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/module.h>
   8 #include <linux/spinlock.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/swap.h>
  11 #include <linux/writeback.h>
  12 #include <linux/pagevec.h>
  13 #include <linux/prefetch.h>
  14 #include <linux/cleancache.h>
  15 #include "extent_io.h"
  16 #include "extent_map.h"
  17 #include "compat.h"
  18 #include "ctree.h"
  19 #include "btrfs_inode.h"
  20 #include "volumes.h"
  21 #include "check-integrity.h"
  22 #include "locking.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26
  27 static LIST_HEAD(buffers);
  28 static LIST_HEAD(states);
  29
  30 #define LEAK_DEBUG 0
  31 #if LEAK_DEBUG
  32 static DEFINE_SPINLOCK(leak_lock);
  33 #endif
  34
  35 #define BUFFER_LRU_MAX 64
  36
  37 struct tree_entry {
  38         u64 start;
  39         u64 end;
  40         struct rb_node rb_node;
  41 };
  42
  43 struct extent_page_data {
  44         struct bio *bio;
  45         struct extent_io_tree *tree;
  46         get_extent_t *get_extent;
  47
  48         /* tells writepage not to lock the state bits for this range
  49          * it still does the unlocking
  50          */
  51         unsigned int extent_locked:1;
  52
  53         /* tells the submit_bio code to use a WRITE_SYNC */
  54         unsigned int sync_io:1;
  55 };
  56
  57 static noinline void flush_write_bio(void *data);
  58
  59 int __init extent_io_init(void)
  60 {
  61         extent_state_cache = kmem_cache_create("extent_state",
  62                         sizeof(struct extent_state), 0,
  63                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  64         if (!extent_state_cache)
  65                 return -ENOMEM;
  66
  67         extent_buffer_cache = kmem_cache_create("extent_buffers",
  68                         sizeof(struct extent_buffer), 0,
  69                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  70         if (!extent_buffer_cache)
  71                 goto free_state_cache;
  72         return 0;
  73
  74 free_state_cache:
  75         kmem_cache_destroy(extent_state_cache);
  76         return -ENOMEM;
  77 }
  78
  79 void extent_io_exit(void)
  80 {
  81         struct extent_state *state;
  82         struct extent_buffer *eb;
  83
  84         while (!list_empty(&states)) {
  85                 state = list_entry(states.next, struct extent_state, leak_list);
  86                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  87                        "state %lu in tree %p refs %d\n",
  88                        (unsigned long long)state->start,
  89                        (unsigned long long)state->end,
  90                        state->state, state->tree, atomic_read(&state->refs));
  91                 list_del(&state->leak_list);
  92                 kmem_cache_free(extent_state_cache, state);
  93
  94         }
  95
  96         while (!list_empty(&buffers)) {
  97                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  98                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  99                        "refs %d\n", (unsigned long long)eb->start,
 100                        eb->len, atomic_read(&eb->refs));
 101                 list_del(&eb->leak_list);
 102                 kmem_cache_free(extent_buffer_cache, eb);
 103         }
 104         if (extent_state_cache)
 105                 kmem_cache_destroy(extent_state_cache);
 106         if (extent_buffer_cache)
 107                 kmem_cache_destroy(extent_buffer_cache);
 108 }
 109
 110 void extent_io_tree_init(struct extent_io_tree *tree,
 111                          struct address_space *mapping)
 112 {
 113         tree->state = RB_ROOT;
 114         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 115         tree->ops = NULL;
 116         tree->dirty_bytes = 0;
 117         spin_lock_init(&tree->lock);
 118         spin_lock_init(&tree->buffer_lock);
 119         tree->mapping = mapping;
 120 }
 121
 122 static struct extent_state *alloc_extent_state(gfp_t mask)
 123 {
 124         struct extent_state *state;
 125 #if LEAK_DEBUG
 126         unsigned long flags;
 127 #endif
 128
 129         state = kmem_cache_alloc(extent_state_cache, mask);
 130         if (!state)
 131                 return state;
 132         state->state = 0;
 133         state->private = 0;
 134         state->tree = NULL;
 135 #if LEAK_DEBUG
 136         spin_lock_irqsave(&leak_lock, flags);
 137         list_add(&state->leak_list, &states);
 138         spin_unlock_irqrestore(&leak_lock, flags);
 139 #endif
 140         atomic_set(&state->refs, 1);
 141         init_waitqueue_head(&state->wq);
 142         return state;
 143 }
 144
 145 void free_extent_state(struct extent_state *state)
 146 {
 147         if (!state)
 148                 return;
 149         if (atomic_dec_and_test(&state->refs)) {
 150 #if LEAK_DEBUG
 151                 unsigned long flags;
 152 #endif
 153                 WARN_ON(state->tree);
 154 #if LEAK_DEBUG
 155                 spin_lock_irqsave(&leak_lock, flags);
 156                 list_del(&state->leak_list);
 157                 spin_unlock_irqrestore(&leak_lock, flags);
 158 #endif
 159                 kmem_cache_free(extent_state_cache, state);
 160         }
 161 }
 162
 163 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 164                                    struct rb_node *node)
 165 {
 166         struct rb_node **p = &root->rb_node;
 167         struct rb_node *parent = NULL;
 168         struct tree_entry *entry;
 169
 170         while (*p) {
 171                 parent = *p;
 172                 entry = rb_entry(parent, struct tree_entry, rb_node);
 173
 174                 if (offset < entry->start)
 175                         p = &(*p)->rb_left;
 176                 else if (offset > entry->end)
 177                         p = &(*p)->rb_right;
 178                 else
 179                         return parent;
 180         }
 181
 182         entry = rb_entry(node, struct tree_entry, rb_node);
 183         rb_link_node(node, parent, p);
 184         rb_insert_color(node, root);
 185         return NULL;
 186 }
 187
 188 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 189                                      struct rb_node **prev_ret,
 190                                      struct rb_node **next_ret)
 191 {
 192         struct rb_root *root = &tree->state;
 193         struct rb_node *n = root->rb_node;
 194         struct rb_node *prev = NULL;
 195         struct rb_node *orig_prev = NULL;
 196         struct tree_entry *entry;
 197         struct tree_entry *prev_entry = NULL;
 198
 199         while (n) {
 200                 entry = rb_entry(n, struct tree_entry, rb_node);
 201                 prev = n;
 202                 prev_entry = entry;
 203
 204                 if (offset < entry->start)
 205                         n = n->rb_left;
 206                 else if (offset > entry->end)
 207                         n = n->rb_right;
 208                 else
 209                         return n;
 210         }
 211
 212         if (prev_ret) {
 213                 orig_prev = prev;
 214                 while (prev && offset > prev_entry->end) {
 215                         prev = rb_next(prev);
 216                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 217                 }
 218                 *prev_ret = prev;
 219                 prev = orig_prev;
 220         }
 221
 222         if (next_ret) {
 223                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 224                 while (prev && offset < prev_entry->start) {
 225                         prev = rb_prev(prev);
 226                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 227                 }
 228                 *next_ret = prev;
 229         }
 230         return NULL;
 231 }
 232
 233 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 234                                           u64 offset)
 235 {
 236         struct rb_node *prev = NULL;
 237         struct rb_node *ret;
 238
 239         ret = __etree_search(tree, offset, &prev, NULL);
 240         if (!ret)
 241                 return prev;
 242         return ret;
 243 }
 244
 245 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 246                      struct extent_state *other)
 247 {
 248         if (tree->ops && tree->ops->merge_extent_hook)
 249                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 250                                              other);
 251 }
 252
 253 /*
 254  * utility function to look for merge candidates inside a given range.
 255  * Any extents with matching state are merged together into a single
 256  * extent in the tree.  Extents with EXTENT_IO in their state field
 257  * are not merged because the end_io handlers need to be able to do
 258  * operations on them without sleeping (or doing allocations/splits).
 259  *
 260  * This should be called with the tree lock held.
 261  */
 262 static void merge_state(struct extent_io_tree *tree,
 263                         struct extent_state *state)
 264 {
 265         struct extent_state *other;
 266         struct rb_node *other_node;
 267
 268         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 269                 return;
 270
 271         other_node = rb_prev(&state->rb_node);
 272         if (other_node) {
 273                 other = rb_entry(other_node, struct extent_state, rb_node);
 274                 if (other->end == state->start - 1 &&
 275                     other->state == state->state) {
 276                         merge_cb(tree, state, other);
 277                         state->start = other->start;
 278                         other->tree = NULL;
 279                         rb_erase(&other->rb_node, &tree->state);
 280                         free_extent_state(other);
 281                 }
 282         }
 283         other_node = rb_next(&state->rb_node);
 284         if (other_node) {
 285                 other = rb_entry(other_node, struct extent_state, rb_node);
 286                 if (other->start == state->end + 1 &&
 287                     other->state == state->state) {
 288                         merge_cb(tree, state, other);
 289                         state->end = other->end;
 290                         other->tree = NULL;
 291                         rb_erase(&other->rb_node, &tree->state);
 292                         free_extent_state(other);
 293                 }
 294         }
 295 }
 296
 297 static void set_state_cb(struct extent_io_tree *tree,
 298                          struct extent_state *state, int *bits)
 299 {
 300         if (tree->ops && tree->ops->set_bit_hook)
 301                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 302 }
 303
 304 static void clear_state_cb(struct extent_io_tree *tree,
 305                            struct extent_state *state, int *bits)
 306 {
 307         if (tree->ops && tree->ops->clear_bit_hook)
 308                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 309 }
 310
 311 static void set_state_bits(struct extent_io_tree *tree,
 312                            struct extent_state *state, int *bits);
 313
 314 /*
 315  * insert an extent_state struct into the tree.  'bits' are set on the
 316  * struct before it is inserted.
 317  *
 318  * This may return -EEXIST if the extent is already there, in which case the
 319  * state struct is freed.
 320  *
 321  * The tree lock is not taken internally.  This is a utility function and
 322  * probably isn't what you want to call (see set/clear_extent_bit).
 323  */
 324 static int insert_state(struct extent_io_tree *tree,
 325                         struct extent_state *state, u64 start, u64 end,
 326                         int *bits)
 327 {
 328         struct rb_node *node;
 329
 330         if (end < start) {
 331                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
 332                        (unsigned long long)end,
 333                        (unsigned long long)start);
 334                 WARN_ON(1);
 335         }
 336         state->start = start;
 337         state->end = end;
 338
 339         set_state_bits(tree, state, bits);
 340
 341         node = tree_insert(&tree->state, end, &state->rb_node);
 342         if (node) {
 343                 struct extent_state *found;
 344                 found = rb_entry(node, struct extent_state, rb_node);
 345                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 346                        "%llu %llu\n", (unsigned long long)found->start,
 347                        (unsigned long long)found->end,
 348                        (unsigned long long)start, (unsigned long long)end);
 349                 return -EEXIST;
 350         }
 351         state->tree = tree;
 352         merge_state(tree, state);
 353         return 0;
 354 }
 355
 356 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 357                      u64 split)
 358 {
 359         if (tree->ops && tree->ops->split_extent_hook)
 360                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 361 }
 362
 363 /*
 364  * split a given extent state struct in two, inserting the preallocated
 365  * struct 'prealloc' as the newly created second half.  'split' indicates an
 366  * offset inside 'orig' where it should be split.
 367  *
 368  * Before calling,
 369  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 370  * are two extent state structs in the tree:
 371  * prealloc: [orig->start, split - 1]
 372  * orig: [ split, orig->end ]
 373  *
 374  * The tree locks are not taken by this function. They need to be held
 375  * by the caller.
 376  */
 377 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 378                        struct extent_state *prealloc, u64 split)
 379 {
 380         struct rb_node *node;
 381
 382         split_cb(tree, orig, split);
 383
 384         prealloc->start = orig->start;
 385         prealloc->end = split - 1;
 386         prealloc->state = orig->state;
 387         orig->start = split;
 388
 389         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 390         if (node) {
 391                 free_extent_state(prealloc);
 392                 return -EEXIST;
 393         }
 394         prealloc->tree = tree;
 395         return 0;
 396 }
 397
 398 /*
 399  * utility function to clear some bits in an extent state struct.
 400  * it will optionally wake up any one waiting on this state (wake == 1), or
 401  * forcibly remove the state from the tree (delete == 1).
 402  *
 403  * If no bits are set on the state struct after clearing things, the
 404  * struct is freed and removed from the tree
 405  */
 406 static int clear_state_bit(struct extent_io_tree *tree,
 407                             struct extent_state *state,
 408                             int *bits, int wake)
 409 {
 410         int bits_to_clear = *bits & ~EXTENT_CTLBITS;
 411         int ret = state->state & bits_to_clear;
 412
 413         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 414                 u64 range = state->end - state->start + 1;
 415                 WARN_ON(range > tree->dirty_bytes);
 416                 tree->dirty_bytes -= range;
 417         }
 418         clear_state_cb(tree, state, bits);
 419         state->state &= ~bits_to_clear;
 420         if (wake)
 421                 wake_up(&state->wq);
 422         if (state->state == 0) {
 423                 if (state->tree) {
 424                         rb_erase(&state->rb_node, &tree->state);
 425                         state->tree = NULL;
 426                         free_extent_state(state);
 427                 } else {
 428                         WARN_ON(1);
 429                 }
 430         } else {
 431                 merge_state(tree, state);
 432         }
 433         return ret;
 434 }
 435
 436 static struct extent_state *
 437 alloc_extent_state_atomic(struct extent_state *prealloc)
 438 {
 439         if (!prealloc)
 440                 prealloc = alloc_extent_state(GFP_ATOMIC);
 441
 442         return prealloc;
 443 }
 444
 445 /*
 446  * clear some bits on a range in the tree.  This may require splitting
 447  * or inserting elements in the tree, so the gfp mask is used to
 448  * indicate which allocations or sleeping are allowed.
 449  *
 450  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 451  * the given range from the tree regardless of state (ie for truncate).
 452  *
 453  * the range [start, end] is inclusive.
 454  *
 455  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 456  * bits were already set, or zero if none of the bits were already set.
 457  */
 458 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 459                      int bits, int wake, int delete,
 460                      struct extent_state **cached_state,
 461                      gfp_t mask)
 462 {
 463         struct extent_state *state;
 464         struct extent_state *cached;
 465         struct extent_state *prealloc = NULL;
 466         struct rb_node *next_node;
 467         struct rb_node *node;
 468         u64 last_end;
 469         int err;
 470         int set = 0;
 471         int clear = 0;
 472
 473         if (delete)
 474                 bits |= ~EXTENT_CTLBITS;
 475         bits |= EXTENT_FIRST_DELALLOC;
 476
 477         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 478                 clear = 1;
 479 again:
 480         if (!prealloc && (mask & __GFP_WAIT)) {
 481                 prealloc = alloc_extent_state(mask);
 482                 if (!prealloc)
 483                         return -ENOMEM;
 484         }
 485
 486         spin_lock(&tree->lock);
 487         if (cached_state) {
 488                 cached = *cached_state;
 489
 490                 if (clear) {
 491                         *cached_state = NULL;
 492                         cached_state = NULL;
 493                 }
 494
 495                 if (cached && cached->tree && cached->start <= start &&
 496                     cached->end > start) {
 497                         if (clear)
 498                                 atomic_dec(&cached->refs);
 499                         state = cached;
 500                         goto hit_next;
 501                 }
 502                 if (clear)
 503                         free_extent_state(cached);
 504         }
 505         /*
 506          * this search will find the extents that end after
 507          * our range starts
 508          */
 509         node = tree_search(tree, start);
 510         if (!node)
 511                 goto out;
 512         state = rb_entry(node, struct extent_state, rb_node);
 513 hit_next:
 514         if (state->start > end)
 515                 goto out;
 516         WARN_ON(state->end < start);
 517         last_end = state->end;
 518
 519         if (state->end < end && !need_resched())
 520                 next_node = rb_next(&state->rb_node);
 521         else
 522                 next_node = NULL;
 523
 524         /* the state doesn't have the wanted bits, go ahead */
 525         if (!(state->state & bits))
 526                 goto next;
 527
 528         /*
 529          *     | ---- desired range ---- |
 530          *  | state | or
 531          *  | ------------- state -------------- |
 532          *
 533          * We need to split the extent we found, and may flip
 534          * bits on second half.
 535          *
 536          * If the extent we found extends past our range, we
 537          * just split and search again.  It'll get split again
 538          * the next time though.
 539          *
 540          * If the extent we found is inside our range, we clear
 541          * the desired bit on it.
 542          */
 543
 544         if (state->start < start) {
 545                 prealloc = alloc_extent_state_atomic(prealloc);
 546                 BUG_ON(!prealloc);
 547                 err = split_state(tree, state, prealloc, start);
 548                 BUG_ON(err == -EEXIST);
 549                 prealloc = NULL;
 550                 if (err)
 551                         goto out;
 552                 if (state->end <= end) {
 553                         set |= clear_state_bit(tree, state, &bits, wake);
 554                         if (last_end == (u64)-1)
 555                                 goto out;
 556                         start = last_end + 1;
 557                 }
 558                 goto search_again;
 559         }
 560         /*
 561          * | ---- desired range ---- |
 562          *                        | state |
 563          * We need to split the extent, and clear the bit
 564          * on the first half
 565          */
 566         if (state->start <= end && state->end > end) {
 567                 prealloc = alloc_extent_state_atomic(prealloc);
 568                 BUG_ON(!prealloc);
 569                 err = split_state(tree, state, prealloc, end + 1);
 570                 BUG_ON(err == -EEXIST);
 571                 if (wake)
 572                         wake_up(&state->wq);
 573
 574                 set |= clear_state_bit(tree, prealloc, &bits, wake);
 575
 576                 prealloc = NULL;
 577                 goto out;
 578         }
 579
 580         set |= clear_state_bit(tree, state, &bits, wake);
 581 next:
 582         if (last_end == (u64)-1)
 583                 goto out;
 584         start = last_end + 1;
 585         if (start <= end && next_node) {
 586                 state = rb_entry(next_node, struct extent_state,
 587                                  rb_node);
 588                 goto hit_next;
 589         }
 590         goto search_again;
 591
 592 out:
 593         spin_unlock(&tree->lock);
 594         if (prealloc)
 595                 free_extent_state(prealloc);
 596
 597         return set;
 598
 599 search_again:
 600         if (start > end)
 601                 goto out;
 602         spin_unlock(&tree->lock);
 603         if (mask & __GFP_WAIT)
 604                 cond_resched();
 605         goto again;
 606 }
 607
 608 static int wait_on_state(struct extent_io_tree *tree,
 609                          struct extent_state *state)
 610                 __releases(tree->lock)
 611                 __acquires(tree->lock)
 612 {
 613         DEFINE_WAIT(wait);
 614         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 615         spin_unlock(&tree->lock);
 616         schedule();
 617         spin_lock(&tree->lock);
 618         finish_wait(&state->wq, &wait);
 619         return 0;
 620 }
 621
 622 /*
 623  * waits for one or more bits to clear on a range in the state tree.
 624  * The range [start, end] is inclusive.
 625  * The tree lock is taken by this function
 626  */
 627 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 628 {
 629         struct extent_state *state;
 630         struct rb_node *node;
 631
 632         spin_lock(&tree->lock);
 633 again:
 634         while (1) {
 635                 /*
 636                  * this search will find all the extents that end after
 637                  * our range starts
 638                  */
 639                 node = tree_search(tree, start);
 640                 if (!node)
 641                         break;
 642
 643                 state = rb_entry(node, struct extent_state, rb_node);
 644
 645                 if (state->start > end)
 646                         goto out;
 647
 648                 if (state->state & bits) {
 649                         start = state->start;
 650                         atomic_inc(&state->refs);
 651                         wait_on_state(tree, state);
 652                         free_extent_state(state);
 653                         goto again;
 654                 }
 655                 start = state->end + 1;
 656
 657                 if (start > end)
 658                         break;
 659
 660                 cond_resched_lock(&tree->lock);
 661         }
 662 out:
 663         spin_unlock(&tree->lock);
 664         return 0;
 665 }
 666
 667 static void set_state_bits(struct extent_io_tree *tree,
 668                            struct extent_state *state,
 669                            int *bits)
 670 {
 671         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 672
 673         set_state_cb(tree, state, bits);
 674         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 675                 u64 range = state->end - state->start + 1;
 676                 tree->dirty_bytes += range;
 677         }
 678         state->state |= bits_to_set;
 679 }
 680
 681 static void cache_state(struct extent_state *state,
 682                         struct extent_state **cached_ptr)
 683 {
 684         if (cached_ptr && !(*cached_ptr)) {
 685                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 686                         *cached_ptr = state;
 687                         atomic_inc(&state->refs);
 688                 }
 689         }
 690 }
 691
 692 static void uncache_state(struct extent_state **cached_ptr)
 693 {
 694         if (cached_ptr && (*cached_ptr)) {
 695                 struct extent_state *state = *cached_ptr;
 696                 *cached_ptr = NULL;
 697                 free_extent_state(state);
 698         }
 699 }
 700
 701 /*
 702  * set some bits on a range in the tree.  This may require allocations or
 703  * sleeping, so the gfp mask is used to indicate what is allowed.
 704  *
 705  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 706  * part of the range already has the desired bits set.  The start of the
 707  * existing range is returned in failed_start in this case.
 708  *
 709  * [start, end] is inclusive This takes the tree lock.
 710  */
 711
 712 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 713                    int bits, int exclusive_bits, u64 *failed_start,
 714                    struct extent_state **cached_state, gfp_t mask)
 715 {
 716         struct extent_state *state;
 717         struct extent_state *prealloc = NULL;
 718         struct rb_node *node;
 719         int err = 0;
 720         u64 last_start;
 721         u64 last_end;
 722
 723         bits |= EXTENT_FIRST_DELALLOC;
 724 again:
 725         if (!prealloc && (mask & __GFP_WAIT)) {
 726                 prealloc = alloc_extent_state(mask);
 727                 BUG_ON(!prealloc);
 728         }
 729
 730         spin_lock(&tree->lock);
 731         if (cached_state && *cached_state) {
 732                 state = *cached_state;
 733                 if (state->start <= start && state->end > start &&
 734                     state->tree) {
 735                         node = &state->rb_node;
 736                         goto hit_next;
 737                 }
 738         }
 739         /*
 740          * this search will find all the extents that end after
 741          * our range starts.
 742          */
 743         node = tree_search(tree, start);
 744         if (!node) {
 745                 prealloc = alloc_extent_state_atomic(prealloc);
 746                 BUG_ON(!prealloc);
 747                 err = insert_state(tree, prealloc, start, end, &bits);
 748                 prealloc = NULL;
 749                 BUG_ON(err == -EEXIST);
 750                 goto out;
 751         }
 752         state = rb_entry(node, struct extent_state, rb_node);
 753 hit_next:
 754         last_start = state->start;
 755         last_end = state->end;
 756
 757         /*
 758          * | ---- desired range ---- |
 759          * | state |
 760          *
 761          * Just lock what we found and keep going
 762          */
 763         if (state->start == start && state->end <= end) {
 764                 struct rb_node *next_node;
 765                 if (state->state & exclusive_bits) {
 766                         *failed_start = state->start;
 767                         err = -EEXIST;
 768                         goto out;
 769                 }
 770
 771                 set_state_bits(tree, state, &bits);
 772
 773                 cache_state(state, cached_state);
 774                 merge_state(tree, state);
 775                 if (last_end == (u64)-1)
 776                         goto out;
 777
 778                 start = last_end + 1;
 779                 next_node = rb_next(&state->rb_node);
 780                 if (next_node && start < end && prealloc && !need_resched()) {
 781                         state = rb_entry(next_node, struct extent_state,
 782                                          rb_node);
 783                         if (state->start == start)
 784                                 goto hit_next;
 785                 }
 786                 goto search_again;
 787         }
 788
 789         /*
 790          *     | ---- desired range ---- |
 791          * | state |
 792          *   or
 793          * | ------------- state -------------- |
 794          *
 795          * We need to split the extent we found, and may flip bits on
 796          * second half.
 797          *
 798          * If the extent we found extends past our
 799          * range, we just split and search again.  It'll get split
 800          * again the next time though.
 801          *
 802          * If the extent we found is inside our range, we set the
 803          * desired bit on it.
 804          */
 805         if (state->start < start) {
 806                 if (state->state & exclusive_bits) {
 807                         *failed_start = start;
 808                         err = -EEXIST;
 809                         goto out;
 810                 }
 811
 812                 prealloc = alloc_extent_state_atomic(prealloc);
 813                 BUG_ON(!prealloc);
 814                 err = split_state(tree, state, prealloc, start);
 815                 BUG_ON(err == -EEXIST);
 816                 prealloc = NULL;
 817                 if (err)
 818                         goto out;
 819                 if (state->end <= end) {
 820                         set_state_bits(tree, state, &bits);
 821                         cache_state(state, cached_state);
 822                         merge_state(tree, state);
 823                         if (last_end == (u64)-1)
 824                                 goto out;
 825                         start = last_end + 1;
 826                 }
 827                 goto search_again;
 828         }
 829         /*
 830          * | ---- desired range ---- |
 831          *     | state | or               | state |
 832          *
 833          * There's a hole, we need to insert something in it and
 834          * ignore the extent we found.
 835          */
 836         if (state->start > start) {
 837                 u64 this_end;
 838                 if (end < last_start)
 839                         this_end = end;
 840                 else
 841                         this_end = last_start - 1;
 842
 843                 prealloc = alloc_extent_state_atomic(prealloc);
 844                 BUG_ON(!prealloc);
 845
 846                 /*
 847                  * Avoid to free 'prealloc' if it can be merged with
 848                  * the later extent.
 849                  */
 850                 err = insert_state(tree, prealloc, start, this_end,
 851                                    &bits);
 852                 BUG_ON(err == -EEXIST);
 853                 if (err) {
 854                         free_extent_state(prealloc);
 855                         prealloc = NULL;
 856                         goto out;
 857                 }
 858                 cache_state(prealloc, cached_state);
 859                 prealloc = NULL;
 860                 start = this_end + 1;
 861                 goto search_again;
 862         }
 863         /*
 864          * | ---- desired range ---- |
 865          *                        | state |
 866          * We need to split the extent, and set the bit
 867          * on the first half
 868          */
 869         if (state->start <= end && state->end > end) {
 870                 if (state->state & exclusive_bits) {
 871                         *failed_start = start;
 872                         err = -EEXIST;
 873                         goto out;
 874                 }
 875
 876                 prealloc = alloc_extent_state_atomic(prealloc);
 877                 BUG_ON(!prealloc);
 878                 err = split_state(tree, state, prealloc, end + 1);
 879                 BUG_ON(err == -EEXIST);
 880
 881                 set_state_bits(tree, prealloc, &bits);
 882                 cache_state(prealloc, cached_state);
 883                 merge_state(tree, prealloc);
 884                 prealloc = NULL;
 885                 goto out;
 886         }
 887
 888         goto search_again;
 889
 890 out:
 891         spin_unlock(&tree->lock);
 892         if (prealloc)
 893                 free_extent_state(prealloc);
 894
 895         return err;
 896
 897 search_again:
 898         if (start > end)
 899                 goto out;
 900         spin_unlock(&tree->lock);
 901         if (mask & __GFP_WAIT)
 902                 cond_resched();
 903         goto again;
 904 }
 905
 906 /**
 907  * convert_extent - convert all bits in a given range from one bit to another
 908  * @tree:       the io tree to search
 909  * @start:      the start offset in bytes
 910  * @end:        the end offset in bytes (inclusive)
 911  * @bits:       the bits to set in this range
 912  * @clear_bits: the bits to clear in this range
 913  * @mask:       the allocation mask
 914  *
 915  * This will go through and set bits for the given range.  If any states exist
 916  * already in this range they are set with the given bit and cleared of the
 917  * clear_bits.  This is only meant to be used by things that are mergeable, ie
 918  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
 919  * boundary bits like LOCK.
 920  */
 921 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 922                        int bits, int clear_bits, gfp_t mask)
 923 {
 924         struct extent_state *state;
 925         struct extent_state *prealloc = NULL;
 926         struct rb_node *node;
 927         int err = 0;
 928         u64 last_start;
 929         u64 last_end;
 930
 931 again:
 932         if (!prealloc && (mask & __GFP_WAIT)) {
 933                 prealloc = alloc_extent_state(mask);
 934                 if (!prealloc)
 935                         return -ENOMEM;
 936         }
 937
 938         spin_lock(&tree->lock);
 939         /*
 940          * this search will find all the extents that end after
 941          * our range starts.
 942          */
 943         node = tree_search(tree, start);
 944         if (!node) {
 945                 prealloc = alloc_extent_state_atomic(prealloc);
 946                 if (!prealloc) {
 947                         err = -ENOMEM;
 948                         goto out;
 949                 }
 950                 err = insert_state(tree, prealloc, start, end, &bits);
 951                 prealloc = NULL;
 952                 BUG_ON(err == -EEXIST);
 953                 goto out;
 954         }
 955         state = rb_entry(node, struct extent_state, rb_node);
 956 hit_next:
 957         last_start = state->start;
 958         last_end = state->end;
 959
 960         /*
 961          * | ---- desired range ---- |
 962          * | state |
 963          *
 964          * Just lock what we found and keep going
 965          */
 966         if (state->start == start && state->end <= end) {
 967                 struct rb_node *next_node;
 968
 969                 set_state_bits(tree, state, &bits);
 970                 clear_state_bit(tree, state, &clear_bits, 0);
 971                 if (last_end == (u64)-1)
 972                         goto out;
 973
 974                 start = last_end + 1;
 975                 next_node = rb_next(&state->rb_node);
 976                 if (next_node && start < end && prealloc && !need_resched()) {
 977                         state = rb_entry(next_node, struct extent_state,
 978                                          rb_node);
 979                         if (state->start == start)
 980                                 goto hit_next;
 981                 }
 982                 goto search_again;
 983         }
 984
 985         /*
 986          *     | ---- desired range ---- |
 987          * | state |
 988          *   or
 989          * | ------------- state -------------- |
 990          *
 991          * We need to split the extent we found, and may flip bits on
 992          * second half.
 993          *
 994          * If the extent we found extends past our
 995          * range, we just split and search again.  It'll get split
 996          * again the next time though.
 997          *
 998          * If the extent we found is inside our range, we set the
 999          * desired bit on it.
1000          */
1001         if (state->start < start) {
1002                 prealloc = alloc_extent_state_atomic(prealloc);
1003                 if (!prealloc) {
1004                         err = -ENOMEM;
1005                         goto out;
1006                 }
1007                 err = split_state(tree, state, prealloc, start);
1008                 BUG_ON(err == -EEXIST);
1009                 prealloc = NULL;
1010                 if (err)
1011                         goto out;
1012                 if (state->end <= end) {
1013                         set_state_bits(tree, state, &bits);
1014                         clear_state_bit(tree, state, &clear_bits, 0);
1015                         if (last_end == (u64)-1)
1016                                 goto out;
1017                         start = last_end + 1;
1018                 }
1019                 goto search_again;
1020         }
1021         /*
1022          * | ---- desired range ---- |
1023          *     | state | or               | state |
1024          *
1025          * There's a hole, we need to insert something in it and
1026          * ignore the extent we found.
1027          */
1028         if (state->start > start) {
1029                 u64 this_end;
1030                 if (end < last_start)
1031                         this_end = end;
1032                 else
1033                         this_end = last_start - 1;
1034
1035                 prealloc = alloc_extent_state_atomic(prealloc);
1036                 if (!prealloc) {
1037                         err = -ENOMEM;
1038                         goto out;
1039                 }
1040
1041                 /*
1042                  * Avoid to free 'prealloc' if it can be merged with
1043                  * the later extent.
1044                  */
1045                 err = insert_state(tree, prealloc, start, this_end,
1046                                    &bits);
1047                 BUG_ON(err == -EEXIST);
1048                 if (err) {
1049                         free_extent_state(prealloc);
1050                         prealloc = NULL;
1051                         goto out;
1052                 }
1053                 prealloc = NULL;
1054                 start = this_end + 1;
1055                 goto search_again;
1056         }
1057         /*
1058          * | ---- desired range ---- |
1059          *                        | state |
1060          * We need to split the extent, and set the bit
1061          * on the first half
1062          */
1063         if (state->start <= end && state->end > end) {
1064                 prealloc = alloc_extent_state_atomic(prealloc);
1065                 if (!prealloc) {
1066                         err = -ENOMEM;
1067                         goto out;
1068                 }
1069
1070                 err = split_state(tree, state, prealloc, end + 1);
1071                 BUG_ON(err == -EEXIST);
1072
1073                 set_state_bits(tree, prealloc, &bits);
1074                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1075                 prealloc = NULL;
1076                 goto out;
1077         }
1078
1079         goto search_again;
1080
1081 out:
1082         spin_unlock(&tree->lock);
1083         if (prealloc)
1084                 free_extent_state(prealloc);
1085
1086         return err;
1087
1088 search_again:
1089         if (start > end)
1090                 goto out;
1091         spin_unlock(&tree->lock);
1092         if (mask & __GFP_WAIT)
1093                 cond_resched();
1094         goto again;
1095 }
1096
1097 /* wrappers around set/clear extent bit */
1098 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1099                      gfp_t mask)
1100 {
1101         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
1102                               NULL, mask);
1103 }
1104
1105 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1106                     int bits, gfp_t mask)
1107 {
1108         return set_extent_bit(tree, start, end, bits, 0, NULL,
1109                               NULL, mask);
1110 }
1111
1112 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1113                       int bits, gfp_t mask)
1114 {
1115         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1116 }
1117
1118 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1119                         struct extent_state **cached_state, gfp_t mask)
1120 {
1121         return set_extent_bit(tree, start, end,
1122                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1123                               0, NULL, cached_state, mask);
1124 }
1125
1126 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1127                        gfp_t mask)
1128 {
1129         return clear_extent_bit(tree, start, end,
1130                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1131                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1132 }
1133
1134 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1135                      gfp_t mask)
1136 {
1137         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
1138                               NULL, mask);
1139 }
1140
1141 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1142                         struct extent_state **cached_state, gfp_t mask)
1143 {
1144         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
1145                               NULL, cached_state, mask);
1146 }
1147
1148 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
1149                                  u64 end, struct extent_state **cached_state,
1150                                  gfp_t mask)
1151 {
1152         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1153                                 cached_state, mask);
1154 }
1155
1156 /*
1157  * either insert or lock state struct between start and end use mask to tell
1158  * us if waiting is desired.
1159  */
1160 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1161                      int bits, struct extent_state **cached_state, gfp_t mask)
1162 {
1163         int err;
1164         u64 failed_start;
1165         while (1) {
1166                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1167                                      EXTENT_LOCKED, &failed_start,
1168                                      cached_state, mask);
1169                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
1170                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1171                         start = failed_start;
1172                 } else {
1173                         break;
1174                 }
1175                 WARN_ON(start > end);
1176         }
1177         return err;
1178 }
1179
1180 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1181 {
1182         return lock_extent_bits(tree, start, end, 0, NULL, mask);
1183 }
1184
1185 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1186                     gfp_t mask)
1187 {
1188         int err;
1189         u64 failed_start;
1190
1191         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1192                              &failed_start, NULL, mask);
1193         if (err == -EEXIST) {
1194                 if (failed_start > start)
1195                         clear_extent_bit(tree, start, failed_start - 1,
1196                                          EXTENT_LOCKED, 1, 0, NULL, mask);
1197                 return 0;
1198         }
1199         return 1;
1200 }
1201
1202 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1203                          struct extent_state **cached, gfp_t mask)
1204 {
1205         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1206                                 mask);
1207 }
1208
1209 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1210 {
1211         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1212                                 mask);
1213 }
1214
1215 /*
1216  * helper function to set both pages and extents in the tree writeback
1217  */
1218 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1219 {
1220         unsigned long index = start >> PAGE_CACHE_SHIFT;
1221         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1222         struct page *page;
1223
1224         while (index <= end_index) {
1225                 page = find_get_page(tree->mapping, index);
1226                 BUG_ON(!page);
1227                 set_page_writeback(page);
1228                 page_cache_release(page);
1229                 index++;
1230         }
1231         return 0;
1232 }
1233
1234 /* find the first state struct with 'bits' set after 'start', and
1235  * return it.  tree->lock must be held.  NULL will returned if
1236  * nothing was found after 'start'
1237  */
1238 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1239                                                  u64 start, int bits)
1240 {
1241         struct rb_node *node;
1242         struct extent_state *state;
1243
1244         /*
1245          * this search will find all the extents that end after
1246          * our range starts.
1247          */
1248         node = tree_search(tree, start);
1249         if (!node)
1250                 goto out;
1251
1252         while (1) {
1253                 state = rb_entry(node, struct extent_state, rb_node);
1254                 if (state->end >= start && (state->state & bits))
1255                         return state;
1256
1257                 node = rb_next(node);
1258                 if (!node)
1259                         break;
1260         }
1261 out:
1262         return NULL;
1263 }
1264
1265 /*
1266  * find the first offset in the io tree with 'bits' set. zero is
1267  * returned if we find something, and *start_ret and *end_ret are
1268  * set to reflect the state struct that was found.
1269  *
1270  * If nothing was found, 1 is returned, < 0 on error
1271  */
1272 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1273                           u64 *start_ret, u64 *end_ret, int bits)
1274 {
1275         struct extent_state *state;
1276         int ret = 1;
1277
1278         spin_lock(&tree->lock);
1279         state = find_first_extent_bit_state(tree, start, bits);
1280         if (state) {
1281                 *start_ret = state->start;
1282                 *end_ret = state->end;
1283                 ret = 0;
1284         }
1285         spin_unlock(&tree->lock);
1286         return ret;
1287 }
1288
1289 /*
1290  * find a contiguous range of bytes in the file marked as delalloc, not
1291  * more than 'max_bytes'.  start and end are used to return the range,
1292  *
1293  * 1 is returned if we find something, 0 if nothing was in the tree
1294  */
1295 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1296                                         u64 *start, u64 *end, u64 max_bytes,
1297                                         struct extent_state **cached_state)
1298 {
1299         struct rb_node *node;
1300         struct extent_state *state;
1301         u64 cur_start = *start;
1302         u64 found = 0;
1303         u64 total_bytes = 0;
1304
1305         spin_lock(&tree->lock);
1306
1307         /*
1308          * this search will find all the extents that end after
1309          * our range starts.
1310          */
1311         node = tree_search(tree, cur_start);
1312         if (!node) {
1313                 if (!found)
1314                         *end = (u64)-1;
1315                 goto out;
1316         }
1317
1318         while (1) {
1319                 state = rb_entry(node, struct extent_state, rb_node);
1320                 if (found && (state->start != cur_start ||
1321                               (state->state & EXTENT_BOUNDARY))) {
1322                         goto out;
1323                 }
1324                 if (!(state->state & EXTENT_DELALLOC)) {
1325                         if (!found)
1326                                 *end = state->end;
1327                         goto out;
1328                 }
1329                 if (!found) {
1330                         *start = state->start;
1331                         *cached_state = state;
1332                         atomic_inc(&state->refs);
1333                 }
1334                 found++;
1335                 *end = state->end;
1336                 cur_start = state->end + 1;
1337                 node = rb_next(node);
1338                 if (!node)
1339                         break;
1340                 total_bytes += state->end - state->start + 1;
1341                 if (total_bytes >= max_bytes)
1342                         break;
1343         }
1344 out:
1345         spin_unlock(&tree->lock);
1346         return found;
1347 }
1348
1349 static noinline int __unlock_for_delalloc(struct inode *inode,
1350                                           struct page *locked_page,
1351                                           u64 start, u64 end)
1352 {
1353         int ret;
1354         struct page *pages[16];
1355         unsigned long index = start >> PAGE_CACHE_SHIFT;
1356         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1357         unsigned long nr_pages = end_index - index + 1;
1358         int i;
1359
1360         if (index == locked_page->index && end_index == index)
1361                 return 0;
1362
1363         while (nr_pages > 0) {
1364                 ret = find_get_pages_contig(inode->i_mapping, index,
1365                                      min_t(unsigned long, nr_pages,
1366                                      ARRAY_SIZE(pages)), pages);
1367                 for (i = 0; i < ret; i++) {
1368                         if (pages[i] != locked_page)
1369                                 unlock_page(pages[i]);
1370                         page_cache_release(pages[i]);
1371                 }
1372                 nr_pages -= ret;
1373                 index += ret;
1374                 cond_resched();
1375         }
1376         return 0;
1377 }
1378
1379 static noinline int lock_delalloc_pages(struct inode *inode,
1380                                         struct page *locked_page,
1381                                         u64 delalloc_start,
1382                                         u64 delalloc_end)
1383 {
1384         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1385         unsigned long start_index = index;
1386         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1387         unsigned long pages_locked = 0;
1388         struct page *pages[16];
1389         unsigned long nrpages;
1390         int ret;
1391         int i;
1392
1393         /* the caller is responsible for locking the start index */
1394         if (index == locked_page->index && index == end_index)
1395                 return 0;
1396
1397         /* skip the page at the start index */
1398         nrpages = end_index - index + 1;
1399         while (nrpages > 0) {
1400                 ret = find_get_pages_contig(inode->i_mapping, index,
1401                                      min_t(unsigned long,
1402                                      nrpages, ARRAY_SIZE(pages)), pages);
1403                 if (ret == 0) {
1404                         ret = -EAGAIN;
1405                         goto done;
1406                 }
1407                 /* now we have an array of pages, lock them all */
1408                 for (i = 0; i < ret; i++) {
1409                         /*
1410                          * the caller is taking responsibility for
1411                          * locked_page
1412                          */
1413                         if (pages[i] != locked_page) {
1414                                 lock_page(pages[i]);
1415                                 if (!PageDirty(pages[i]) ||
1416                                     pages[i]->mapping != inode->i_mapping) {
1417                                         ret = -EAGAIN;
1418                                         unlock_page(pages[i]);
1419                                         page_cache_release(pages[i]);
1420                                         goto done;
1421                                 }
1422                         }
1423                         page_cache_release(pages[i]);
1424                         pages_locked++;
1425                 }
1426                 nrpages -= ret;
1427                 index += ret;
1428                 cond_resched();
1429         }
1430         ret = 0;
1431 done:
1432         if (ret && pages_locked) {
1433                 __unlock_for_delalloc(inode, locked_page,
1434                               delalloc_start,
1435                               ((u64)(start_index + pages_locked - 1)) <<
1436                               PAGE_CACHE_SHIFT);
1437         }
1438         return ret;
1439 }
1440
1441 /*
1442  * find a contiguous range of bytes in the file marked as delalloc, not
1443  * more than 'max_bytes'.  start and end are used to return the range,
1444  *
1445  * 1 is returned if we find something, 0 if nothing was in the tree
1446  */
1447 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1448                                              struct extent_io_tree *tree,
1449                                              struct page *locked_page,
1450                                              u64 *start, u64 *end,
1451                                              u64 max_bytes)
1452 {
1453         u64 delalloc_start;
1454         u64 delalloc_end;
1455         u64 found;
1456         struct extent_state *cached_state = NULL;
1457         int ret;
1458         int loops = 0;
1459
1460 again:
1461         /* step one, find a bunch of delalloc bytes starting at start */
1462         delalloc_start = *start;
1463         delalloc_end = 0;
1464         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1465                                     max_bytes, &cached_state);
1466         if (!found || delalloc_end <= *start) {
1467                 *start = delalloc_start;
1468                 *end = delalloc_end;
1469                 free_extent_state(cached_state);
1470                 return found;
1471         }
1472
1473         /*
1474          * start comes from the offset of locked_page.  We have to lock
1475          * pages in order, so we can't process delalloc bytes before
1476          * locked_page
1477          */
1478         if (delalloc_start < *start)
1479                 delalloc_start = *start;
1480
1481         /*
1482          * make sure to limit the number of pages we try to lock down
1483          * if we're looping.
1484          */
1485         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1486                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1487
1488         /* step two, lock all the pages after the page that has start */
1489         ret = lock_delalloc_pages(inode, locked_page,
1490                                   delalloc_start, delalloc_end);
1491         if (ret == -EAGAIN) {
1492                 /* some of the pages are gone, lets avoid looping by
1493                  * shortening the size of the delalloc range we're searching
1494                  */
1495                 free_extent_state(cached_state);
1496                 if (!loops) {
1497                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1498                         max_bytes = PAGE_CACHE_SIZE - offset;
1499                         loops = 1;
1500                         goto again;
1501                 } else {
1502                         found = 0;
1503                         goto out_failed;
1504                 }
1505         }
1506         BUG_ON(ret);
1507
1508         /* step three, lock the state bits for the whole range */
1509         lock_extent_bits(tree, delalloc_start, delalloc_end,
1510                          0, &cached_state, GFP_NOFS);
1511
1512         /* then test to make sure it is all still delalloc */
1513         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1514                              EXTENT_DELALLOC, 1, cached_state);
1515         if (!ret) {
1516                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1517                                      &cached_state, GFP_NOFS);
1518                 __unlock_for_delalloc(inode, locked_page,
1519                               delalloc_start, delalloc_end);
1520                 cond_resched();
1521                 goto again;
1522         }
1523         free_extent_state(cached_state);
1524         *start = delalloc_start;
1525         *end = delalloc_end;
1526 out_failed:
1527         return found;
1528 }
1529
1530 int extent_clear_unlock_delalloc(struct inode *inode,
1531                                 struct extent_io_tree *tree,
1532                                 u64 start, u64 end, struct page *locked_page,
1533                                 unsigned long op)
1534 {
1535         int ret;
1536         struct page *pages[16];
1537         unsigned long index = start >> PAGE_CACHE_SHIFT;
1538         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1539         unsigned long nr_pages = end_index - index + 1;
1540         int i;
1541         int clear_bits = 0;
1542
1543         if (op & EXTENT_CLEAR_UNLOCK)
1544                 clear_bits |= EXTENT_LOCKED;
1545         if (op & EXTENT_CLEAR_DIRTY)
1546                 clear_bits |= EXTENT_DIRTY;
1547
1548         if (op & EXTENT_CLEAR_DELALLOC)
1549                 clear_bits |= EXTENT_DELALLOC;
1550
1551         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1552         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1553                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1554                     EXTENT_SET_PRIVATE2)))
1555                 return 0;
1556
1557         while (nr_pages > 0) {
1558                 ret = find_get_pages_contig(inode->i_mapping, index,
1559                                      min_t(unsigned long,
1560                                      nr_pages, ARRAY_SIZE(pages)), pages);
1561                 for (i = 0; i < ret; i++) {
1562
1563                         if (op & EXTENT_SET_PRIVATE2)
1564                                 SetPagePrivate2(pages[i]);
1565
1566                         if (pages[i] == locked_page) {
1567                                 page_cache_release(pages[i]);
1568                                 continue;
1569                         }
1570                         if (op & EXTENT_CLEAR_DIRTY)
1571                                 clear_page_dirty_for_io(pages[i]);
1572                         if (op & EXTENT_SET_WRITEBACK)
1573                                 set_page_writeback(pages[i]);
1574                         if (op & EXTENT_END_WRITEBACK)
1575                                 end_page_writeback(pages[i]);
1576                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1577                                 unlock_page(pages[i]);
1578                         page_cache_release(pages[i]);
1579                 }
1580                 nr_pages -= ret;
1581                 index += ret;
1582                 cond_resched();
1583         }
1584         return 0;
1585 }
1586
1587 /*
1588  * count the number of bytes in the tree that have a given bit(s)
1589  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1590  * cached.  The total number found is returned.
1591  */
1592 u64 count_range_bits(struct extent_io_tree *tree,
1593                      u64 *start, u64 search_end, u64 max_bytes,
1594                      unsigned long bits, int contig)
1595 {
1596         struct rb_node *node;
1597         struct extent_state *state;
1598         u64 cur_start = *start;
1599         u64 total_bytes = 0;
1600         u64 last = 0;
1601         int found = 0;
1602
1603         if (search_end <= cur_start) {
1604                 WARN_ON(1);
1605                 return 0;
1606         }
1607
1608         spin_lock(&tree->lock);
1609         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1610                 total_bytes = tree->dirty_bytes;
1611                 goto out;
1612         }
1613         /*
1614          * this search will find all the extents that end after
1615          * our range starts.
1616          */
1617         node = tree_search(tree, cur_start);
1618         if (!node)
1619                 goto out;
1620
1621         while (1) {
1622                 state = rb_entry(node, struct extent_state, rb_node);
1623                 if (state->start > search_end)
1624                         break;
1625                 if (contig && found && state->start > last + 1)
1626                         break;
1627                 if (state->end >= cur_start && (state->state & bits) == bits) {
1628                         total_bytes += min(search_end, state->end) + 1 -
1629                                        max(cur_start, state->start);
1630                         if (total_bytes >= max_bytes)
1631                                 break;
1632                         if (!found) {
1633                                 *start = max(cur_start, state->start);
1634                                 found = 1;
1635                         }
1636                         last = state->end;
1637                 } else if (contig && found) {
1638                         break;
1639                 }
1640                 node = rb_next(node);
1641                 if (!node)
1642                         break;
1643         }
1644 out:
1645         spin_unlock(&tree->lock);
1646         return total_bytes;
1647 }
1648
1649 /*
1650  * set the private field for a given byte offset in the tree.  If there isn't
1651  * an extent_state there already, this does nothing.
1652  */
1653 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1654 {
1655         struct rb_node *node;
1656         struct extent_state *state;
1657         int ret = 0;
1658
1659         spin_lock(&tree->lock);
1660         /*
1661          * this search will find all the extents that end after
1662          * our range starts.
1663          */
1664         node = tree_search(tree, start);
1665         if (!node) {
1666                 ret = -ENOENT;
1667                 goto out;
1668         }
1669         state = rb_entry(node, struct extent_state, rb_node);
1670         if (state->start != start) {
1671                 ret = -ENOENT;
1672                 goto out;
1673         }
1674         state->private = private;
1675 out:
1676         spin_unlock(&tree->lock);
1677         return ret;
1678 }
1679
1680 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1681 {
1682         struct rb_node *node;
1683         struct extent_state *state;
1684         int ret = 0;
1685
1686         spin_lock(&tree->lock);
1687         /*
1688          * this search will find all the extents that end after
1689          * our range starts.
1690          */
1691         node = tree_search(tree, start);
1692         if (!node) {
1693                 ret = -ENOENT;
1694                 goto out;
1695         }
1696         state = rb_entry(node, struct extent_state, rb_node);
1697         if (state->start != start) {
1698                 ret = -ENOENT;
1699                 goto out;
1700         }
1701         *private = state->private;
1702 out:
1703         spin_unlock(&tree->lock);
1704         return ret;
1705 }
1706
1707 /*
1708  * searches a range in the state tree for a given mask.
1709  * If 'filled' == 1, this returns 1 only if every extent in the tree
1710  * has the bits set.  Otherwise, 1 is returned if any bit in the
1711  * range is found set.
1712  */
1713 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1714                    int bits, int filled, struct extent_state *cached)
1715 {
1716         struct extent_state *state = NULL;
1717         struct rb_node *node;
1718         int bitset = 0;
1719
1720         spin_lock(&tree->lock);
1721         if (cached && cached->tree && cached->start <= start &&
1722             cached->end > start)
1723                 node = &cached->rb_node;
1724         else
1725                 node = tree_search(tree, start);
1726         while (node && start <= end) {
1727                 state = rb_entry(node, struct extent_state, rb_node);
1728
1729                 if (filled && state->start > start) {
1730                         bitset = 0;
1731                         break;
1732                 }
1733
1734                 if (state->start > end)
1735                         break;
1736
1737                 if (state->state & bits) {
1738                         bitset = 1;
1739                         if (!filled)
1740                                 break;
1741                 } else if (filled) {
1742                         bitset = 0;
1743                         break;
1744                 }
1745
1746                 if (state->end == (u64)-1)
1747                         break;
1748
1749                 start = state->end + 1;
1750                 if (start > end)
1751                         break;
1752                 node = rb_next(node);
1753                 if (!node) {
1754                         if (filled)
1755                                 bitset = 0;
1756                         break;
1757                 }
1758         }
1759         spin_unlock(&tree->lock);
1760         return bitset;
1761 }
1762
1763 /*
1764  * helper function to set a given page up to date if all the
1765  * extents in the tree for that page are up to date
1766  */
1767 static int check_page_uptodate(struct extent_io_tree *tree,
1768                                struct page *page)
1769 {
1770         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1771         u64 end = start + PAGE_CACHE_SIZE - 1;
1772         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1773                 SetPageUptodate(page);
1774         return 0;
1775 }
1776
1777 /*
1778  * helper function to unlock a page if all the extents in the tree
1779  * for that page are unlocked
1780  */
1781 static int check_page_locked(struct extent_io_tree *tree,
1782                              struct page *page)
1783 {
1784         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1785         u64 end = start + PAGE_CACHE_SIZE - 1;
1786         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1787                 unlock_page(page);
1788         return 0;
1789 }
1790
1791 /*
1792  * helper function to end page writeback if all the extents
1793  * in the tree for that page are done with writeback
1794  */
1795 static int check_page_writeback(struct extent_io_tree *tree,
1796                              struct page *page)
1797 {
1798         end_page_writeback(page);
1799         return 0;
1800 }
1801
1802 /*
1803  * When IO fails, either with EIO or csum verification fails, we
1804  * try other mirrors that might have a good copy of the data.  This
1805  * io_failure_record is used to record state as we go through all the
1806  * mirrors.  If another mirror has good data, the page is set up to date
1807  * and things continue.  If a good mirror can't be found, the original
1808  * bio end_io callback is called to indicate things have failed.
1809  */
1810 struct io_failure_record {
1811         struct page *page;
1812         u64 start;
1813         u64 len;
1814         u64 logical;
1815         unsigned long bio_flags;
1816         int this_mirror;
1817         int failed_mirror;
1818         int in_validation;
1819 };
1820
1821 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1822                                 int did_repair)
1823 {
1824         int ret;
1825         int err = 0;
1826         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1827
1828         set_state_private(failure_tree, rec->start, 0);
1829         ret = clear_extent_bits(failure_tree, rec->start,
1830                                 rec->start + rec->len - 1,
1831                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1832         if (ret)
1833                 err = ret;
1834
1835         if (did_repair) {
1836                 ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1837                                         rec->start + rec->len - 1,
1838                                         EXTENT_DAMAGED, GFP_NOFS);
1839                 if (ret && !err)
1840                         err = ret;
1841         }
1842
1843         kfree(rec);
1844         return err;
1845 }
1846
1847 static void repair_io_failure_callback(struct bio *bio, int err)
1848 {
1849         complete(bio->bi_private);
1850 }
1851
1852 /*
1853  * this bypasses the standard btrfs submit functions deliberately, as
1854  * the standard behavior is to write all copies in a raid setup. here we only
1855  * want to write the one bad copy. so we do the mapping for ourselves and issue
1856  * submit_bio directly.
1857  * to avoid any synchonization issues, wait for the data after writing, which
1858  * actually prevents the read that triggered the error from finishing.
1859  * currently, there can be no more than two copies of every data bit. thus,
1860  * exactly one rewrite is required.
1861  */
1862 int repair_io_failure(struct btrfs_mapping_tree *map_tree, u64 start,
1863                         u64 length, u64 logical, struct page *page,
1864                         int mirror_num)
1865 {
1866         struct bio *bio;
1867         struct btrfs_device *dev;
1868         DECLARE_COMPLETION_ONSTACK(compl);
1869         u64 map_length = 0;
1870         u64 sector;
1871         struct btrfs_bio *bbio = NULL;
1872         int ret;
1873
1874         BUG_ON(!mirror_num);
1875
1876         bio = bio_alloc(GFP_NOFS, 1);
1877         if (!bio)
1878                 return -EIO;
1879         bio->bi_private = &compl;
1880         bio->bi_end_io = repair_io_failure_callback;
1881         bio->bi_size = 0;
1882         map_length = length;
1883
1884         ret = btrfs_map_block(map_tree, WRITE, logical,
1885                               &map_length, &bbio, mirror_num);
1886         if (ret) {
1887                 bio_put(bio);
1888                 return -EIO;
1889         }
1890         BUG_ON(mirror_num != bbio->mirror_num);
1891         sector = bbio->stripes[mirror_num-1].physical >> 9;
1892         bio->bi_sector = sector;
1893         dev = bbio->stripes[mirror_num-1].dev;
1894         kfree(bbio);
1895         if (!dev || !dev->bdev || !dev->writeable) {
1896                 bio_put(bio);
1897                 return -EIO;
1898         }
1899         bio->bi_bdev = dev->bdev;
1900         bio_add_page(bio, page, length, start-page_offset(page));
1901         btrfsic_submit_bio(WRITE_SYNC, bio);
1902         wait_for_completion(&compl);
1903
1904         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
1905                 /* try to remap that extent elsewhere? */
1906                 bio_put(bio);
1907                 return -EIO;
1908         }
1909
1910         printk(KERN_INFO "btrfs read error corrected: ino %lu off %llu (dev %s "
1911                         "sector %llu)\n", page->mapping->host->i_ino, start,
1912                         dev->name, sector);
1913
1914         bio_put(bio);
1915         return 0;
1916 }
1917
1918 /*
1919  * each time an IO finishes, we do a fast check in the IO failure tree
1920  * to see if we need to process or clean up an io_failure_record
1921  */
1922 static int clean_io_failure(u64 start, struct page *page)
1923 {
1924         u64 private;
1925         u64 private_failure;
1926         struct io_failure_record *failrec;
1927         struct btrfs_mapping_tree *map_tree;
1928         struct extent_state *state;
1929         int num_copies;
1930         int did_repair = 0;
1931         int ret;
1932         struct inode *inode = page->mapping->host;
1933
1934         private = 0;
1935         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
1936                                 (u64)-1, 1, EXTENT_DIRTY, 0);
1937         if (!ret)
1938                 return 0;
1939
1940         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
1941                                 &private_failure);
1942         if (ret)
1943                 return 0;
1944
1945         failrec = (struct io_failure_record *)(unsigned long) private_failure;
1946         BUG_ON(!failrec->this_mirror);
1947
1948         if (failrec->in_validation) {
1949                 /* there was no real error, just free the record */
1950                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
1951                          failrec->start);
1952                 did_repair = 1;
1953                 goto out;
1954         }
1955
1956         spin_lock(&BTRFS_I(inode)->io_tree.lock);
1957         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
1958                                             failrec->start,
1959                                             EXTENT_LOCKED);
1960         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
1961
1962         if (state && state->start == failrec->start) {
1963                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
1964                 num_copies = btrfs_num_copies(map_tree, failrec->logical,
1965                                                 failrec->len);
1966                 if (num_copies > 1)  {
1967                         ret = repair_io_failure(map_tree, start, failrec->len,
1968                                                 failrec->logical, page,
1969                                                 failrec->failed_mirror);
1970                         did_repair = !ret;
1971                 }
1972         }
1973
1974 out:
1975         if (!ret)
1976                 ret = free_io_failure(inode, failrec, did_repair);
1977
1978         return ret;
1979 }
1980
1981 /*
1982  * this is a generic handler for readpage errors (default
1983  * readpage_io_failed_hook). if other copies exist, read those and write back
1984  * good data to the failed position. does not investigate in remapping the
1985  * failed extent elsewhere, hoping the device will be smart enough to do this as
1986  * needed
1987  */
1988
1989 static int bio_readpage_error(struct bio *failed_bio, struct page *page,
1990                                 u64 start, u64 end, int failed_mirror,
1991                                 struct extent_state *state)
1992 {
1993         struct io_failure_record *failrec = NULL;
1994         u64 private;
1995         struct extent_map *em;
1996         struct inode *inode = page->mapping->host;
1997         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1998         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1999         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2000         struct bio *bio;
2001         int num_copies;
2002         int ret;
2003         int read_mode;
2004         u64 logical;
2005
2006         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2007
2008         ret = get_state_private(failure_tree, start, &private);
2009         if (ret) {
2010                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2011                 if (!failrec)
2012                         return -ENOMEM;
2013                 failrec->start = start;
2014                 failrec->len = end - start + 1;
2015                 failrec->this_mirror = 0;
2016                 failrec->bio_flags = 0;
2017                 failrec->in_validation = 0;
2018
2019                 read_lock(&em_tree->lock);
2020                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2021                 if (!em) {
2022                         read_unlock(&em_tree->lock);
2023                         kfree(failrec);
2024                         return -EIO;
2025                 }
2026
2027                 if (em->start > start || em->start + em->len < start) {
2028                         free_extent_map(em);
2029                         em = NULL;
2030                 }
2031                 read_unlock(&em_tree->lock);
2032
2033                 if (!em || IS_ERR(em)) {
2034                         kfree(failrec);
2035                         return -EIO;
2036                 }
2037                 logical = start - em->start;
2038                 logical = em->block_start + logical;
2039                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2040                         logical = em->block_start;
2041                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2042                         extent_set_compress_type(&failrec->bio_flags,
2043                                                  em->compress_type);
2044                 }
2045                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2046                          "len=%llu\n", logical, start, failrec->len);
2047                 failrec->logical = logical;
2048                 free_extent_map(em);
2049
2050                 /* set the bits in the private failure tree */
2051                 ret = set_extent_bits(failure_tree, start, end,
2052                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2053                 if (ret >= 0)
2054                         ret = set_state_private(failure_tree, start,
2055                                                 (u64)(unsigned long)failrec);
2056                 /* set the bits in the inode's tree */
2057                 if (ret >= 0)
2058                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2059                                                 GFP_NOFS);
2060                 if (ret < 0) {
2061                         kfree(failrec);
2062                         return ret;
2063                 }
2064         } else {
2065                 failrec = (struct io_failure_record *)(unsigned long)private;
2066                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2067                          "start=%llu, len=%llu, validation=%d\n",
2068                          failrec->logical, failrec->start, failrec->len,
2069                          failrec->in_validation);
2070                 /*
2071                  * when data can be on disk more than twice, add to failrec here
2072                  * (e.g. with a list for failed_mirror) to make
2073                  * clean_io_failure() clean all those errors at once.
2074                  */
2075         }
2076         num_copies = btrfs_num_copies(
2077                               &BTRFS_I(inode)->root->fs_info->mapping_tree,
2078                               failrec->logical, failrec->len);
2079         if (num_copies == 1) {
2080                 /*
2081                  * we only have a single copy of the data, so don't bother with
2082                  * all the retry and error correction code that follows. no
2083                  * matter what the error is, it is very likely to persist.
2084                  */
2085                 pr_debug("bio_readpage_error: cannot repair, num_copies == 1. "
2086                          "state=%p, num_copies=%d, next_mirror %d, "
2087                          "failed_mirror %d\n", state, num_copies,
2088                          failrec->this_mirror, failed_mirror);
2089                 free_io_failure(inode, failrec, 0);
2090                 return -EIO;
2091         }
2092
2093         if (!state) {
2094                 spin_lock(&tree->lock);
2095                 state = find_first_extent_bit_state(tree, failrec->start,
2096                                                     EXTENT_LOCKED);
2097                 if (state && state->start != failrec->start)
2098                         state = NULL;
2099                 spin_unlock(&tree->lock);
2100         }
2101
2102         /*
2103          * there are two premises:
2104          *      a) deliver good data to the caller
2105          *      b) correct the bad sectors on disk
2106          */
2107         if (failed_bio->bi_vcnt > 1) {
2108                 /*
2109                  * to fulfill b), we need to know the exact failing sectors, as
2110                  * we don't want to rewrite any more than the failed ones. thus,
2111                  * we need separate read requests for the failed bio
2112                  *
2113                  * if the following BUG_ON triggers, our validation request got
2114                  * merged. we need separate requests for our algorithm to work.
2115                  */
2116                 BUG_ON(failrec->in_validation);
2117                 failrec->in_validation = 1;
2118                 failrec->this_mirror = failed_mirror;
2119                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2120         } else {
2121                 /*
2122                  * we're ready to fulfill a) and b) alongside. get a good copy
2123                  * of the failed sector and if we succeed, we have setup
2124                  * everything for repair_io_failure to do the rest for us.
2125                  */
2126                 if (failrec->in_validation) {
2127                         BUG_ON(failrec->this_mirror != failed_mirror);
2128                         failrec->in_validation = 0;
2129                         failrec->this_mirror = 0;
2130                 }
2131                 failrec->failed_mirror = failed_mirror;
2132                 failrec->this_mirror++;
2133                 if (failrec->this_mirror == failed_mirror)
2134                         failrec->this_mirror++;
2135                 read_mode = READ_SYNC;
2136         }
2137
2138         if (!state || failrec->this_mirror > num_copies) {
2139                 pr_debug("bio_readpage_error: (fail) state=%p, num_copies=%d, "
2140                          "next_mirror %d, failed_mirror %d\n", state,
2141                          num_copies, failrec->this_mirror, failed_mirror);
2142                 free_io_failure(inode, failrec, 0);
2143                 return -EIO;
2144         }
2145
2146         bio = bio_alloc(GFP_NOFS, 1);
2147         bio->bi_private = state;
2148         bio->bi_end_io = failed_bio->bi_end_io;
2149         bio->bi_sector = failrec->logical >> 9;
2150         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2151         bio->bi_size = 0;
2152
2153         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2154
2155         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2156                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2157                  failrec->this_mirror, num_copies, failrec->in_validation);
2158
2159         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2160                                          failrec->this_mirror,
2161                                          failrec->bio_flags, 0);
2162         return ret;
2163 }
2164
2165 /* lots and lots of room for performance fixes in the end_bio funcs */
2166
2167 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2168 {
2169         int uptodate = (err == 0);
2170         struct extent_io_tree *tree;
2171         int ret;
2172
2173         tree = &BTRFS_I(page->mapping->host)->io_tree;
2174
2175         if (tree->ops && tree->ops->writepage_end_io_hook) {
2176                 ret = tree->ops->writepage_end_io_hook(page, start,
2177                                                end, NULL, uptodate);
2178                 if (ret)
2179                         uptodate = 0;
2180         }
2181
2182         if (!uptodate && tree->ops &&
2183             tree->ops->writepage_io_failed_hook) {
2184                 ret = tree->ops->writepage_io_failed_hook(NULL, page,
2185                                                  start, end, NULL);
2186                 /* Writeback already completed */
2187                 if (ret == 0)
2188                         return 1;
2189         }
2190
2191         if (!uptodate) {
2192                 clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
2193                 ClearPageUptodate(page);
2194                 SetPageError(page);
2195         }
2196         return 0;
2197 }
2198
2199 /*
2200  * after a writepage IO is done, we need to:
2201  * clear the uptodate bits on error
2202  * clear the writeback bits in the extent tree for this IO
2203  * end_page_writeback if the page has no more pending IO
2204  *
2205  * Scheduling is not allowed, so the extent state tree is expected
2206  * to have one and only one object corresponding to this IO.
2207  */
2208 static void end_bio_extent_writepage(struct bio *bio, int err)
2209 {
2210         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2211         struct extent_io_tree *tree;
2212         u64 start;
2213         u64 end;
2214         int whole_page;
2215
2216         do {
2217                 struct page *page = bvec->bv_page;
2218                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2219
2220                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2221                          bvec->bv_offset;
2222                 end = start + bvec->bv_len - 1;
2223
2224                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2225                         whole_page = 1;
2226                 else
2227                         whole_page = 0;
2228
2229                 if (--bvec >= bio->bi_io_vec)
2230                         prefetchw(&bvec->bv_page->flags);
2231
2232                 if (end_extent_writepage(page, err, start, end))
2233                         continue;
2234
2235                 if (whole_page)
2236                         end_page_writeback(page);
2237                 else
2238                         check_page_writeback(tree, page);
2239         } while (bvec >= bio->bi_io_vec);
2240
2241         bio_put(bio);
2242 }
2243
2244 /*
2245  * after a readpage IO is done, we need to:
2246  * clear the uptodate bits on error
2247  * set the uptodate bits if things worked
2248  * set the page up to date if all extents in the tree are uptodate
2249  * clear the lock bit in the extent tree
2250  * unlock the page if there are no other extents locked for it
2251  *
2252  * Scheduling is not allowed, so the extent state tree is expected
2253  * to have one and only one object corresponding to this IO.
2254  */
2255 static void end_bio_extent_readpage(struct bio *bio, int err)
2256 {
2257         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2258         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2259         struct bio_vec *bvec = bio->bi_io_vec;
2260         struct extent_io_tree *tree;
2261         u64 start;
2262         u64 end;
2263         int whole_page;
2264         int ret;
2265
2266         if (err)
2267                 uptodate = 0;
2268
2269         do {
2270                 struct page *page = bvec->bv_page;
2271                 struct extent_state *cached = NULL;
2272                 struct extent_state *state;
2273
2274                 pr_debug("end_bio_extent_readpage: bi_vcnt=%d, idx=%d, err=%d, "
2275                          "mirror=%ld\n", bio->bi_vcnt, bio->bi_idx, err,
2276                          (long int)bio->bi_bdev);
2277                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2278
2279                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2280                         bvec->bv_offset;
2281                 end = start + bvec->bv_len - 1;
2282
2283                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2284                         whole_page = 1;
2285                 else
2286                         whole_page = 0;
2287
2288                 if (++bvec <= bvec_end)
2289                         prefetchw(&bvec->bv_page->flags);
2290
2291                 spin_lock(&tree->lock);
2292                 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
2293                 if (state && state->start == start) {
2294                         /*
2295                          * take a reference on the state, unlock will drop
2296                          * the ref
2297                          */
2298                         cache_state(state, &cached);
2299                 }
2300                 spin_unlock(&tree->lock);
2301
2302                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
2303                         ret = tree->ops->readpage_end_io_hook(page, start, end,
2304                                                               state);
2305                         if (ret)
2306                                 uptodate = 0;
2307                         else
2308                                 clean_io_failure(start, page);
2309                 }
2310                 if (!uptodate) {
2311                         int failed_mirror;
2312                         failed_mirror = (int)(unsigned long)bio->bi_bdev;
2313                         /*
2314                          * The generic bio_readpage_error handles errors the
2315                          * following way: If possible, new read requests are
2316                          * created and submitted and will end up in
2317                          * end_bio_extent_readpage as well (if we're lucky, not
2318                          * in the !uptodate case). In that case it returns 0 and
2319                          * we just go on with the next page in our bio. If it
2320                          * can't handle the error it will return -EIO and we
2321                          * remain responsible for that page.
2322                          */
2323                         ret = bio_readpage_error(bio, page, start, end,
2324                                                         failed_mirror, NULL);
2325                         if (ret == 0) {
2326 error_handled:
2327                                 uptodate =
2328                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2329                                 if (err)
2330                                         uptodate = 0;
2331                                 uncache_state(&cached);
2332                                 continue;
2333                         }
2334                         if (tree->ops && tree->ops->readpage_io_failed_hook) {
2335                                 ret = tree->ops->readpage_io_failed_hook(
2336                                                         bio, page, start, end,
2337                                                         failed_mirror, state);
2338                                 if (ret == 0)
2339                                         goto error_handled;
2340                         }
2341                 }
2342
2343                 if (uptodate && tree->track_uptodate) {
2344                         set_extent_uptodate(tree, start, end, &cached,
2345                                             GFP_ATOMIC);
2346                 }
2347                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2348
2349                 if (whole_page) {
2350                         if (uptodate) {
2351                                 SetPageUptodate(page);
2352                         } else {
2353                                 ClearPageUptodate(page);
2354                                 SetPageError(page);
2355                         }
2356                         unlock_page(page);
2357                 } else {
2358                         if (uptodate) {
2359                                 check_page_uptodate(tree, page);
2360                         } else {
2361                                 ClearPageUptodate(page);
2362                                 SetPageError(page);
2363                         }
2364                         check_page_locked(tree, page);
2365                 }
2366         } while (bvec <= bvec_end);
2367
2368         bio_put(bio);
2369 }
2370
2371 struct bio *
2372 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2373                 gfp_t gfp_flags)
2374 {
2375         struct bio *bio;
2376
2377         bio = bio_alloc(gfp_flags, nr_vecs);
2378
2379         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2380                 while (!bio && (nr_vecs /= 2))
2381                         bio = bio_alloc(gfp_flags, nr_vecs);
2382         }
2383
2384         if (bio) {
2385                 bio->bi_size = 0;
2386                 bio->bi_bdev = bdev;
2387                 bio->bi_sector = first_sector;
2388         }
2389         return bio;
2390 }
2391
2392 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
2393                           unsigned long bio_flags)
2394 {
2395         int ret = 0;
2396         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2397         struct page *page = bvec->bv_page;
2398         struct extent_io_tree *tree = bio->bi_private;
2399         u64 start;
2400
2401         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
2402
2403         bio->bi_private = NULL;
2404
2405         bio_get(bio);
2406
2407         if (tree->ops && tree->ops->submit_bio_hook)
2408                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2409                                            mirror_num, bio_flags, start);
2410         else
2411                 btrfsic_submit_bio(rw, bio);
2412
2413         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2414                 ret = -EOPNOTSUPP;
2415         bio_put(bio);
2416         return ret;
2417 }
2418
2419 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2420                               struct page *page, sector_t sector,
2421                               size_t size, unsigned long offset,
2422                               struct block_device *bdev,
2423                               struct bio **bio_ret,
2424                               unsigned long max_pages,
2425                               bio_end_io_t end_io_func,
2426                               int mirror_num,
2427                               unsigned long prev_bio_flags,
2428                               unsigned long bio_flags)
2429 {
2430         int ret = 0;
2431         struct bio *bio;
2432         int nr;
2433         int contig = 0;
2434         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2435         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2436         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2437
2438         if (bio_ret && *bio_ret) {
2439                 bio = *bio_ret;
2440                 if (old_compressed)
2441                         contig = bio->bi_sector == sector;
2442                 else
2443                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
2444                                 sector;
2445
2446                 if (prev_bio_flags != bio_flags || !contig ||
2447                     (tree->ops && tree->ops->merge_bio_hook &&
2448                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
2449                                                bio_flags)) ||
2450                     bio_add_page(bio, page, page_size, offset) < page_size) {
2451                         ret = submit_one_bio(rw, bio, mirror_num,
2452                                              prev_bio_flags);
2453                         bio = NULL;
2454                 } else {
2455                         return 0;
2456                 }
2457         }
2458         if (this_compressed)
2459                 nr = BIO_MAX_PAGES;
2460         else
2461                 nr = bio_get_nr_vecs(bdev);
2462
2463         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2464         if (!bio)
2465                 return -ENOMEM;
2466
2467         bio_add_page(bio, page, page_size, offset);
2468         bio->bi_end_io = end_io_func;
2469         bio->bi_private = tree;
2470
2471         if (bio_ret)
2472                 *bio_ret = bio;
2473         else
2474                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2475
2476         return ret;
2477 }
2478
2479 void attach_extent_buffer_page(struct extent_buffer *eb, struct page *page)
2480 {
2481         if (!PagePrivate(page)) {
2482                 SetPagePrivate(page);
2483                 page_cache_get(page);
2484                 set_page_private(page, (unsigned long)eb);
2485         } else {
2486                 WARN_ON(page->private != (unsigned long)eb);
2487         }
2488 }
2489
2490 void set_page_extent_mapped(struct page *page)
2491 {
2492         if (!PagePrivate(page)) {
2493                 SetPagePrivate(page);
2494                 page_cache_get(page);
2495                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2496         }
2497 }
2498
2499 /*
2500  * basic readpage implementation.  Locked extent state structs are inserted
2501  * into the tree that are removed when the IO is done (by the end_io
2502  * handlers)
2503  */
2504 static int __extent_read_full_page(struct extent_io_tree *tree,
2505                                    struct page *page,
2506                                    get_extent_t *get_extent,
2507                                    struct bio **bio, int mirror_num,
2508                                    unsigned long *bio_flags)
2509 {
2510         struct inode *inode = page->mapping->host;
2511         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2512         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2513         u64 end;
2514         u64 cur = start;
2515         u64 extent_offset;
2516         u64 last_byte = i_size_read(inode);
2517         u64 block_start;
2518         u64 cur_end;
2519         sector_t sector;
2520         struct extent_map *em;
2521         struct block_device *bdev;
2522         struct btrfs_ordered_extent *ordered;
2523         int ret;
2524         int nr = 0;
2525         size_t pg_offset = 0;
2526         size_t iosize;
2527         size_t disk_io_size;
2528         size_t blocksize = inode->i_sb->s_blocksize;
2529         unsigned long this_bio_flag = 0;
2530
2531         set_page_extent_mapped(page);
2532
2533         if (!PageUptodate(page)) {
2534                 if (cleancache_get_page(page) == 0) {
2535                         BUG_ON(blocksize != PAGE_SIZE);
2536                         goto out;
2537                 }
2538         }
2539
2540         end = page_end;
2541         while (1) {
2542                 lock_extent(tree, start, end, GFP_NOFS);
2543                 ordered = btrfs_lookup_ordered_extent(inode, start);
2544                 if (!ordered)
2545                         break;
2546                 unlock_extent(tree, start, end, GFP_NOFS);
2547                 btrfs_start_ordered_extent(inode, ordered, 1);
2548                 btrfs_put_ordered_extent(ordered);
2549         }
2550
2551         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2552                 char *userpage;
2553                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2554
2555                 if (zero_offset) {
2556                         iosize = PAGE_CACHE_SIZE - zero_offset;
2557                         userpage = kmap_atomic(page, KM_USER0);
2558                         memset(userpage + zero_offset, 0, iosize);
2559                         flush_dcache_page(page);
2560                         kunmap_atomic(userpage, KM_USER0);
2561                 }
2562         }
2563         while (cur <= end) {
2564                 if (cur >= last_byte) {
2565                         char *userpage;
2566                         struct extent_state *cached = NULL;
2567
2568                         iosize = PAGE_CACHE_SIZE - pg_offset;
2569                         userpage = kmap_atomic(page, KM_USER0);
2570                         memset(userpage + pg_offset, 0, iosize);
2571                         flush_dcache_page(page);
2572                         kunmap_atomic(userpage, KM_USER0);
2573                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2574                                             &cached, GFP_NOFS);
2575                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2576                                              &cached, GFP_NOFS);
2577                         break;
2578                 }
2579                 em = get_extent(inode, page, pg_offset, cur,
2580                                 end - cur + 1, 0);
2581                 if (IS_ERR_OR_NULL(em)) {
2582                         SetPageError(page);
2583                         unlock_extent(tree, cur, end, GFP_NOFS);
2584                         break;
2585                 }
2586                 extent_offset = cur - em->start;
2587                 BUG_ON(extent_map_end(em) <= cur);
2588                 BUG_ON(end < cur);
2589
2590                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2591                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2592                         extent_set_compress_type(&this_bio_flag,
2593                                                  em->compress_type);
2594                 }
2595
2596                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2597                 cur_end = min(extent_map_end(em) - 1, end);
2598                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2599                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2600                         disk_io_size = em->block_len;
2601                         sector = em->block_start >> 9;
2602                 } else {
2603                         sector = (em->block_start + extent_offset) >> 9;
2604                         disk_io_size = iosize;
2605                 }
2606                 bdev = em->bdev;
2607                 block_start = em->block_start;
2608                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2609                         block_start = EXTENT_MAP_HOLE;
2610                 free_extent_map(em);
2611                 em = NULL;
2612
2613                 /* we've found a hole, just zero and go on */
2614                 if (block_start == EXTENT_MAP_HOLE) {
2615                         char *userpage;
2616                         struct extent_state *cached = NULL;
2617
2618                         userpage = kmap_atomic(page, KM_USER0);
2619                         memset(userpage + pg_offset, 0, iosize);
2620                         flush_dcache_page(page);
2621                         kunmap_atomic(userpage, KM_USER0);
2622
2623                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2624                                             &cached, GFP_NOFS);
2625                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2626                                              &cached, GFP_NOFS);
2627                         cur = cur + iosize;
2628                         pg_offset += iosize;
2629                         continue;
2630                 }
2631                 /* the get_extent function already copied into the page */
2632                 if (test_range_bit(tree, cur, cur_end,
2633                                    EXTENT_UPTODATE, 1, NULL)) {
2634                         check_page_uptodate(tree, page);
2635                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2636                         cur = cur + iosize;
2637                         pg_offset += iosize;
2638                         continue;
2639                 }
2640                 /* we have an inline extent but it didn't get marked up
2641                  * to date.  Error out
2642                  */
2643                 if (block_start == EXTENT_MAP_INLINE) {
2644                         SetPageError(page);
2645                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2646                         cur = cur + iosize;
2647                         pg_offset += iosize;
2648                         continue;
2649                 }
2650
2651                 ret = 0;
2652                 if (tree->ops && tree->ops->readpage_io_hook) {
2653                         ret = tree->ops->readpage_io_hook(page, cur,
2654                                                           cur + iosize - 1);
2655                 }
2656                 if (!ret) {
2657                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2658                         pnr -= page->index;
2659                         ret = submit_extent_page(READ, tree, page,
2660                                          sector, disk_io_size, pg_offset,
2661                                          bdev, bio, pnr,
2662                                          end_bio_extent_readpage, mirror_num,
2663                                          *bio_flags,
2664                                          this_bio_flag);
2665                         nr++;
2666                         *bio_flags = this_bio_flag;
2667                 }
2668                 if (ret)
2669                         SetPageError(page);
2670                 cur = cur + iosize;
2671                 pg_offset += iosize;
2672         }
2673 out:
2674         if (!nr) {
2675                 if (!PageError(page))
2676                         SetPageUptodate(page);
2677                 unlock_page(page);
2678         }
2679         return 0;
2680 }
2681
2682 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2683                             get_extent_t *get_extent, int mirror_num)
2684 {
2685         struct bio *bio = NULL;
2686         unsigned long bio_flags = 0;
2687         int ret;
2688
2689         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
2690                                       &bio_flags);
2691         if (bio)
2692                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
2693         return ret;
2694 }
2695
2696 static noinline void update_nr_written(struct page *page,
2697                                       struct writeback_control *wbc,
2698                                       unsigned long nr_written)
2699 {
2700         wbc->nr_to_write -= nr_written;
2701         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2702             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2703                 page->mapping->writeback_index = page->index + nr_written;
2704 }
2705
2706 /*
2707  * the writepage semantics are similar to regular writepage.  extent
2708  * records are inserted to lock ranges in the tree, and as dirty areas
2709  * are found, they are marked writeback.  Then the lock bits are removed
2710  * and the end_io handler clears the writeback ranges
2711  */
2712 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2713                               void *data)
2714 {
2715         struct inode *inode = page->mapping->host;
2716         struct extent_page_data *epd = data;
2717         struct extent_io_tree *tree = epd->tree;
2718         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2719         u64 delalloc_start;
2720         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2721         u64 end;
2722         u64 cur = start;
2723         u64 extent_offset;
2724         u64 last_byte = i_size_read(inode);
2725         u64 block_start;
2726         u64 iosize;
2727         sector_t sector;
2728         struct extent_state *cached_state = NULL;
2729         struct extent_map *em;
2730         struct block_device *bdev;
2731         int ret;
2732         int nr = 0;
2733         size_t pg_offset = 0;
2734         size_t blocksize;
2735         loff_t i_size = i_size_read(inode);
2736         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2737         u64 nr_delalloc;
2738         u64 delalloc_end;
2739         int page_started;
2740         int compressed;
2741         int write_flags;
2742         unsigned long nr_written = 0;
2743         bool fill_delalloc = true;
2744
2745         if (wbc->sync_mode == WB_SYNC_ALL)
2746                 write_flags = WRITE_SYNC;
2747         else
2748                 write_flags = WRITE;
2749
2750         trace___extent_writepage(page, inode, wbc);
2751
2752         WARN_ON(!PageLocked(page));
2753
2754         ClearPageError(page);
2755
2756         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2757         if (page->index > end_index ||
2758            (page->index == end_index && !pg_offset)) {
2759                 page->mapping->a_ops->invalidatepage(page, 0);
2760                 unlock_page(page);
2761                 return 0;
2762         }
2763
2764         if (page->index == end_index) {
2765                 char *userpage;
2766
2767                 userpage = kmap_atomic(page, KM_USER0);
2768                 memset(userpage + pg_offset, 0,
2769                        PAGE_CACHE_SIZE - pg_offset);
2770                 kunmap_atomic(userpage, KM_USER0);
2771                 flush_dcache_page(page);
2772         }
2773         pg_offset = 0;
2774
2775         set_page_extent_mapped(page);
2776
2777         if (!tree->ops || !tree->ops->fill_delalloc)
2778                 fill_delalloc = false;
2779
2780         delalloc_start = start;
2781         delalloc_end = 0;
2782         page_started = 0;
2783         if (!epd->extent_locked && fill_delalloc) {
2784                 u64 delalloc_to_write = 0;
2785                 /*
2786                  * make sure the wbc mapping index is at least updated
2787                  * to this page.
2788                  */
2789                 update_nr_written(page, wbc, 0);
2790
2791                 while (delalloc_end < page_end) {
2792                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2793                                                        page,
2794                                                        &delalloc_start,
2795                                                        &delalloc_end,
2796                                                        128 * 1024 * 1024);
2797                         if (nr_delalloc == 0) {
2798                                 delalloc_start = delalloc_end + 1;
2799                                 continue;
2800                         }
2801                         ret = tree->ops->fill_delalloc(inode, page,
2802                                                        delalloc_start,
2803                                                        delalloc_end,
2804                                                        &page_started,
2805                                                        &nr_written);
2806                         BUG_ON(ret);
2807                         /*
2808                          * delalloc_end is already one less than the total
2809                          * length, so we don't subtract one from
2810                          * PAGE_CACHE_SIZE
2811                          */
2812                         delalloc_to_write += (delalloc_end - delalloc_start +
2813                                               PAGE_CACHE_SIZE) >>
2814                                               PAGE_CACHE_SHIFT;
2815                         delalloc_start = delalloc_end + 1;
2816                 }
2817                 if (wbc->nr_to_write < delalloc_to_write) {
2818                         int thresh = 8192;
2819
2820                         if (delalloc_to_write < thresh * 2)
2821                                 thresh = delalloc_to_write;
2822                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2823                                                  thresh);
2824                 }
2825
2826                 /* did the fill delalloc function already unlock and start
2827                  * the IO?
2828                  */
2829                 if (page_started) {
2830                         ret = 0;
2831                         /*
2832                          * we've unlocked the page, so we can't update
2833                          * the mapping's writeback index, just update
2834                          * nr_to_write.
2835                          */
2836                         wbc->nr_to_write -= nr_written;
2837                         goto done_unlocked;
2838                 }
2839         }
2840         if (tree->ops && tree->ops->writepage_start_hook) {
2841                 ret = tree->ops->writepage_start_hook(page, start,
2842                                                       page_end);
2843                 if (ret) {
2844                         /* Fixup worker will requeue */
2845                         if (ret == -EBUSY)
2846                                 wbc->pages_skipped++;
2847                         else
2848                                 redirty_page_for_writepage(wbc, page);
2849                         update_nr_written(page, wbc, nr_written);
2850                         unlock_page(page);
2851                         ret = 0;
2852                         goto done_unlocked;
2853                 }
2854         }
2855
2856         /*
2857          * we don't want to touch the inode after unlocking the page,
2858          * so we update the mapping writeback index now
2859          */
2860         update_nr_written(page, wbc, nr_written + 1);
2861
2862         end = page_end;
2863         if (last_byte <= start) {
2864                 if (tree->ops && tree->ops->writepage_end_io_hook)
2865                         tree->ops->writepage_end_io_hook(page, start,
2866                                                          page_end, NULL, 1);
2867                 goto done;
2868         }
2869
2870         blocksize = inode->i_sb->s_blocksize;
2871
2872         while (cur <= end) {
2873                 if (cur >= last_byte) {
2874                         if (tree->ops && tree->ops->writepage_end_io_hook)
2875                                 tree->ops->writepage_end_io_hook(page, cur,
2876                                                          page_end, NULL, 1);
2877                         break;
2878                 }
2879                 em = epd->get_extent(inode, page, pg_offset, cur,
2880                                      end - cur + 1, 1);
2881                 if (IS_ERR_OR_NULL(em)) {
2882                         SetPageError(page);
2883                         break;
2884                 }
2885
2886                 extent_offset = cur - em->start;
2887                 BUG_ON(extent_map_end(em) <= cur);
2888                 BUG_ON(end < cur);
2889                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2890                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2891                 sector = (em->block_start + extent_offset) >> 9;
2892                 bdev = em->bdev;
2893                 block_start = em->block_start;
2894                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2895                 free_extent_map(em);
2896                 em = NULL;
2897
2898                 /*
2899                  * compressed and inline extents are written through other
2900                  * paths in the FS
2901                  */
2902                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2903                     block_start == EXTENT_MAP_INLINE) {
2904                         /*
2905                          * end_io notification does not happen here for
2906                          * compressed extents
2907                          */
2908                         if (!compressed && tree->ops &&
2909                             tree->ops->writepage_end_io_hook)
2910                                 tree->ops->writepage_end_io_hook(page, cur,
2911                                                          cur + iosize - 1,
2912                                                          NULL, 1);
2913                         else if (compressed) {
2914                                 /* we don't want to end_page_writeback on
2915                                  * a compressed extent.  this happens
2916                                  * elsewhere
2917                                  */
2918                                 nr++;
2919                         }
2920
2921                         cur += iosize;
2922                         pg_offset += iosize;
2923                         continue;
2924                 }
2925                 /* leave this out until we have a page_mkwrite call */
2926                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2927                                    EXTENT_DIRTY, 0, NULL)) {
2928                         cur = cur + iosize;
2929                         pg_offset += iosize;
2930                         continue;
2931                 }
2932
2933                 if (tree->ops && tree->ops->writepage_io_hook) {
2934                         ret = tree->ops->writepage_io_hook(page, cur,
2935                                                 cur + iosize - 1);
2936                 } else {
2937                         ret = 0;
2938                 }
2939                 if (ret) {
2940                         SetPageError(page);
2941                 } else {
2942                         unsigned long max_nr = end_index + 1;
2943
2944                         set_range_writeback(tree, cur, cur + iosize - 1);
2945                         if (!PageWriteback(page)) {
2946                                 printk(KERN_ERR "btrfs warning page %lu not "
2947                                        "writeback, cur %llu end %llu\n",
2948                                        page->index, (unsigned long long)cur,
2949                                        (unsigned long long)end);
2950                         }
2951
2952                         ret = submit_extent_page(write_flags, tree, page,
2953                                                  sector, iosize, pg_offset,
2954                                                  bdev, &epd->bio, max_nr,
2955                                                  end_bio_extent_writepage,
2956                                                  0, 0, 0);
2957                         if (ret)
2958                                 SetPageError(page);
2959                 }
2960                 cur = cur + iosize;
2961                 pg_offset += iosize;
2962                 nr++;
2963         }
2964 done:
2965         if (nr == 0) {
2966                 /* make sure the mapping tag for page dirty gets cleared */
2967                 set_page_writeback(page);
2968                 end_page_writeback(page);
2969         }
2970         unlock_page(page);
2971
2972 done_unlocked:
2973
2974         /* drop our reference on any cached states */
2975         free_extent_state(cached_state);
2976         return 0;
2977 }
2978
2979 static int eb_wait(void *word)
2980 {
2981         io_schedule();
2982         return 0;
2983 }
2984
2985 static void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
2986 {
2987         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
2988                     TASK_UNINTERRUPTIBLE);
2989 }
2990
2991 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
2992                                      struct btrfs_fs_info *fs_info,
2993                                      struct extent_page_data *epd)
2994 {
2995         unsigned long i, num_pages;
2996         int flush = 0;
2997         int ret = 0;
2998
2999         if (!btrfs_try_tree_write_lock(eb)) {
3000                 flush = 1;
3001                 flush_write_bio(epd);
3002                 btrfs_tree_lock(eb);
3003         }
3004
3005         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3006                 btrfs_tree_unlock(eb);
3007                 if (!epd->sync_io)
3008                         return 0;
3009                 if (!flush) {
3010                         flush_write_bio(epd);
3011                         flush = 1;
3012                 }
3013                 wait_on_extent_buffer_writeback(eb);
3014                 btrfs_tree_lock(eb);
3015                 if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3016                         printk(KERN_ERR "Um, ok?\n");
3017                         btrfs_tree_unlock(eb);
3018                         return 0;
3019                 }
3020         }
3021
3022         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3023                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3024                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3025                 spin_lock(&fs_info->delalloc_lock);
3026                 if (fs_info->dirty_metadata_bytes >= eb->len)
3027                         fs_info->dirty_metadata_bytes -= eb->len;
3028                 else
3029                         WARN_ON(1);
3030                 spin_unlock(&fs_info->delalloc_lock);
3031                 ret = 1;
3032         }
3033
3034         btrfs_tree_unlock(eb);
3035
3036         if (!ret)
3037                 return ret;
3038
3039         num_pages = num_extent_pages(eb->start, eb->len);
3040         for (i = 0; i < num_pages; i++) {
3041                 struct page *p = extent_buffer_page(eb, i);
3042
3043                 if (!trylock_page(p)) {
3044                         if (!flush) {
3045                                 flush_write_bio(epd);
3046                                 flush = 1;
3047                         }
3048                         lock_page(p);
3049                 }
3050         }
3051
3052         return ret;
3053 }
3054
3055 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3056 {
3057         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3058         smp_mb__after_clear_bit();
3059         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3060 }
3061
3062 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3063 {
3064         int uptodate = err == 0;
3065         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
3066         struct extent_buffer *eb;
3067         int done;
3068
3069         do {
3070                 struct page *page = bvec->bv_page;
3071
3072                 bvec--;
3073                 eb = (struct extent_buffer *)page->private;
3074                 BUG_ON(!eb);
3075                 done = atomic_dec_and_test(&eb->io_pages);
3076
3077                 if (!uptodate || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3078                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3079                         ClearPageUptodate(page);
3080                         SetPageError(page);
3081                 }
3082
3083                 end_page_writeback(page);
3084
3085                 if (!done)
3086                         continue;
3087
3088                 end_extent_buffer_writeback(eb);
3089         } while (bvec >= bio->bi_io_vec);
3090
3091         bio_put(bio);
3092
3093 }
3094
3095 static int write_one_eb(struct extent_buffer *eb,
3096                         struct btrfs_fs_info *fs_info,
3097                         struct writeback_control *wbc,
3098                         struct extent_page_data *epd)
3099 {
3100         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3101         u64 offset = eb->start;
3102         unsigned long i, num_pages;
3103         int rw = (epd->sync_io ? WRITE_SYNC : WRITE);
3104         int ret;
3105
3106         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3107         num_pages = num_extent_pages(eb->start, eb->len);
3108         atomic_set(&eb->io_pages, num_pages);
3109         for (i = 0; i < num_pages; i++) {
3110                 struct page *p = extent_buffer_page(eb, i);
3111
3112                 clear_page_dirty_for_io(p);
3113                 set_page_writeback(p);
3114                 ret = submit_extent_page(rw, eb->tree, p, offset >> 9,
3115                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3116                                          -1, end_bio_extent_buffer_writepage,
3117                                          0, 0, 0);
3118                 if (ret) {
3119                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3120                         SetPageError(p);
3121                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3122                                 end_extent_buffer_writeback(eb);
3123                         ret = -EIO;
3124                         break;
3125                 }
3126                 offset += PAGE_CACHE_SIZE;
3127                 update_nr_written(p, wbc, 1);
3128                 unlock_page(p);
3129         }
3130
3131         if (unlikely(ret)) {
3132                 for (; i < num_pages; i++) {
3133                         struct page *p = extent_buffer_page(eb, i);
3134                         unlock_page(p);
3135                 }
3136         }
3137
3138         return ret;
3139 }
3140
3141 int btree_write_cache_pages(struct address_space *mapping,
3142                                    struct writeback_control *wbc)
3143 {
3144         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3145         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3146         struct extent_buffer *eb, *prev_eb = NULL;
3147         struct extent_page_data epd = {
3148                 .bio = NULL,
3149                 .tree = tree,
3150                 .extent_locked = 0,
3151                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3152         };
3153         int ret = 0;
3154         int done = 0;
3155         int nr_to_write_done = 0;
3156         struct pagevec pvec;
3157         int nr_pages;
3158         pgoff_t index;
3159         pgoff_t end;            /* Inclusive */
3160         int scanned = 0;
3161         int tag;
3162
3163         pagevec_init(&pvec, 0);
3164         if (wbc->range_cyclic) {
3165                 index = mapping->writeback_index; /* Start from prev offset */
3166                 end = -1;
3167         } else {
3168                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3169                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3170                 scanned = 1;
3171         }
3172         if (wbc->sync_mode == WB_SYNC_ALL)
3173                 tag = PAGECACHE_TAG_TOWRITE;
3174         else
3175                 tag = PAGECACHE_TAG_DIRTY;
3176 retry:
3177         if (wbc->sync_mode == WB_SYNC_ALL)
3178                 tag_pages_for_writeback(mapping, index, end);
3179         while (!done && !nr_to_write_done && (index <= end) &&
3180                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3181                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3182                 unsigned i;
3183
3184                 scanned = 1;
3185                 for (i = 0; i < nr_pages; i++) {
3186                         struct page *page = pvec.pages[i];
3187
3188                         if (!PagePrivate(page))
3189                                 continue;
3190
3191                         if (!wbc->range_cyclic && page->index > end) {
3192                                 done = 1;
3193                                 break;
3194                         }
3195
3196                         eb = (struct extent_buffer *)page->private;
3197                         if (!eb) {
3198                                 WARN_ON(1);
3199                                 continue;
3200                         }
3201
3202                         if (eb == prev_eb)
3203                                 continue;
3204
3205                         if (!atomic_inc_not_zero(&eb->refs)) {
3206                                 WARN_ON(1);
3207                                 continue;
3208                         }
3209
3210                         prev_eb = eb;
3211                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3212                         if (!ret) {
3213                                 free_extent_buffer(eb);
3214                                 continue;
3215                         }
3216
3217                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3218                         if (ret) {
3219                                 done = 1;
3220                                 free_extent_buffer(eb);
3221                                 break;
3222                         }
3223                         free_extent_buffer(eb);
3224
3225                         /*
3226                          * the filesystem may choose to bump up nr_to_write.
3227                          * We have to make sure to honor the new nr_to_write
3228                          * at any time
3229                          */
3230                         nr_to_write_done = wbc->nr_to_write <= 0;
3231                 }
3232                 pagevec_release(&pvec);
3233                 cond_resched();
3234         }
3235         if (!scanned && !done) {
3236                 /*
3237                  * We hit the last page and there is more work to be done: wrap
3238                  * back to the start of the file
3239                  */
3240                 scanned = 1;
3241                 index = 0;
3242                 goto retry;
3243         }
3244         flush_write_bio(&epd);
3245         return ret;
3246 }
3247
3248 /**
3249  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3250  * @mapping: address space structure to write
3251  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3252  * @writepage: function called for each page
3253  * @data: data passed to writepage function
3254  *
3255  * If a page is already under I/O, write_cache_pages() skips it, even
3256  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3257  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3258  * and msync() need to guarantee that all the data which was dirty at the time
3259  * the call was made get new I/O started against them.  If wbc->sync_mode is
3260  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3261  * existing IO to complete.
3262  */
3263 static int extent_write_cache_pages(struct extent_io_tree *tree,
3264                              struct address_space *mapping,
3265                              struct writeback_control *wbc,
3266                              writepage_t writepage, void *data,
3267                              void (*flush_fn)(void *))
3268 {
3269         int ret = 0;
3270         int done = 0;
3271         int nr_to_write_done = 0;
3272         struct pagevec pvec;
3273         int nr_pages;
3274         pgoff_t index;
3275         pgoff_t end;            /* Inclusive */
3276         int scanned = 0;
3277         int tag;
3278
3279         pagevec_init(&pvec, 0);
3280         if (wbc->range_cyclic) {
3281                 index = mapping->writeback_index; /* Start from prev offset */
3282                 end = -1;
3283         } else {
3284                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3285                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3286                 scanned = 1;
3287         }
3288         if (wbc->sync_mode == WB_SYNC_ALL)
3289                 tag = PAGECACHE_TAG_TOWRITE;
3290         else
3291                 tag = PAGECACHE_TAG_DIRTY;
3292 retry:
3293         if (wbc->sync_mode == WB_SYNC_ALL)
3294                 tag_pages_for_writeback(mapping, index, end);
3295         while (!done && !nr_to_write_done && (index <= end) &&
3296                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3297                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3298                 unsigned i;
3299
3300                 scanned = 1;
3301                 for (i = 0; i < nr_pages; i++) {
3302                         struct page *page = pvec.pages[i];
3303
3304                         /*
3305                          * At this point we hold neither mapping->tree_lock nor
3306                          * lock on the page itself: the page may be truncated or
3307                          * invalidated (changing page->mapping to NULL), or even
3308                          * swizzled back from swapper_space to tmpfs file
3309                          * mapping
3310                          */
3311                         if (tree->ops &&
3312                             tree->ops->write_cache_pages_lock_hook) {
3313                                 tree->ops->write_cache_pages_lock_hook(page,
3314                                                                data, flush_fn);
3315                         } else {
3316                                 if (!trylock_page(page)) {
3317                                         flush_fn(data);
3318                                         lock_page(page);
3319                                 }
3320                         }
3321
3322                         if (unlikely(page->mapping != mapping)) {
3323                                 unlock_page(page);
3324                                 continue;
3325                         }
3326
3327                         if (!wbc->range_cyclic && page->index > end) {
3328                                 done = 1;
3329                                 unlock_page(page);
3330                                 continue;
3331                         }
3332
3333                         if (wbc->sync_mode != WB_SYNC_NONE) {
3334                                 if (PageWriteback(page))
3335                                         flush_fn(data);
3336                                 wait_on_page_writeback(page);
3337                         }
3338
3339                         if (PageWriteback(page) ||
3340                             !clear_page_dirty_for_io(page)) {
3341                                 unlock_page(page);
3342                                 continue;
3343                         }
3344
3345                         ret = (*writepage)(page, wbc, data);
3346
3347                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3348                                 unlock_page(page);
3349                                 ret = 0;
3350                         }
3351                         if (ret)
3352                                 done = 1;
3353
3354                         /*
3355                          * the filesystem may choose to bump up nr_to_write.
3356                          * We have to make sure to honor the new nr_to_write
3357                          * at any time
3358                          */
3359                         nr_to_write_done = wbc->nr_to_write <= 0;
3360                 }
3361                 pagevec_release(&pvec);
3362                 cond_resched();
3363         }
3364         if (!scanned && !done) {
3365                 /*
3366                  * We hit the last page and there is more work to be done: wrap
3367                  * back to the start of the file
3368                  */
3369                 scanned = 1;
3370                 index = 0;
3371                 goto retry;
3372         }
3373         return ret;
3374 }
3375
3376 static void flush_epd_write_bio(struct extent_page_data *epd)
3377 {
3378         if (epd->bio) {
3379                 if (epd->sync_io)
3380                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
3381                 else
3382                         submit_one_bio(WRITE, epd->bio, 0, 0);
3383                 epd->bio = NULL;
3384         }
3385 }
3386
3387 static noinline void flush_write_bio(void *data)
3388 {
3389         struct extent_page_data *epd = data;
3390         flush_epd_write_bio(epd);
3391 }
3392
3393 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3394                           get_extent_t *get_extent,
3395                           struct writeback_control *wbc)
3396 {
3397         int ret;
3398         struct extent_page_data epd = {
3399                 .bio = NULL,
3400                 .tree = tree,
3401                 .get_extent = get_extent,
3402                 .extent_locked = 0,
3403                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3404         };
3405
3406         ret = __extent_writepage(page, wbc, &epd);
3407
3408         flush_epd_write_bio(&epd);
3409         return ret;
3410 }
3411
3412 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3413                               u64 start, u64 end, get_extent_t *get_extent,
3414                               int mode)
3415 {
3416         int ret = 0;
3417         struct address_space *mapping = inode->i_mapping;
3418         struct page *page;
3419         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3420                 PAGE_CACHE_SHIFT;
3421
3422         struct extent_page_data epd = {
3423                 .bio = NULL,
3424                 .tree = tree,
3425                 .get_extent = get_extent,
3426                 .extent_locked = 1,
3427                 .sync_io = mode == WB_SYNC_ALL,
3428         };
3429         struct writeback_control wbc_writepages = {
3430                 .sync_mode      = mode,
3431                 .nr_to_write    = nr_pages * 2,
3432                 .range_start    = start,
3433                 .range_end      = end + 1,
3434         };
3435
3436         while (start <= end) {
3437                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3438                 if (clear_page_dirty_for_io(page))
3439                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3440                 else {
3441                         if (tree->ops && tree->ops->writepage_end_io_hook)
3442                                 tree->ops->writepage_end_io_hook(page, start,
3443                                                  start + PAGE_CACHE_SIZE - 1,
3444                                                  NULL, 1);
3445                         unlock_page(page);
3446                 }
3447                 page_cache_release(page);
3448                 start += PAGE_CACHE_SIZE;
3449         }
3450
3451         flush_epd_write_bio(&epd);
3452         return ret;
3453 }
3454
3455 int extent_writepages(struct extent_io_tree *tree,
3456                       struct address_space *mapping,
3457                       get_extent_t *get_extent,
3458                       struct writeback_control *wbc)
3459 {
3460         int ret = 0;
3461         struct extent_page_data epd = {
3462                 .bio = NULL,
3463                 .tree = tree,
3464                 .get_extent = get_extent,
3465                 .extent_locked = 0,
3466                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3467         };
3468
3469         ret = extent_write_cache_pages(tree, mapping, wbc,
3470                                        __extent_writepage, &epd,
3471                                        flush_write_bio);
3472         flush_epd_write_bio(&epd);
3473         return ret;
3474 }
3475
3476 int extent_readpages(struct extent_io_tree *tree,
3477                      struct address_space *mapping,
3478                      struct list_head *pages, unsigned nr_pages,
3479                      get_extent_t get_extent)
3480 {
3481         struct bio *bio = NULL;
3482         unsigned page_idx;
3483         unsigned long bio_flags = 0;
3484
3485         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3486                 struct page *page = list_entry(pages->prev, struct page, lru);
3487
3488                 prefetchw(&page->flags);
3489                 list_del(&page->lru);
3490                 if (!add_to_page_cache_lru(page, mapping,
3491                                         page->index, GFP_NOFS)) {
3492                         __extent_read_full_page(tree, page, get_extent,
3493                                                 &bio, 0, &bio_flags);
3494                 }
3495                 page_cache_release(page);
3496         }
3497         BUG_ON(!list_empty(pages));
3498         if (bio)
3499                 submit_one_bio(READ, bio, 0, bio_flags);
3500         return 0;
3501 }
3502
3503 /*
3504  * basic invalidatepage code, this waits on any locked or writeback
3505  * ranges corresponding to the page, and then deletes any extent state
3506  * records from the tree
3507  */
3508 int extent_invalidatepage(struct extent_io_tree *tree,
3509                           struct page *page, unsigned long offset)
3510 {
3511         struct extent_state *cached_state = NULL;
3512         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
3513         u64 end = start + PAGE_CACHE_SIZE - 1;
3514         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3515
3516         start += (offset + blocksize - 1) & ~(blocksize - 1);
3517         if (start > end)
3518                 return 0;
3519
3520         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
3521         wait_on_page_writeback(page);
3522         clear_extent_bit(tree, start, end,
3523                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3524                          EXTENT_DO_ACCOUNTING,
3525                          1, 1, &cached_state, GFP_NOFS);
3526         return 0;
3527 }
3528
3529 /*
3530  * a helper for releasepage, this tests for areas of the page that
3531  * are locked or under IO and drops the related state bits if it is safe
3532  * to drop the page.
3533  */
3534 int try_release_extent_state(struct extent_map_tree *map,
3535                              struct extent_io_tree *tree, struct page *page,
3536                              gfp_t mask)
3537 {
3538         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3539         u64 end = start + PAGE_CACHE_SIZE - 1;
3540         int ret = 1;
3541
3542         if (test_range_bit(tree, start, end,
3543                            EXTENT_IOBITS, 0, NULL))
3544                 ret = 0;
3545         else {
3546                 if ((mask & GFP_NOFS) == GFP_NOFS)
3547                         mask = GFP_NOFS;
3548                 /*
3549                  * at this point we can safely clear everything except the
3550                  * locked bit and the nodatasum bit
3551                  */
3552                 ret = clear_extent_bit(tree, start, end,
3553                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3554                                  0, 0, NULL, mask);
3555
3556                 /* if clear_extent_bit failed for enomem reasons,
3557                  * we can't allow the release to continue.
3558                  */
3559                 if (ret < 0)
3560                         ret = 0;
3561                 else
3562                         ret = 1;
3563         }
3564         return ret;
3565 }
3566
3567 /*
3568  * a helper for releasepage.  As long as there are no locked extents
3569  * in the range corresponding to the page, both state records and extent
3570  * map records are removed
3571  */
3572 int try_release_extent_mapping(struct extent_map_tree *map,
3573                                struct extent_io_tree *tree, struct page *page,
3574                                gfp_t mask)
3575 {
3576         struct extent_map *em;
3577         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3578         u64 end = start + PAGE_CACHE_SIZE - 1;
3579
3580         if ((mask & __GFP_WAIT) &&
3581             page->mapping->host->i_size > 16 * 1024 * 1024) {
3582                 u64 len;
3583                 while (start <= end) {
3584                         len = end - start + 1;
3585                         write_lock(&map->lock);
3586                         em = lookup_extent_mapping(map, start, len);
3587                         if (!em) {
3588                                 write_unlock(&map->lock);
3589                                 break;
3590                         }
3591                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3592                             em->start != start) {
3593                                 write_unlock(&map->lock);
3594                                 free_extent_map(em);
3595                                 break;
3596                         }
3597                         if (!test_range_bit(tree, em->start,
3598                                             extent_map_end(em) - 1,
3599                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
3600                                             0, NULL)) {
3601                                 remove_extent_mapping(map, em);
3602                                 /* once for the rb tree */
3603                                 free_extent_map(em);
3604                         }
3605                         start = extent_map_end(em);
3606                         write_unlock(&map->lock);
3607
3608                         /* once for us */
3609                         free_extent_map(em);
3610                 }
3611         }
3612         return try_release_extent_state(map, tree, page, mask);
3613 }
3614
3615 /*
3616  * helper function for fiemap, which doesn't want to see any holes.
3617  * This maps until we find something past 'last'
3618  */
3619 static struct extent_map *get_extent_skip_holes(struct inode *inode,
3620                                                 u64 offset,
3621                                                 u64 last,
3622                                                 get_extent_t *get_extent)
3623 {
3624         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
3625         struct extent_map *em;
3626         u64 len;
3627
3628         if (offset >= last)
3629                 return NULL;
3630
3631         while(1) {
3632                 len = last - offset;
3633                 if (len == 0)
3634                         break;
3635                 len = (len + sectorsize - 1) & ~(sectorsize - 1);
3636                 em = get_extent(inode, NULL, 0, offset, len, 0);
3637                 if (IS_ERR_OR_NULL(em))
3638                         return em;
3639
3640                 /* if this isn't a hole return it */
3641                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
3642                     em->block_start != EXTENT_MAP_HOLE) {
3643                         return em;
3644                 }
3645
3646                 /* this is a hole, advance to the next extent */
3647                 offset = extent_map_end(em);
3648                 free_extent_map(em);
3649                 if (offset >= last)
3650                         break;
3651         }
3652         return NULL;
3653 }
3654
3655 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
3656                 __u64 start, __u64 len, get_extent_t *get_extent)
3657 {
3658         int ret = 0;
3659         u64 off = start;
3660         u64 max = start + len;
3661         u32 flags = 0;
3662         u32 found_type;
3663         u64 last;
3664         u64 last_for_get_extent = 0;
3665         u64 disko = 0;
3666         u64 isize = i_size_read(inode);
3667         struct btrfs_key found_key;
3668         struct extent_map *em = NULL;
3669         struct extent_state *cached_state = NULL;
3670         struct btrfs_path *path;
3671         struct btrfs_file_extent_item *item;
3672         int end = 0;
3673         u64 em_start = 0;
3674         u64 em_len = 0;
3675         u64 em_end = 0;
3676         unsigned long emflags;
3677
3678         if (len == 0)
3679                 return -EINVAL;
3680
3681         path = btrfs_alloc_path();
3682         if (!path)
3683                 return -ENOMEM;
3684         path->leave_spinning = 1;
3685
3686         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
3687         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
3688
3689         /*
3690          * lookup the last file extent.  We're not using i_size here
3691          * because there might be preallocation past i_size
3692          */
3693         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3694                                        path, btrfs_ino(inode), -1, 0);
3695         if (ret < 0) {
3696                 btrfs_free_path(path);
3697                 return ret;
3698         }
3699         WARN_ON(!ret);
3700         path->slots[0]--;
3701         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3702                               struct btrfs_file_extent_item);
3703         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3704         found_type = btrfs_key_type(&found_key);
3705
3706         /* No extents, but there might be delalloc bits */
3707         if (found_key.objectid != btrfs_ino(inode) ||
3708             found_type != BTRFS_EXTENT_DATA_KEY) {
3709                 /* have to trust i_size as the end */
3710                 last = (u64)-1;
3711                 last_for_get_extent = isize;
3712         } else {
3713                 /*
3714                  * remember the start of the last extent.  There are a
3715                  * bunch of different factors that go into the length of the
3716                  * extent, so its much less complex to remember where it started
3717                  */
3718                 last = found_key.offset;
3719                 last_for_get_extent = last + 1;
3720         }
3721         btrfs_free_path(path);
3722
3723         /*
3724          * we might have some extents allocated but more delalloc past those
3725          * extents.  so, we trust isize unless the start of the last extent is
3726          * beyond isize
3727          */
3728         if (last < isize) {
3729                 last = (u64)-1;
3730                 last_for_get_extent = isize;
3731         }
3732
3733         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3734                          &cached_state, GFP_NOFS);
3735
3736         em = get_extent_skip_holes(inode, start, last_for_get_extent,
3737                                    get_extent);
3738         if (!em)
3739                 goto out;
3740         if (IS_ERR(em)) {
3741                 ret = PTR_ERR(em);
3742                 goto out;
3743         }
3744
3745         while (!end) {
3746                 u64 offset_in_extent;
3747
3748                 /* break if the extent we found is outside the range */
3749                 if (em->start >= max || extent_map_end(em) < off)
3750                         break;
3751
3752                 /*
3753                  * get_extent may return an extent that starts before our
3754                  * requested range.  We have to make sure the ranges
3755                  * we return to fiemap always move forward and don't
3756                  * overlap, so adjust the offsets here
3757                  */
3758                 em_start = max(em->start, off);
3759
3760                 /*
3761                  * record the offset from the start of the extent
3762                  * for adjusting the disk offset below
3763                  */
3764                 offset_in_extent = em_start - em->start;
3765                 em_end = extent_map_end(em);
3766                 em_len = em_end - em_start;
3767                 emflags = em->flags;
3768                 disko = 0;
3769                 flags = 0;
3770
3771                 /*
3772                  * bump off for our next call to get_extent
3773                  */
3774                 off = extent_map_end(em);
3775                 if (off >= max)
3776                         end = 1;
3777
3778                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3779                         end = 1;
3780                         flags |= FIEMAP_EXTENT_LAST;
3781                 } else if (em->block_start == EXTENT_MAP_INLINE) {
3782                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
3783                                   FIEMAP_EXTENT_NOT_ALIGNED);
3784                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3785                         flags |= (FIEMAP_EXTENT_DELALLOC |
3786                                   FIEMAP_EXTENT_UNKNOWN);
3787                 } else {
3788                         disko = em->block_start + offset_in_extent;
3789                 }
3790                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3791                         flags |= FIEMAP_EXTENT_ENCODED;
3792
3793                 free_extent_map(em);
3794                 em = NULL;
3795                 if ((em_start >= last) || em_len == (u64)-1 ||
3796                    (last == (u64)-1 && isize <= em_end)) {
3797                         flags |= FIEMAP_EXTENT_LAST;
3798                         end = 1;
3799                 }
3800
3801                 /* now scan forward to see if this is really the last extent. */
3802                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
3803                                            get_extent);
3804                 if (IS_ERR(em)) {
3805                         ret = PTR_ERR(em);
3806                         goto out;
3807                 }
3808                 if (!em) {
3809                         flags |= FIEMAP_EXTENT_LAST;
3810                         end = 1;
3811                 }
3812                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3813                                               em_len, flags);
3814                 if (ret)
3815                         goto out_free;
3816         }
3817 out_free:
3818         free_extent_map(em);
3819 out:
3820         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3821                              &cached_state, GFP_NOFS);
3822         return ret;
3823 }
3824
3825 inline struct page *extent_buffer_page(struct extent_buffer *eb,
3826                                               unsigned long i)
3827 {
3828         return eb->pages[i];
3829 }
3830
3831 inline unsigned long num_extent_pages(u64 start, u64 len)
3832 {
3833         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3834                 (start >> PAGE_CACHE_SHIFT);
3835 }
3836
3837 static void __free_extent_buffer(struct extent_buffer *eb)
3838 {
3839 #if LEAK_DEBUG
3840         unsigned long flags;
3841         spin_lock_irqsave(&leak_lock, flags);
3842         list_del(&eb->leak_list);
3843         spin_unlock_irqrestore(&leak_lock, flags);
3844 #endif
3845         if (eb->pages && eb->pages != eb->inline_pages)
3846                 kfree(eb->pages);
3847         kmem_cache_free(extent_buffer_cache, eb);
3848 }
3849
3850 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3851                                                    u64 start,
3852                                                    unsigned long len,
3853                                                    gfp_t mask)
3854 {
3855         struct extent_buffer *eb = NULL;
3856 #if LEAK_DEBUG
3857         unsigned long flags;
3858 #endif
3859
3860         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3861         if (eb == NULL)
3862                 return NULL;
3863         eb->start = start;
3864         eb->len = len;
3865         eb->tree = tree;
3866         rwlock_init(&eb->lock);
3867         atomic_set(&eb->write_locks, 0);
3868         atomic_set(&eb->read_locks, 0);
3869         atomic_set(&eb->blocking_readers, 0);
3870         atomic_set(&eb->blocking_writers, 0);
3871         atomic_set(&eb->spinning_readers, 0);
3872         atomic_set(&eb->spinning_writers, 0);
3873         eb->lock_nested = 0;
3874         init_waitqueue_head(&eb->write_lock_wq);
3875         init_waitqueue_head(&eb->read_lock_wq);
3876
3877 #if LEAK_DEBUG
3878         spin_lock_irqsave(&leak_lock, flags);
3879         list_add(&eb->leak_list, &buffers);
3880         spin_unlock_irqrestore(&leak_lock, flags);
3881 #endif
3882         spin_lock_init(&eb->refs_lock);
3883         atomic_set(&eb->refs, 1);
3884         atomic_set(&eb->io_pages, 0);
3885
3886         if (len > MAX_INLINE_EXTENT_BUFFER_SIZE) {
3887                 struct page **pages;
3888                 int num_pages = (len + PAGE_CACHE_SIZE - 1) >>
3889                         PAGE_CACHE_SHIFT;
3890                 pages = kzalloc(num_pages, mask);
3891                 if (!pages) {
3892                         __free_extent_buffer(eb);
3893                         return NULL;
3894                 }
3895                 eb->pages = pages;
3896         } else {
3897                 eb->pages = eb->inline_pages;
3898         }
3899
3900         return eb;
3901 }
3902
3903 static int extent_buffer_under_io(struct extent_buffer *eb)
3904 {
3905         return (atomic_read(&eb->io_pages) ||
3906                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
3907                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3908 }
3909
3910 /*
3911  * Helper for releasing extent buffer page.
3912  */
3913 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3914                                                 unsigned long start_idx)
3915 {
3916         unsigned long index;
3917         struct page *page;
3918
3919         BUG_ON(extent_buffer_under_io(eb));
3920
3921         index = num_extent_pages(eb->start, eb->len);
3922         if (start_idx >= index)
3923                 return;
3924
3925         do {
3926                 index--;
3927                 page = extent_buffer_page(eb, index);
3928                 if (page) {
3929                         spin_lock(&page->mapping->private_lock);
3930                         /*
3931                          * We do this since we'll remove the pages after we've
3932                          * removed the eb from the radix tree, so we could race
3933                          * and have this page now attached to the new eb.  So
3934                          * only clear page_private if it's still connected to
3935                          * this eb.
3936                          */
3937                         if (PagePrivate(page) &&
3938                             page->private == (unsigned long)eb) {
3939                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3940                                 BUG_ON(PageDirty(page));
3941                                 BUG_ON(PageWriteback(page));
3942                                 /*
3943                                  * We need to make sure we haven't be attached
3944                                  * to a new eb.
3945                                  */
3946                                 ClearPagePrivate(page);
3947                                 set_page_private(page, 0);
3948                                 /* One for the page private */
3949                                 page_cache_release(page);
3950                         }
3951                         spin_unlock(&page->mapping->private_lock);
3952
3953                         /* One for when we alloced the page */
3954                         page_cache_release(page);
3955                 }
3956         } while (index != start_idx);
3957 }
3958
3959 /*
3960  * Helper for releasing the extent buffer.
3961  */
3962 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3963 {
3964         btrfs_release_extent_buffer_page(eb, 0);
3965         __free_extent_buffer(eb);
3966 }
3967
3968 static void check_buffer_tree_ref(struct extent_buffer *eb)
3969 {
3970         /* the ref bit is tricky.  We have to make sure it is set
3971          * if we have the buffer dirty.   Otherwise the
3972          * code to free a buffer can end up dropping a dirty
3973          * page
3974          *
3975          * Once the ref bit is set, it won't go away while the
3976          * buffer is dirty or in writeback, and it also won't
3977          * go away while we have the reference count on the
3978          * eb bumped.
3979          *
3980          * We can't just set the ref bit without bumping the
3981          * ref on the eb because free_extent_buffer might
3982          * see the ref bit and try to clear it.  If this happens
3983          * free_extent_buffer might end up dropping our original
3984          * ref by mistake and freeing the page before we are able
3985          * to add one more ref.
3986          *
3987          * So bump the ref count first, then set the bit.  If someone
3988          * beat us to it, drop the ref we added.
3989          */
3990         if (!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
3991                 atomic_inc(&eb->refs);
3992                 if (test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3993                         atomic_dec(&eb->refs);
3994         }
3995 }
3996
3997 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
3998 {
3999         unsigned long num_pages, i;
4000
4001         check_buffer_tree_ref(eb);
4002
4003         num_pages = num_extent_pages(eb->start, eb->len);
4004         for (i = 0; i < num_pages; i++) {
4005                 struct page *p = extent_buffer_page(eb, i);
4006                 mark_page_accessed(p);
4007         }
4008 }
4009
4010 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
4011                                           u64 start, unsigned long len)
4012 {
4013         unsigned long num_pages = num_extent_pages(start, len);
4014         unsigned long i;
4015         unsigned long index = start >> PAGE_CACHE_SHIFT;
4016         struct extent_buffer *eb;
4017         struct extent_buffer *exists = NULL;
4018         struct page *p;
4019         struct address_space *mapping = tree->mapping;
4020         int uptodate = 1;
4021         int ret;
4022
4023         rcu_read_lock();
4024         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4025         if (eb && atomic_inc_not_zero(&eb->refs)) {
4026                 rcu_read_unlock();
4027                 mark_extent_buffer_accessed(eb);
4028                 return eb;
4029         }
4030         rcu_read_unlock();
4031
4032         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
4033         if (!eb)
4034                 return NULL;
4035
4036         for (i = 0; i < num_pages; i++, index++) {
4037                 p = find_or_create_page(mapping, index, GFP_NOFS);
4038                 if (!p) {
4039                         WARN_ON(1);
4040                         goto free_eb;
4041                 }
4042
4043                 spin_lock(&mapping->private_lock);
4044                 if (PagePrivate(p)) {
4045                         /*
4046                          * We could have already allocated an eb for this page
4047                          * and attached one so lets see if we can get a ref on
4048                          * the existing eb, and if we can we know it's good and
4049                          * we can just return that one, else we know we can just
4050                          * overwrite page->private.
4051                          */
4052                         exists = (struct extent_buffer *)p->private;
4053                         if (atomic_inc_not_zero(&exists->refs)) {
4054                                 spin_unlock(&mapping->private_lock);
4055                                 unlock_page(p);
4056                                 mark_extent_buffer_accessed(exists);
4057                                 goto free_eb;
4058                         }
4059
4060                         /*
4061                          * Do this so attach doesn't complain and we need to
4062                          * drop the ref the old guy had.
4063                          */
4064                         ClearPagePrivate(p);
4065                         WARN_ON(PageDirty(p));
4066                         page_cache_release(p);
4067                 }
4068                 attach_extent_buffer_page(eb, p);
4069                 spin_unlock(&mapping->private_lock);
4070                 WARN_ON(PageDirty(p));
4071                 mark_page_accessed(p);
4072                 eb->pages[i] = p;
4073                 if (!PageUptodate(p))
4074                         uptodate = 0;
4075
4076                 /*
4077                  * see below about how we avoid a nasty race with release page
4078                  * and why we unlock later
4079                  */
4080         }
4081         if (uptodate)
4082                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4083 again:
4084         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4085         if (ret)
4086                 goto free_eb;
4087
4088         spin_lock(&tree->buffer_lock);
4089         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
4090         if (ret == -EEXIST) {
4091                 exists = radix_tree_lookup(&tree->buffer,
4092                                                 start >> PAGE_CACHE_SHIFT);
4093                 if (!atomic_inc_not_zero(&exists->refs)) {
4094                         spin_unlock(&tree->buffer_lock);
4095                         radix_tree_preload_end();
4096                         exists = NULL;
4097                         goto again;
4098                 }
4099                 spin_unlock(&tree->buffer_lock);
4100                 radix_tree_preload_end();
4101                 mark_extent_buffer_accessed(exists);
4102                 goto free_eb;
4103         }
4104         /* add one reference for the tree */
4105         spin_lock(&eb->refs_lock);
4106         check_buffer_tree_ref(eb);
4107         spin_unlock(&eb->refs_lock);
4108         spin_unlock(&tree->buffer_lock);
4109         radix_tree_preload_end();
4110
4111         /*
4112          * there is a race where release page may have
4113          * tried to find this extent buffer in the radix
4114          * but failed.  It will tell the VM it is safe to
4115          * reclaim the, and it will clear the page private bit.
4116          * We must make sure to set the page private bit properly
4117          * after the extent buffer is in the radix tree so
4118          * it doesn't get lost
4119          */
4120         SetPageChecked(eb->pages[0]);
4121         for (i = 1; i < num_pages; i++) {
4122                 p = extent_buffer_page(eb, i);
4123                 ClearPageChecked(p);
4124                 unlock_page(p);
4125         }
4126         unlock_page(eb->pages[0]);
4127         return eb;
4128
4129 free_eb:
4130         for (i = 0; i < num_pages; i++) {
4131                 if (eb->pages[i])
4132                         unlock_page(eb->pages[i]);
4133         }
4134
4135         if (!atomic_dec_and_test(&eb->refs))
4136                 return exists;
4137         btrfs_release_extent_buffer(eb);
4138         return exists;
4139 }
4140
4141 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
4142                                          u64 start, unsigned long len)
4143 {
4144         struct extent_buffer *eb;
4145
4146         rcu_read_lock();
4147         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4148         if (eb && atomic_inc_not_zero(&eb->refs)) {
4149                 rcu_read_unlock();
4150                 mark_extent_buffer_accessed(eb);
4151                 return eb;
4152         }
4153         rcu_read_unlock();
4154
4155         return NULL;
4156 }
4157
4158 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4159 {
4160         struct extent_buffer *eb =
4161                         container_of(head, struct extent_buffer, rcu_head);
4162
4163         __free_extent_buffer(eb);
4164 }
4165
4166 /* Expects to have eb->eb_lock already held */
4167 static void release_extent_buffer(struct extent_buffer *eb, gfp_t mask)
4168 {
4169         WARN_ON(atomic_read(&eb->refs) == 0);
4170         if (atomic_dec_and_test(&eb->refs)) {
4171                 struct extent_io_tree *tree = eb->tree;
4172
4173                 spin_unlock(&eb->refs_lock);
4174
4175                 spin_lock(&tree->buffer_lock);
4176                 radix_tree_delete(&tree->buffer,
4177                                   eb->start >> PAGE_CACHE_SHIFT);
4178                 spin_unlock(&tree->buffer_lock);
4179
4180                 /* Should be safe to release our pages at this point */
4181                 btrfs_release_extent_buffer_page(eb, 0);
4182
4183                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4184                 return;
4185         }
4186         spin_unlock(&eb->refs_lock);
4187 }
4188
4189 void free_extent_buffer(struct extent_buffer *eb)
4190 {
4191         if (!eb)
4192                 return;
4193
4194         spin_lock(&eb->refs_lock);
4195         if (atomic_read(&eb->refs) == 2 &&
4196             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4197             !extent_buffer_under_io(eb) &&
4198             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4199                 atomic_dec(&eb->refs);
4200
4201         /*
4202          * I know this is terrible, but it's temporary until we stop tracking
4203          * the uptodate bits and such for the extent buffers.
4204          */
4205         release_extent_buffer(eb, GFP_ATOMIC);
4206 }
4207
4208 void free_extent_buffer_stale(struct extent_buffer *eb)
4209 {
4210         if (!eb)
4211                 return;
4212
4213         spin_lock(&eb->refs_lock);
4214         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4215
4216         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4217             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4218                 atomic_dec(&eb->refs);
4219         release_extent_buffer(eb, GFP_NOFS);
4220 }
4221
4222 int clear_extent_buffer_dirty(struct extent_buffer *eb)
4223 {
4224         unsigned long i;
4225         unsigned long num_pages;
4226         struct page *page;
4227
4228         num_pages = num_extent_pages(eb->start, eb->len);
4229         WARN_ON(atomic_read(&eb->refs) == 0);
4230
4231         for (i = 0; i < num_pages; i++) {
4232                 page = extent_buffer_page(eb, i);
4233                 if (!PageDirty(page))
4234                         continue;
4235
4236                 lock_page(page);
4237                 WARN_ON(!PagePrivate(page));
4238
4239                 clear_page_dirty_for_io(page);
4240                 spin_lock_irq(&page->mapping->tree_lock);
4241                 if (!PageDirty(page)) {
4242                         radix_tree_tag_clear(&page->mapping->page_tree,
4243                                                 page_index(page),
4244                                                 PAGECACHE_TAG_DIRTY);
4245                 }
4246                 spin_unlock_irq(&page->mapping->tree_lock);
4247                 ClearPageError(page);
4248                 unlock_page(page);
4249         }
4250         WARN_ON(atomic_read(&eb->refs) == 0);
4251         return 0;
4252 }
4253
4254 int set_extent_buffer_dirty(struct extent_buffer *eb)
4255 {
4256         unsigned long i;
4257         unsigned long num_pages;
4258         int was_dirty = 0;
4259
4260         check_buffer_tree_ref(eb);
4261
4262         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4263
4264         num_pages = num_extent_pages(eb->start, eb->len);
4265         WARN_ON(atomic_read(&eb->refs) == 0);
4266         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4267
4268         for (i = 0; i < num_pages; i++)
4269                 set_page_dirty(extent_buffer_page(eb, i));
4270         return was_dirty;
4271 }
4272
4273 static int range_straddles_pages(u64 start, u64 len)
4274 {
4275         if (len < PAGE_CACHE_SIZE)
4276                 return 1;
4277         if (start & (PAGE_CACHE_SIZE - 1))
4278                 return 1;
4279         if ((start + len) & (PAGE_CACHE_SIZE - 1))
4280                 return 1;
4281         return 0;
4282 }
4283
4284 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4285 {
4286         unsigned long i;
4287         struct page *page;
4288         unsigned long num_pages;
4289
4290         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4291         num_pages = num_extent_pages(eb->start, eb->len);
4292         for (i = 0; i < num_pages; i++) {
4293                 page = extent_buffer_page(eb, i);
4294                 if (page)
4295                         ClearPageUptodate(page);
4296         }
4297         return 0;
4298 }
4299
4300 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4301 {
4302         unsigned long i;
4303         struct page *page;
4304         unsigned long num_pages;
4305
4306         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4307         num_pages = num_extent_pages(eb->start, eb->len);
4308         for (i = 0; i < num_pages; i++) {
4309                 page = extent_buffer_page(eb, i);
4310                 SetPageUptodate(page);
4311         }
4312         return 0;
4313 }
4314
4315 int extent_range_uptodate(struct extent_io_tree *tree,
4316                           u64 start, u64 end)
4317 {
4318         struct page *page;
4319         int ret;
4320         int pg_uptodate = 1;
4321         int uptodate;
4322         unsigned long index;
4323
4324         if (range_straddles_pages(start, end - start + 1)) {
4325                 ret = test_range_bit(tree, start, end,
4326                                      EXTENT_UPTODATE, 1, NULL);
4327                 if (ret)
4328                         return 1;
4329         }
4330         while (start <= end) {
4331                 index = start >> PAGE_CACHE_SHIFT;
4332                 page = find_get_page(tree->mapping, index);
4333                 if (!page)
4334                         return 1;
4335                 uptodate = PageUptodate(page);
4336                 page_cache_release(page);
4337                 if (!uptodate) {
4338                         pg_uptodate = 0;
4339                         break;
4340                 }
4341                 start += PAGE_CACHE_SIZE;
4342         }
4343         return pg_uptodate;
4344 }
4345
4346 int extent_buffer_uptodate(struct extent_buffer *eb)
4347 {
4348         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4349 }
4350
4351 int read_extent_buffer_pages(struct extent_io_tree *tree,
4352                              struct extent_buffer *eb, u64 start, int wait,
4353                              get_extent_t *get_extent, int mirror_num)
4354 {
4355         unsigned long i;
4356         unsigned long start_i;
4357         struct page *page;
4358         int err;
4359         int ret = 0;
4360         int locked_pages = 0;
4361         int all_uptodate = 1;
4362         unsigned long num_pages;
4363         unsigned long num_reads = 0;
4364         struct bio *bio = NULL;
4365         unsigned long bio_flags = 0;
4366
4367         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4368                 return 0;
4369
4370         if (start) {
4371                 WARN_ON(start < eb->start);
4372                 start_i = (start >> PAGE_CACHE_SHIFT) -
4373                         (eb->start >> PAGE_CACHE_SHIFT);
4374         } else {
4375                 start_i = 0;
4376         }
4377
4378         num_pages = num_extent_pages(eb->start, eb->len);
4379         for (i = start_i; i < num_pages; i++) {
4380                 page = extent_buffer_page(eb, i);
4381                 if (wait == WAIT_NONE) {
4382                         if (!trylock_page(page))
4383                                 goto unlock_exit;
4384                 } else {
4385                         lock_page(page);
4386                 }
4387                 locked_pages++;
4388                 if (!PageUptodate(page)) {
4389                         num_reads++;
4390                         all_uptodate = 0;
4391                 }
4392         }
4393         if (all_uptodate) {
4394                 if (start_i == 0)
4395                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4396                 goto unlock_exit;
4397         }
4398
4399         atomic_set(&eb->io_pages, num_reads);
4400         for (i = start_i; i < num_pages; i++) {
4401                 page = extent_buffer_page(eb, i);
4402                 if (!PageUptodate(page)) {
4403                         ClearPageError(page);
4404                         err = __extent_read_full_page(tree, page,
4405                                                       get_extent, &bio,
4406                                                       mirror_num, &bio_flags);
4407                         if (err)
4408                                 ret = err;
4409                 } else {
4410                         unlock_page(page);
4411                 }
4412         }
4413
4414         if (bio)
4415                 submit_one_bio(READ, bio, mirror_num, bio_flags);
4416
4417         if (ret || wait != WAIT_COMPLETE)
4418                 return ret;
4419
4420         for (i = start_i; i < num_pages; i++) {
4421                 page = extent_buffer_page(eb, i);
4422                 wait_on_page_locked(page);
4423                 if (!PageUptodate(page))
4424                         ret = -EIO;
4425         }
4426
4427         return ret;
4428
4429 unlock_exit:
4430         i = start_i;
4431         while (locked_pages > 0) {
4432                 page = extent_buffer_page(eb, i);
4433                 i++;
4434                 unlock_page(page);
4435                 locked_pages--;
4436         }
4437         return ret;
4438 }
4439
4440 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4441                         unsigned long start,
4442                         unsigned long len)
4443 {
4444         size_t cur;
4445         size_t offset;
4446         struct page *page;
4447         char *kaddr;
4448         char *dst = (char *)dstv;
4449         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4450         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4451
4452         WARN_ON(start > eb->len);
4453         WARN_ON(start + len > eb->start + eb->len);
4454
4455         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4456
4457         while (len > 0) {
4458                 page = extent_buffer_page(eb, i);
4459
4460                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4461                 kaddr = page_address(page);
4462                 memcpy(dst, kaddr + offset, cur);
4463
4464                 dst += cur;
4465                 len -= cur;
4466                 offset = 0;
4467                 i++;
4468         }
4469 }
4470
4471 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4472                                unsigned long min_len, char **map,
4473                                unsigned long *map_start,
4474                                unsigned long *map_len)
4475 {
4476         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4477         char *kaddr;
4478         struct page *p;
4479         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4480         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4481         unsigned long end_i = (start_offset + start + min_len - 1) >>
4482                 PAGE_CACHE_SHIFT;
4483
4484         if (i != end_i)
4485                 return -EINVAL;
4486
4487         if (i == 0) {
4488                 offset = start_offset;
4489                 *map_start = 0;
4490         } else {
4491                 offset = 0;
4492                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4493         }
4494
4495         if (start + min_len > eb->len) {
4496                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4497                        "wanted %lu %lu\n", (unsigned long long)eb->start,
4498                        eb->len, start, min_len);
4499                 WARN_ON(1);
4500                 return -EINVAL;
4501         }
4502
4503         p = extent_buffer_page(eb, i);
4504         kaddr = page_address(p);
4505         *map = kaddr + offset;
4506         *map_len = PAGE_CACHE_SIZE - offset;
4507         return 0;
4508 }
4509
4510 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4511                           unsigned long start,
4512                           unsigned long len)
4513 {
4514         size_t cur;
4515         size_t offset;
4516         struct page *page;
4517         char *kaddr;
4518         char *ptr = (char *)ptrv;
4519         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4520         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4521         int ret = 0;
4522
4523         WARN_ON(start > eb->len);
4524         WARN_ON(start + len > eb->start + eb->len);
4525
4526         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4527
4528         while (len > 0) {
4529                 page = extent_buffer_page(eb, i);
4530
4531                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4532
4533                 kaddr = page_address(page);
4534                 ret = memcmp(ptr, kaddr + offset, cur);
4535                 if (ret)
4536                         break;
4537
4538                 ptr += cur;
4539                 len -= cur;
4540                 offset = 0;
4541                 i++;
4542         }
4543         return ret;
4544 }
4545
4546 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4547                          unsigned long start, unsigned long len)
4548 {
4549         size_t cur;
4550         size_t offset;
4551         struct page *page;
4552         char *kaddr;
4553         char *src = (char *)srcv;
4554         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4555         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4556
4557         WARN_ON(start > eb->len);
4558         WARN_ON(start + len > eb->start + eb->len);
4559
4560         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4561
4562         while (len > 0) {
4563                 page = extent_buffer_page(eb, i);
4564                 WARN_ON(!PageUptodate(page));
4565
4566                 cur = min(len, PAGE_CACHE_SIZE - offset);
4567                 kaddr = page_address(page);
4568                 memcpy(kaddr + offset, src, cur);
4569
4570                 src += cur;
4571                 len -= cur;
4572                 offset = 0;
4573                 i++;
4574         }
4575 }
4576
4577 void memset_extent_buffer(struct extent_buffer *eb, char c,
4578                           unsigned long start, unsigned long len)
4579 {
4580         size_t cur;
4581         size_t offset;
4582         struct page *page;
4583         char *kaddr;
4584         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4585         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4586
4587         WARN_ON(start > eb->len);
4588         WARN_ON(start + len > eb->start + eb->len);
4589
4590         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4591
4592         while (len > 0) {
4593                 page = extent_buffer_page(eb, i);
4594                 WARN_ON(!PageUptodate(page));
4595
4596                 cur = min(len, PAGE_CACHE_SIZE - offset);
4597                 kaddr = page_address(page);
4598                 memset(kaddr + offset, c, cur);
4599
4600                 len -= cur;
4601                 offset = 0;
4602                 i++;
4603         }
4604 }
4605
4606 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
4607                         unsigned long dst_offset, unsigned long src_offset,
4608                         unsigned long len)
4609 {
4610         u64 dst_len = dst->len;
4611         size_t cur;
4612         size_t offset;
4613         struct page *page;
4614         char *kaddr;
4615         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4616         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4617
4618         WARN_ON(src->len != dst_len);
4619
4620         offset = (start_offset + dst_offset) &
4621                 ((unsigned long)PAGE_CACHE_SIZE - 1);
4622
4623         while (len > 0) {
4624                 page = extent_buffer_page(dst, i);
4625                 WARN_ON(!PageUptodate(page));
4626
4627                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
4628
4629                 kaddr = page_address(page);
4630                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4631
4632                 src_offset += cur;
4633                 len -= cur;
4634                 offset = 0;
4635                 i++;
4636         }
4637 }
4638
4639 static void move_pages(struct page *dst_page, struct page *src_page,
4640                        unsigned long dst_off, unsigned long src_off,
4641                        unsigned long len)
4642 {
4643         char *dst_kaddr = page_address(dst_page);
4644         if (dst_page == src_page) {
4645                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
4646         } else {
4647                 char *src_kaddr = page_address(src_page);
4648                 char *p = dst_kaddr + dst_off + len;
4649                 char *s = src_kaddr + src_off + len;
4650
4651                 while (len--)
4652                         *--p = *--s;
4653         }
4654 }
4655
4656 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4657 {
4658         unsigned long distance = (src > dst) ? src - dst : dst - src;
4659         return distance < len;
4660 }
4661
4662 static void copy_pages(struct page *dst_page, struct page *src_page,
4663                        unsigned long dst_off, unsigned long src_off,
4664                        unsigned long len)
4665 {
4666         char *dst_kaddr = page_address(dst_page);
4667         char *src_kaddr;
4668         int must_memmove = 0;
4669
4670         if (dst_page != src_page) {
4671                 src_kaddr = page_address(src_page);
4672         } else {
4673                 src_kaddr = dst_kaddr;
4674                 if (areas_overlap(src_off, dst_off, len))
4675                         must_memmove = 1;
4676         }
4677
4678         if (must_memmove)
4679                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
4680         else
4681                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
4682 }
4683
4684 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4685                            unsigned long src_offset, unsigned long len)
4686 {
4687         size_t cur;
4688         size_t dst_off_in_page;
4689         size_t src_off_in_page;
4690         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4691         unsigned long dst_i;
4692         unsigned long src_i;
4693
4694         if (src_offset + len > dst->len) {
4695                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4696                        "len %lu dst len %lu\n", src_offset, len, dst->len);
4697                 BUG_ON(1);
4698         }
4699         if (dst_offset + len > dst->len) {
4700                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4701                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
4702                 BUG_ON(1);
4703         }
4704
4705         while (len > 0) {
4706                 dst_off_in_page = (start_offset + dst_offset) &
4707                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4708                 src_off_in_page = (start_offset + src_offset) &
4709                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4710
4711                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4712                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
4713
4714                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
4715                                                src_off_in_page));
4716                 cur = min_t(unsigned long, cur,
4717                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
4718
4719                 copy_pages(extent_buffer_page(dst, dst_i),
4720                            extent_buffer_page(dst, src_i),
4721                            dst_off_in_page, src_off_in_page, cur);
4722
4723                 src_offset += cur;
4724                 dst_offset += cur;
4725                 len -= cur;
4726         }
4727 }
4728
4729 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4730                            unsigned long src_offset, unsigned long len)
4731 {
4732         size_t cur;
4733         size_t dst_off_in_page;
4734         size_t src_off_in_page;
4735         unsigned long dst_end = dst_offset + len - 1;
4736         unsigned long src_end = src_offset + len - 1;
4737         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4738         unsigned long dst_i;
4739         unsigned long src_i;
4740
4741         if (src_offset + len > dst->len) {
4742                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4743                        "len %lu len %lu\n", src_offset, len, dst->len);
4744                 BUG_ON(1);
4745         }
4746         if (dst_offset + len > dst->len) {
4747                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4748                        "len %lu len %lu\n", dst_offset, len, dst->len);
4749                 BUG_ON(1);
4750         }
4751         if (dst_offset < src_offset) {
4752                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4753                 return;
4754         }
4755         while (len > 0) {
4756                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4757                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4758
4759                 dst_off_in_page = (start_offset + dst_end) &
4760                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4761                 src_off_in_page = (start_offset + src_end) &
4762                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4763
4764                 cur = min_t(unsigned long, len, src_off_in_page + 1);
4765                 cur = min(cur, dst_off_in_page + 1);
4766                 move_pages(extent_buffer_page(dst, dst_i),
4767                            extent_buffer_page(dst, src_i),
4768                            dst_off_in_page - cur + 1,
4769                            src_off_in_page - cur + 1, cur);
4770
4771                 dst_end -= cur;
4772                 src_end -= cur;
4773                 len -= cur;
4774         }
4775 }
4776
4777 int try_release_extent_buffer(struct page *page, gfp_t mask)
4778 {
4779         struct extent_buffer *eb;
4780
4781         /*
4782          * We need to make sure noboody is attaching this page to an eb right
4783          * now.
4784          */
4785         spin_lock(&page->mapping->private_lock);
4786         if (!PagePrivate(page)) {
4787                 spin_unlock(&page->mapping->private_lock);
4788                 return 1;
4789         }
4790
4791         eb = (struct extent_buffer *)page->private;
4792         BUG_ON(!eb);
4793
4794         /*
4795          * This is a little awful but should be ok, we need to make sure that
4796          * the eb doesn't disappear out from under us while we're looking at
4797          * this page.
4798          */
4799         spin_lock(&eb->refs_lock);
4800         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4801                 spin_unlock(&eb->refs_lock);
4802                 spin_unlock(&page->mapping->private_lock);
4803                 return 0;
4804         }
4805         spin_unlock(&page->mapping->private_lock);
4806
4807         if ((mask & GFP_NOFS) == GFP_NOFS)
4808                 mask = GFP_NOFS;
4809
4810         /*
4811          * If tree ref isn't set then we know the ref on this eb is a real ref,
4812          * so just return, this page will likely be freed soon anyway.
4813          */
4814         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4815                 spin_unlock(&eb->refs_lock);
4816                 return 0;
4817         }
4818         release_extent_buffer(eb, mask);
4819
4820         return 1;
4821 }