         int refs;
4841         int old;
4842         if (!eb)
4843                 return;
4844
4845         while (1) {
4846                 refs = atomic_read(&eb->refs);
4847                 if (refs <= 3)
4848                         break;
4849                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
4850                 if (old == refs)
4851                         return;
4852         }
4853
4854         spin_lock(&eb->refs_lock);
4855         if (atomic_read(&eb->refs) == 2 &&
4856             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4857                 atomic_dec(&eb->refs);
4858
4859         if (atomic_read(&eb->refs) == 2 &&
4860             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4861             !extent_buffer_under_io(eb) &&
4862             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4863                 atomic_dec(&eb->refs);
4864
4865         /*
4866          * I know this is terrible, but it's temporary until we stop tracking
4867          * the uptodate bits and such for the extent buffers.
4868          */
4869         release_extent_buffer(eb);
4870 }
4871
4872 void free_extent_buffer_stale(struct extent_buffer *eb)
4873 {
4874         if (!eb)
4875                 return;
4876
4877         spin_lock(&eb->refs_lock);
4878         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4879
4880         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4881             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4882                 atomic_dec(&eb->refs);
4883         release_extent_buffer(eb);
4884 }
4885
4886 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4887 {
4888         unsigned long i;
4889         unsigned long num_pages;
4890         struct page *page;
4891
4892         num_pages = num_extent_pages(eb->start, eb->len);
4893
4894         for (i = 0; i < num_pages; i++) {
4895                 page = extent_buffer_page(eb, i);
4896                 if (!PageDirty(page))
4897                         continue;
4898
4899                 lock_page(page);
4900                 WARN_ON(!PagePrivate(page));
4901
4902                 clear_page_dirty_for_io(page);
4903                 spin_lock_irq(&page->mapping->tree_lock);
4904                 if (!PageDirty(page)) {
4905                         radix_tree_tag_clear(&page->mapping->page_tree,
4906                                                 page_index(page),
4907                                                 PAGECACHE_TAG_DIRTY);
4908                 }
4909                 spin_unlock_irq(&page->mapping->tree_lock);
4910                 ClearPageError(page);
4911                 unlock_page(page);
4912         }
4913         WARN_ON(atomic_read(&eb->refs) == 0);
4914 }
4915
4916 int set_extent_buffer_dirty(struct extent_buffer *eb)
4917 {
4918         unsigned long i;
4919         unsigned long num_pages;
4920         int was_dirty = 0;
4921
4922         check_buffer_tree_ref(eb);
4923
4924         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4925
4926         num_pages = num_extent_pages(eb->start, eb->len);
4927         WARN_ON(atomic_read(&eb->refs) == 0);
4928         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4929
4930         for (i = 0; i < num_pages; i++)
4931                 set_page_dirty(extent_buffer_page(eb, i));
4932         return was_dirty;
4933 }
4934
4935 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4936 {
4937         unsigned long i;
4938         struct page *page;
4939         unsigned long num_pages;
4940
4941         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4942         num_pages = num_extent_pages(eb->start, eb->len);
4943         for (i = 0; i < num_pages; i++) {
4944                 page = extent_buffer_page(eb, i);
4945                 if (page)
4946                         ClearPageUptodate(page);
4947         }
4948         return 0;
4949 }
4950
4951 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4952 {
4953         unsigned long i;
4954         struct page *page;
4955         unsigned long num_pages;
4956
4957         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4958         num_pages = num_extent_pages(eb->start, eb->len);
4959         for (i = 0; i < num_pages; i++) {
4960                 page = extent_buffer_page(eb, i);
4961                 SetPageUptodate(page);
4962         }
4963         return 0;
4964 }
4965
4966 int extent_buffer_uptodate(struct extent_buffer *eb)
4967 {
4968         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4969 }
4970
4971 int read_extent_buffer_pages(struct extent_io_tree *tree,
4972                              struct extent_buffer *eb, u64 start, int wait,
4973                              get_extent_t *get_extent, int mirror_num)
4974 {
4975         unsigned long i;
4976         unsigned long start_i;
4977         struct page *page;
4978         int err;
4979         int ret = 0;
4980         int locked_pages = 0;
4981         int all_uptodate = 1;
4982         unsigned long num_pages;
4983         unsigned long num_reads = 0;
4984         struct bio *bio = NULL;
4985         unsigned long bio_flags = 0;
4986
4987         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4988                 return 0;
4989
4990         if (start) {
4991                 WARN_ON(start < eb->start);
4992                 start_i = (start >> PAGE_CACHE_SHIFT) -
4993                         (eb->start >> PAGE_CACHE_SHIFT);
4994         } else {
4995                 start_i = 0;
4996         }
4997
4998         num_pages = num_extent_pages(eb->start, eb->len);
4999         for (i = start_i; i < num_pages; i++) {
5000                 page = extent_buffer_page(eb, i);
5001                 if (wait == WAIT_NONE) {
5002                         if (!trylock_page(page))
5003                                 goto unlock_exit;
5004                 } else {
5005                         lock_page(page);
5006                 }
5007                 locked_pages++;
5008                 if (!PageUptodate(page)) {
5009                         num_reads++;
5010                         all_uptodate = 0;
5011                 }
5012         }
5013         if (all_uptodate) {
5014                 if (start_i == 0)
5015                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5016                 goto unlock_exit;
5017         }
5018
5019         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
5020         eb->read_mirror = 0;
5021         atomic_set(&eb->io_pages, num_reads);
5022         for (i = start_i; i < num_pages; i++) {
5023                 page = extent_buffer_page(eb, i);
5024                 if (!PageUptodate(page)) {
5025                         ClearPageError(page);
5026                         err = __extent_read_full_page(tree, page,
5027                                                       get_extent, &bio,
5028                                                       mirror_num, &bio_flags,
5029                                                       READ | REQ_META);
5030                         if (err)
5031                                 ret = err;
5032                 } else {
5033                         unlock_page(page);
5034                 }
5035         }
5036
5037         if (bio) {
5038                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
5039                                      bio_flags);
5040                 if (err)
5041                         return err;
5042         }
5043
5044         if (ret || wait != WAIT_COMPLETE)
5045                 return ret;
5046
5047         for (i = start_i; i < num_pages; i++) {
5048                 page = extent_buffer_page(eb, i);
5049                 wait_on_page_locked(page);
5050                 if (!PageUptodate(page))
5051                         ret = -EIO;
5052         }
5053
5054         return ret;
5055
5056 unlock_exit:
5057         i = start_i;
5058         while (locked_pages > 0) {
5059                 page = extent_buffer_page(eb, i);
5060                 i++;
5061                 unlock_page(page);
5062                 locked_pages--;
5063         }
5064         return ret;
5065 }
5066
5067 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5068                         unsigned long start,
5069                         unsigned long len)
5070 {
5071         size_t cur;
5072         size_t offset;
5073         struct page *page;
5074         char *kaddr;
5075         char *dst = (char *)dstv;
5076         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5077         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5078
5079         WARN_ON(start > eb->len);
5080         WARN_ON(start + len > eb->start + eb->len);
5081
5082         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5083
5084         while (len > 0) {
5085                 page = extent_buffer_page(eb, i);
5086
5087                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5088                 kaddr = page_address(page);
5089                 memcpy(dst, kaddr + offset, cur);
5090
5091                 dst += cur;
5092                 len -= cur;
5093                 offset = 0;
5094                 i++;
5095         }
5096 }
5097
5098 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5099                         unsigned long start,
5100                         unsigned long len)
5101 {
5102         size_t cur;
5103         size_t offset;
5104         struct page *page;
5105         char *kaddr;
5106         char __user *dst = (char __user *)dstv;
5107         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5108         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5109         int ret = 0;
5110
5111         WARN_ON(start > eb->len);
5112         WARN_ON(start + len > eb->start + eb->len);
5113
5114         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5115
5116         while (len > 0) {
5117                 page = extent_buffer_page(eb, i);
5118
5119                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5120                 kaddr = page_address(page);
5121                 if (copy_to_user(dst, kaddr + offset, cur)) {
5122                         ret = -EFAULT;
5123                         break;
5124                 }
5125
5126                 dst += cur;
5127                 len -= cur;
5128                 offset = 0;
5129                 i++;
5130         }
5131
5132         return ret;
5133 }
5134
5135 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5136                                unsigned long min_len, char **map,
5137                                unsigned long *map_start,
5138                                unsigned long *map_len)
5139 {
5140         size_t offset = start & (PAGE_CACHE_SIZE - 1);
5141         char *kaddr;
5142         struct page *p;
5143         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5144         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5145         unsigned long end_i = (start_offset + start + min_len - 1) >>
5146                 PAGE_CACHE_SHIFT;
5147
5148         if (i != end_i)
5149                 return -EINVAL;
5150
5151         if (i == 0) {
5152                 offset = start_offset;
5153                 *map_start = 0;
5154         } else {
5155                 offset = 0;
5156                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
5157         }
5158
5159         if (start + min_len > eb->len) {
5160                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5161                        "wanted %lu %lu\n",
5162                        eb->start, eb->len, start, min_len);
5163                 return -EINVAL;
5164         }
5165
5166         p = extent_buffer_page(eb, i);
5167         kaddr = page_address(p);
5168         *map = kaddr + offset;
5169         *map_len = PAGE_CACHE_SIZE - offset;
5170         return 0;
5171 }
5172
5173 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5174                           unsigned long start,
5175                           unsigned long len)
5176 {
5177         size_t cur;
5178         size_t offset;
5179         struct page *page;
5180         char *kaddr;
5181         char *ptr = (char *)ptrv;
5182         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5183         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5184         int ret = 0;
5185
5186         WARN_ON(start > eb->len);
5187         WARN_ON(start + len > eb->start + eb->len);
5188
5189         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5190
5191         while (len > 0) {
5192                 page = extent_buffer_page(eb, i);
5193
5194                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5195
5196                 kaddr = page_address(page);
5197                 ret = memcmp(ptr, kaddr + offset, cur);
5198                 if (ret)
5199                         break;
5200
5201                 ptr += cur;
5202                 len -= cur;
5203                 offset = 0;
5204                 i++;
5205         }
5206         return ret;
5207 }
5208
5209 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5210                          unsigned long start, unsigned long len)
5211 {
5212         size_t cur;
5213         size_t offset;
5214         struct page *page;
5215         char *kaddr;
5216         char *src = (char *)srcv;
5217         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5218         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5219
5220         WARN_ON(start > eb->len);
5221         WARN_ON(start + len > eb->start + eb->len);
5222
5223         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5224
5225         while (len > 0) {
5226                 page = extent_buffer_page(eb, i);
5227                 WARN_ON(!PageUptodate(page));
5228
5229                 cur = min(len, PAGE_CACHE_SIZE - offset);
5230                 kaddr = page_address(page);
5231                 memcpy(kaddr + offset, src, cur);
5232
5233                 src += cur;
5234                 len -= cur;
5235                 offset = 0;
5236                 i++;
5237         }
5238 }
5239
5240 void memset_extent_buffer(struct extent_buffer *eb, char c,
5241                           unsigned long start, unsigned long len)
5242 {
5243         size_t cur;
5244         size_t offset;
5245         struct page *page;
5246         char *kaddr;
5247         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5248         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5249
5250         WARN_ON(start > eb->len);
5251         WARN_ON(start + len > eb->start + eb->len);
5252
5253         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5254
5255         while (len > 0) {
5256                 page = extent_buffer_page(eb, i);
5257                 WARN_ON(!PageUptodate(page));
5258
5259                 cur = min(len, PAGE_CACHE_SIZE - offset);
5260                 kaddr = page_address(page);
5261                 memset(kaddr + offset, c, cur);
5262
5263                 len -= cur;
5264                 offset = 0;
5265                 i++;
5266         }
5267 }
5268
5269 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5270                         unsigned long dst_offset, unsigned long src_offset,
5271                         unsigned long len)
5272 {
5273         u64 dst_len = dst->len;
5274         size_t cur;
5275         size_t offset;
5276         struct page *page;
5277         char *kaddr;
5278         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5279         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5280
5281         WARN_ON(src->len != dst_len);
5282
5283         offset = (start_offset + dst_offset) &
5284                 (PAGE_CACHE_SIZE - 1);
5285
5286         while (len > 0) {
5287                 page = extent_buffer_page(dst, i);
5288                 WARN_ON(!PageUptodate(page));
5289
5290                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5291
5292                 kaddr = page_address(page);
5293                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5294
5295                 src_offset += cur;
5296                 len -= cur;
5297                 offset = 0;
5298                 i++;
5299         }
5300 }
5301
5302 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5303 {
5304         unsigned long distance = (src > dst) ? src - dst : dst - src;
5305         return distance < len;
5306 }
5307
5308 static void copy_pages(struct page *dst_page, struct page *src_page,
5309                        unsigned long dst_off, unsigned long src_off,
5310                        unsigned long len)
5311 {
5312         char *dst_kaddr = page_address(dst_page);
5313         char *src_kaddr;
5314         int must_memmove = 0;
5315
5316         if (dst_page != src_page) {
5317                 src_kaddr = page_address(src_page);
5318         } else {
5319                 src_kaddr = dst_kaddr;
5320                 if (areas_overlap(src_off, dst_off, len))
5321                         must_memmove = 1;
5322         }
5323
5324         if (must_memmove)
5325                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5326         else
5327                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5328 }
5329
5330 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5331                            unsigned long src_offset, unsigned long len)
5332 {
5333         size_t cur;
5334         size_t dst_off_in_page;
5335         size_t src_off_in_page;
5336         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5337         unsigned long dst_i;
5338         unsigned long src_i;
5339
5340         if (src_offset + len > dst->len) {
5341                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5342                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5343                 BUG_ON(1);
5344         }
5345         if (dst_offset + len > dst->len) {
5346                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5347                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5348                 BUG_ON(1);
5349         }
5350
5351         while (len > 0) {
5352                 dst_off_in_page = (start_offset + dst_offset) &
5353                         (PAGE_CACHE_SIZE - 1);
5354                 src_off_in_page = (start_offset + src_offset) &
5355                         (PAGE_CACHE_SIZE - 1);
5356
5357                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5358                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5359
5360                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5361                                                src_off_in_page));
5362                 cur = min_t(unsigned long, cur,
5363                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5364
5365                 copy_pages(extent_buffer_page(dst, dst_i),
5366                            extent_buffer_page(dst, src_i),
5367                            dst_off_in_page, src_off_in_page, cur);
5368
5369                 src_offset += cur;
5370                 dst_offset += cur;
5371                 len -= cur;
5372         }
5373 }
5374
5375 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5376                            unsigned long src_offset, unsigned long len)
5377 {
5378         size_t cur;
5379         size_t dst_off_in_page;
5380         size_t src_off_in_page;
5381         unsigned long dst_end = dst_offset + len - 1;
5382         unsigned long src_end = src_offset + len - 1;
5383         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5384         unsigned long dst_i;
5385         unsigned long src_i;
5386
5387         if (src_offset + len > dst->len) {
5388                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5389                        "len %lu len %lu\n", src_offset, len, dst->len);
5390                 BUG_ON(1);
5391         }
5392         if (dst_offset + len > dst->len) {
5393                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5394                        "len %lu len %lu\n", dst_offset, len, dst->len);
5395                 BUG_ON(1);
5396         }
5397         if (dst_offset < src_offset) {
5398                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5399                 return;
5400         }
5401         while (len > 0) {
5402                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5403                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5404
5405                 dst_off_in_page = (start_offset + dst_end) &
5406                         (PAGE_CACHE_SIZE - 1);
5407                 src_off_in_page = (start_offset + src_end) &
5408                         (PAGE_CACHE_SIZE - 1);
5409
5410                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5411                 cur = min(cur, dst_off_in_page + 1);
5412                 copy_pages(extent_buffer_page(dst, dst_i),
5413                            extent_buffer_page(dst, src_i),
5414                            dst_off_in_page - cur + 1,
5415                            src_off_in_page - cur + 1, cur);
5416
5417                 dst_end -= cur;
5418                 src_end -= cur;
5419                 len -= cur;
5420         }
5421 }
5422
5423 int try_release_extent_buffer(struct page *page)
5424 {
5425         struct extent_buffer *eb;
5426
5427         /*
5428          * We need to make sure noboody is attaching this page to an eb right
5429          * now.
5430          */
5431         spin_lock(&page->mapping->private_lock);
5432         if (!PagePrivate(page)) {
5433                 spin_unlock(&page->mapping->private_lock);
5434                 return 1;
5435         }
5436
5437         eb = (struct extent_buffer *)page->private;
5438         BUG_ON(!eb);
5439
5440         /*
5441          * This is a little awful but should be ok, we need to make sure that
5442          * the eb doesn't disappear out from under us while we're looking at
5443          * this page.
5444          */
5445         spin_lock(&eb->refs_lock);
5446         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5447                 spin_unlock(&eb->refs_lock);
5448                 spin_unlock(&page->mapping->private_lock);
5449                 return 0;
5450         }
5451         spin_unlock(&page->mapping->private_lock);
5452
5453         /*
5454          * If tree ref isn't set then we know the ref on this eb is a real ref,
5455          * so just return, this page will likely be freed soon anyway.
5456          */
5457         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5458                 spin_unlock(&eb->refs_lock);
5459                 return 0;
5460         }
5461
5462         return release_extent_buffer(eb);
5463 }