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 static inline struct btrfs_fs_info *
  59 tree_fs_info(struct extent_io_tree *tree)
  60 {
  61         return btrfs_sb(tree->mapping->host->i_sb);
  62 }
  63
  64 int __init extent_io_init(void)
  65 {
  66         extent_state_cache = kmem_cache_create("extent_state",
  67                         sizeof(struct extent_state), 0,
  68                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  69         if (!extent_state_cache)
  70                 return -ENOMEM;
  71
  72         extent_buffer_cache = kmem_cache_create("extent_buffers",
  73                         sizeof(struct extent_buffer), 0,
  74                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  75         if (!extent_buffer_cache)
  76                 goto free_state_cache;
  77         return 0;
  78
  79 free_state_cache:
  80         kmem_cache_destroy(extent_state_cache);
  81         return -ENOMEM;
  82 }
  83
  84 void extent_io_exit(void)
  85 {
  86         struct extent_state *state;
  87         struct extent_buffer *eb;
  88
  89         while (!list_empty(&states)) {
  90                 state = list_entry(states.next, struct extent_state, leak_list);
  91                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  92                        "state %lu in tree %p refs %d\n",
  93                        (unsigned long long)state->start,
  94                        (unsigned long long)state->end,
  95                        state->state, state->tree, atomic_read(&state->refs));
  96                 list_del(&state->leak_list);
  97                 kmem_cache_free(extent_state_cache, state);
  98
  99         }
 100
 101         while (!list_empty(&buffers)) {
 102                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
 103                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
 104                        "refs %d\n", (unsigned long long)eb->start,
 105                        eb->len, atomic_read(&eb->refs));
 106                 list_del(&eb->leak_list);
 107                 kmem_cache_free(extent_buffer_cache, eb);
 108         }
 109         if (extent_state_cache)
 110                 kmem_cache_destroy(extent_state_cache);
 111         if (extent_buffer_cache)
 112                 kmem_cache_destroy(extent_buffer_cache);
 113 }
 114
 115 void extent_io_tree_init(struct extent_io_tree *tree,
 116                          struct address_space *mapping)
 117 {
 118         tree->state = RB_ROOT;
 119         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 120         tree->ops = NULL;
 121         tree->dirty_bytes = 0;
 122         spin_lock_init(&tree->lock);
 123         spin_lock_init(&tree->buffer_lock);
 124         tree->mapping = mapping;
 125 }
 126
 127 static struct extent_state *alloc_extent_state(gfp_t mask)
 128 {
 129         struct extent_state *state;
 130 #if LEAK_DEBUG
 131         unsigned long flags;
 132 #endif
 133
 134         state = kmem_cache_alloc(extent_state_cache, mask);
 135         if (!state)
 136                 return state;
 137         state->state = 0;
 138         state->private = 0;
 139         state->tree = NULL;
 140 #if LEAK_DEBUG
 141         spin_lock_irqsave(&leak_lock, flags);
 142         list_add(&state->leak_list, &states);
 143         spin_unlock_irqrestore(&leak_lock, flags);
 144 #endif
 145         atomic_set(&state->refs, 1);
 146         init_waitqueue_head(&state->wq);
 147         trace_alloc_extent_state(state, mask, _RET_IP_);
 148         return state;
 149 }
 150
 151 void free_extent_state(struct extent_state *state)
 152 {
 153         if (!state)
 154                 return;
 155         if (atomic_dec_and_test(&state->refs)) {
 156 #if LEAK_DEBUG
 157                 unsigned long flags;
 158 #endif
 159                 WARN_ON(state->tree);
 160 #if LEAK_DEBUG
 161                 spin_lock_irqsave(&leak_lock, flags);
 162                 list_del(&state->leak_list);
 163                 spin_unlock_irqrestore(&leak_lock, flags);
 164 #endif
 165                 trace_free_extent_state(state, _RET_IP_);
 166                 kmem_cache_free(extent_state_cache, state);
 167         }
 168 }
 169
 170 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 171                                    struct rb_node *node)
 172 {
 173         struct rb_node **p = &root->rb_node;
 174         struct rb_node *parent = NULL;
 175         struct tree_entry *entry;
 176
 177         while (*p) {
 178                 parent = *p;
 179                 entry = rb_entry(parent, struct tree_entry, rb_node);
 180
 181                 if (offset < entry->start)
 182                         p = &(*p)->rb_left;
 183                 else if (offset > entry->end)
 184                         p = &(*p)->rb_right;
 185                 else
 186                         return parent;
 187         }
 188
 189         entry = rb_entry(node, struct tree_entry, rb_node);
 190         rb_link_node(node, parent, p);
 191         rb_insert_color(node, root);
 192         return NULL;
 193 }
 194
 195 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 196                                      struct rb_node **prev_ret,
 197                                      struct rb_node **next_ret)
 198 {
 199         struct rb_root *root = &tree->state;
 200         struct rb_node *n = root->rb_node;
 201         struct rb_node *prev = NULL;
 202         struct rb_node *orig_prev = NULL;
 203         struct tree_entry *entry;
 204         struct tree_entry *prev_entry = NULL;
 205
 206         while (n) {
 207                 entry = rb_entry(n, struct tree_entry, rb_node);
 208                 prev = n;
 209                 prev_entry = entry;
 210
 211                 if (offset < entry->start)
 212                         n = n->rb_left;
 213                 else if (offset > entry->end)
 214                         n = n->rb_right;
 215                 else
 216                         return n;
 217         }
 218
 219         if (prev_ret) {
 220                 orig_prev = prev;
 221                 while (prev && offset > prev_entry->end) {
 222                         prev = rb_next(prev);
 223                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 224                 }
 225                 *prev_ret = prev;
 226                 prev = orig_prev;
 227         }
 228
 229         if (next_ret) {
 230                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 231                 while (prev && offset < prev_entry->start) {
 232                         prev = rb_prev(prev);
 233                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 234                 }
 235                 *next_ret = prev;
 236         }
 237         return NULL;
 238 }
 239
 240 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 241                                           u64 offset)
 242 {
 243         struct rb_node *prev = NULL;
 244         struct rb_node *ret;
 245
 246         ret = __etree_search(tree, offset, &prev, NULL);
 247         if (!ret)
 248                 return prev;
 249         return ret;
 250 }
 251
 252 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 253                      struct extent_state *other)
 254 {
 255         if (tree->ops && tree->ops->merge_extent_hook)
 256                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 257                                              other);
 258 }
 259
 260 /*
 261  * utility function to look for merge candidates inside a given range.
 262  * Any extents with matching state are merged together into a single
 263  * extent in the tree.  Extents with EXTENT_IO in their state field
 264  * are not merged because the end_io handlers need to be able to do
 265  * operations on them without sleeping (or doing allocations/splits).
 266  *
 267  * This should be called with the tree lock held.
 268  */
 269 static void merge_state(struct extent_io_tree *tree,
 270                         struct extent_state *state)
 271 {
 272         struct extent_state *other;
 273         struct rb_node *other_node;
 274
 275         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 276                 return;
 277
 278         other_node = rb_prev(&state->rb_node);
 279         if (other_node) {
 280                 other = rb_entry(other_node, struct extent_state, rb_node);
 281                 if (other->end == state->start - 1 &&
 282                     other->state == state->state) {
 283                         merge_cb(tree, state, other);
 284                         state->start = other->start;
 285                         other->tree = NULL;
 286                         rb_erase(&other->rb_node, &tree->state);
 287                         free_extent_state(other);
 288                 }
 289         }
 290         other_node = rb_next(&state->rb_node);
 291         if (other_node) {
 292                 other = rb_entry(other_node, struct extent_state, rb_node);
 293                 if (other->start == state->end + 1 &&
 294                     other->state == state->state) {
 295                         merge_cb(tree, state, other);
 296                         state->end = other->end;
 297                         other->tree = NULL;
 298                         rb_erase(&other->rb_node, &tree->state);
 299                         free_extent_state(other);
 300                 }
 301         }
 302 }
 303
 304 static void set_state_cb(struct extent_io_tree *tree,
 305                          struct extent_state *state, int *bits)
 306 {
 307         if (tree->ops && tree->ops->set_bit_hook)
 308                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 309 }
 310
 311 static void clear_state_cb(struct extent_io_tree *tree,
 312                            struct extent_state *state, int *bits)
 313 {
 314         if (tree->ops && tree->ops->clear_bit_hook)
 315                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 316 }
 317
 318 static void set_state_bits(struct extent_io_tree *tree,
 319                            struct extent_state *state, int *bits);
 320
 321 /*
 322  * insert an extent_state struct into the tree.  'bits' are set on the
 323  * struct before it is inserted.
 324  *
 325  * This may return -EEXIST if the extent is already there, in which case the
 326  * state struct is freed.
 327  *
 328  * The tree lock is not taken internally.  This is a utility function and
 329  * probably isn't what you want to call (see set/clear_extent_bit).
 330  */
 331 static int insert_state(struct extent_io_tree *tree,
 332                         struct extent_state *state, u64 start, u64 end,
 333                         int *bits)
 334 {
 335         struct rb_node *node;
 336
 337         if (end < start) {
 338                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
 339                        (unsigned long long)end,
 340                        (unsigned long long)start);
 341                 WARN_ON(1);
 342         }
 343         state->start = start;
 344         state->end = end;
 345
 346         set_state_bits(tree, state, bits);
 347
 348         node = tree_insert(&tree->state, end, &state->rb_node);
 349         if (node) {
 350                 struct extent_state *found;
 351                 found = rb_entry(node, struct extent_state, rb_node);
 352                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 353                        "%llu %llu\n", (unsigned long long)found->start,
 354                        (unsigned long long)found->end,
 355                        (unsigned long long)start, (unsigned long long)end);
 356                 return -EEXIST;
 357         }
 358         state->tree = tree;
 359         merge_state(tree, state);
 360         return 0;
 361 }
 362
 363 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 364                      u64 split)
 365 {
 366         if (tree->ops && tree->ops->split_extent_hook)
 367                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 368 }
 369
 370 /*
 371  * split a given extent state struct in two, inserting the preallocated
 372  * struct 'prealloc' as the newly created second half.  'split' indicates an
 373  * offset inside 'orig' where it should be split.
 374  *
 375  * Before calling,
 376  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 377  * are two extent state structs in the tree:
 378  * prealloc: [orig->start, split - 1]
 379  * orig: [ split, orig->end ]
 380  *
 381  * The tree locks are not taken by this function. They need to be held
 382  * by the caller.
 383  */
 384 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 385                        struct extent_state *prealloc, u64 split)
 386 {
 387         struct rb_node *node;
 388
 389         split_cb(tree, orig, split);
 390
 391         prealloc->start = orig->start;
 392         prealloc->end = split - 1;
 393         prealloc->state = orig->state;
 394         orig->start = split;
 395
 396         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 397         if (node) {
 398                 free_extent_state(prealloc);
 399                 return -EEXIST;
 400         }
 401         prealloc->tree = tree;
 402         return 0;
 403 }
 404
 405 /*
 406  * utility function to clear some bits in an extent state struct.
 407  * it will optionally wake up any one waiting on this state (wake == 1), or
 408  * forcibly remove the state from the tree (delete == 1).
 409  *
 410  * If no bits are set on the state struct after clearing things, the
 411  * struct is freed and removed from the tree
 412  */
 413 static int clear_state_bit(struct extent_io_tree *tree,
 414                             struct extent_state *state,
 415                             int *bits, int wake)
 416 {
 417         int bits_to_clear = *bits & ~EXTENT_CTLBITS;
 418         int ret = state->state & bits_to_clear;
 419
 420         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 421                 u64 range = state->end - state->start + 1;
 422                 WARN_ON(range > tree->dirty_bytes);
 423                 tree->dirty_bytes -= range;
 424         }
 425         clear_state_cb(tree, state, bits);
 426         state->state &= ~bits_to_clear;
 427         if (wake)
 428                 wake_up(&state->wq);
 429         if (state->state == 0) {
 430                 if (state->tree) {
 431                         rb_erase(&state->rb_node, &tree->state);
 432                         state->tree = NULL;
 433                         free_extent_state(state);
 434                 } else {
 435                         WARN_ON(1);
 436                 }
 437         } else {
 438                 merge_state(tree, state);
 439         }
 440         return ret;
 441 }
 442
 443 static struct extent_state *
 444 alloc_extent_state_atomic(struct extent_state *prealloc)
 445 {
 446         if (!prealloc)
 447                 prealloc = alloc_extent_state(GFP_ATOMIC);
 448
 449         return prealloc;
 450 }
 451
 452 void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 453 {
 454         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 455                     "Extent tree was modified by another "
 456                     "thread while locked.");
 457 }
 458
 459 /*
 460  * clear some bits on a range in the tree.  This may require splitting
 461  * or inserting elements in the tree, so the gfp mask is used to
 462  * indicate which allocations or sleeping are allowed.
 463  *
 464  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 465  * the given range from the tree regardless of state (ie for truncate).
 466  *
 467  * the range [start, end] is inclusive.
 468  *
 469  * This takes the tree lock, and returns 0 on success and < 0 on error.
 470  */
 471 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 472                      int bits, int wake, int delete,
 473                      struct extent_state **cached_state,
 474                      gfp_t mask)
 475 {
 476         struct extent_state *state;
 477         struct extent_state *cached;
 478         struct extent_state *prealloc = NULL;
 479         struct rb_node *next_node;
 480         struct rb_node *node;
 481         u64 last_end;
 482         int err;
 483         int clear = 0;
 484
 485         if (delete)
 486                 bits |= ~EXTENT_CTLBITS;
 487         bits |= EXTENT_FIRST_DELALLOC;
 488
 489         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 490                 clear = 1;
 491 again:
 492         if (!prealloc && (mask & __GFP_WAIT)) {
 493                 prealloc = alloc_extent_state(mask);
 494                 if (!prealloc)
 495                         return -ENOMEM;
 496         }
 497
 498         spin_lock(&tree->lock);
 499         if (cached_state) {
 500                 cached = *cached_state;
 501
 502                 if (clear) {
 503                         *cached_state = NULL;
 504                         cached_state = NULL;
 505                 }
 506
 507                 if (cached && cached->tree && cached->start <= start &&
 508                     cached->end > start) {
 509                         if (clear)
 510                                 atomic_dec(&cached->refs);
 511                         state = cached;
 512                         goto hit_next;
 513                 }
 514                 if (clear)
 515                         free_extent_state(cached);
 516         }
 517         /*
 518          * this search will find the extents that end after
 519          * our range starts
 520          */
 521         node = tree_search(tree, start);
 522         if (!node)
 523                 goto out;
 524         state = rb_entry(node, struct extent_state, rb_node);
 525 hit_next:
 526         if (state->start > end)
 527                 goto out;
 528         WARN_ON(state->end < start);
 529         last_end = state->end;
 530
 531         if (state->end < end && !need_resched())
 532                 next_node = rb_next(&state->rb_node);
 533         else
 534                 next_node = NULL;
 535
 536         /* the state doesn't have the wanted bits, go ahead */
 537         if (!(state->state & bits))
 538                 goto next;
 539
 540         /*
 541          *     | ---- desired range ---- |
 542          *  | state | or
 543          *  | ------------- state -------------- |
 544          *
 545          * We need to split the extent we found, and may flip
 546          * bits on second half.
 547          *
 548          * If the extent we found extends past our range, we
 549          * just split and search again.  It'll get split again
 550          * the next time though.
 551          *
 552          * If the extent we found is inside our range, we clear
 553          * the desired bit on it.
 554          */
 555
 556         if (state->start < start) {
 557                 prealloc = alloc_extent_state_atomic(prealloc);
 558                 BUG_ON(!prealloc);
 559                 err = split_state(tree, state, prealloc, start);
 560                 if (err)
 561                         extent_io_tree_panic(tree, err);
 562
 563                 prealloc = NULL;
 564                 if (err)
 565                         goto out;
 566                 if (state->end <= end) {
 567                         clear_state_bit(tree, state, &bits, wake);
 568                         if (last_end == (u64)-1)
 569                                 goto out;
 570                         start = last_end + 1;
 571                 }
 572                 goto search_again;
 573         }
 574         /*
 575          * | ---- desired range ---- |
 576          *                        | state |
 577          * We need to split the extent, and clear the bit
 578          * on the first half
 579          */
 580         if (state->start <= end && state->end > end) {
 581                 prealloc = alloc_extent_state_atomic(prealloc);
 582                 BUG_ON(!prealloc);
 583                 err = split_state(tree, state, prealloc, end + 1);
 584                 if (err)
 585                         extent_io_tree_panic(tree, err);
 586
 587                 if (wake)
 588                         wake_up(&state->wq);
 589
 590                 clear_state_bit(tree, prealloc, &bits, wake);
 591
 592                 prealloc = NULL;
 593                 goto out;
 594         }
 595
 596         clear_state_bit(tree, state, &bits, wake);
 597 next:
 598         if (last_end == (u64)-1)
 599                 goto out;
 600         start = last_end + 1;
 601         if (start <= end && next_node) {
 602                 state = rb_entry(next_node, struct extent_state,
 603                                  rb_node);
 604                 goto hit_next;
 605         }
 606         goto search_again;
 607
 608 out:
 609         spin_unlock(&tree->lock);
 610         if (prealloc)
 611                 free_extent_state(prealloc);
 612
 613         return 0;
 614
 615 search_again:
 616         if (start > end)
 617                 goto out;
 618         spin_unlock(&tree->lock);
 619         if (mask & __GFP_WAIT)
 620                 cond_resched();
 621         goto again;
 622 }
 623
 624 static void wait_on_state(struct extent_io_tree *tree,
 625                           struct extent_state *state)
 626                 __releases(tree->lock)
 627                 __acquires(tree->lock)
 628 {
 629         DEFINE_WAIT(wait);
 630         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 631         spin_unlock(&tree->lock);
 632         schedule();
 633         spin_lock(&tree->lock);
 634         finish_wait(&state->wq, &wait);
 635 }
 636
 637 /*
 638  * waits for one or more bits to clear on a range in the state tree.
 639  * The range [start, end] is inclusive.
 640  * The tree lock is taken by this function
 641  */
 642 void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 643 {
 644         struct extent_state *state;
 645         struct rb_node *node;
 646
 647         spin_lock(&tree->lock);
 648 again:
 649         while (1) {
 650                 /*
 651                  * this search will find all the extents that end after
 652                  * our range starts
 653                  */
 654                 node = tree_search(tree, start);
 655                 if (!node)
 656                         break;
 657
 658                 state = rb_entry(node, struct extent_state, rb_node);
 659
 660                 if (state->start > end)
 661                         goto out;
 662
 663                 if (state->state & bits) {
 664                         start = state->start;
 665                         atomic_inc(&state->refs);
 666                         wait_on_state(tree, state);
 667                         free_extent_state(state);
 668                         goto again;
 669                 }
 670                 start = state->end + 1;
 671
 672                 if (start > end)
 673                         break;
 674
 675                 cond_resched_lock(&tree->lock);
 676         }
 677 out:
 678         spin_unlock(&tree->lock);
 679 }
 680
 681 static void set_state_bits(struct extent_io_tree *tree,
 682                            struct extent_state *state,
 683                            int *bits)
 684 {
 685         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 686
 687         set_state_cb(tree, state, bits);
 688         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 689                 u64 range = state->end - state->start + 1;
 690                 tree->dirty_bytes += range;
 691         }
 692         state->state |= bits_to_set;
 693 }
 694
 695 static void cache_state(struct extent_state *state,
 696                         struct extent_state **cached_ptr)
 697 {
 698         if (cached_ptr && !(*cached_ptr)) {
 699                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 700                         *cached_ptr = state;
 701                         atomic_inc(&state->refs);
 702                 }
 703         }
 704 }
 705
 706 static void uncache_state(struct extent_state **cached_ptr)
 707 {
 708         if (cached_ptr && (*cached_ptr)) {
 709                 struct extent_state *state = *cached_ptr;
 710                 *cached_ptr = NULL;
 711                 free_extent_state(state);
 712         }
 713 }
 714
 715 /*
 716  * set some bits on a range in the tree.  This may require allocations or
 717  * sleeping, so the gfp mask is used to indicate what is allowed.
 718  *
 719  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 720  * part of the range already has the desired bits set.  The start of the
 721  * existing range is returned in failed_start in this case.
 722  *
 723  * [start, end] is inclusive This takes the tree lock.
 724  */
 725
 726 static int __must_check
 727 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 728                  int bits, int exclusive_bits, u64 *failed_start,
 729                  struct extent_state **cached_state, gfp_t mask)
 730 {
 731         struct extent_state *state;
 732         struct extent_state *prealloc = NULL;
 733         struct rb_node *node;
 734         int err = 0;
 735         u64 last_start;
 736         u64 last_end;
 737
 738         bits |= EXTENT_FIRST_DELALLOC;
 739 again:
 740         if (!prealloc && (mask & __GFP_WAIT)) {
 741                 prealloc = alloc_extent_state(mask);
 742                 BUG_ON(!prealloc);
 743         }
 744
 745         spin_lock(&tree->lock);
 746         if (cached_state && *cached_state) {
 747                 state = *cached_state;
 748                 if (state->start <= start && state->end > start &&
 749                     state->tree) {
 750                         node = &state->rb_node;
 751                         goto hit_next;
 752                 }
 753         }
 754         /*
 755          * this search will find all the extents that end after
 756          * our range starts.
 757          */
 758         node = tree_search(tree, start);
 759         if (!node) {
 760                 prealloc = alloc_extent_state_atomic(prealloc);
 761                 BUG_ON(!prealloc);
 762                 err = insert_state(tree, prealloc, start, end, &bits);
 763                 if (err)
 764                         extent_io_tree_panic(tree, err);
 765
 766                 prealloc = NULL;
 767                 goto out;
 768         }
 769         state = rb_entry(node, struct extent_state, rb_node);
 770 hit_next:
 771         last_start = state->start;
 772         last_end = state->end;
 773
 774         /*
 775          * | ---- desired range ---- |
 776          * | state |
 777          *
 778          * Just lock what we found and keep going
 779          */
 780         if (state->start == start && state->end <= end) {
 781                 struct rb_node *next_node;
 782                 if (state->state & exclusive_bits) {
 783                         *failed_start = state->start;
 784                         err = -EEXIST;
 785                         goto out;
 786                 }
 787
 788                 set_state_bits(tree, state, &bits);
 789
 790                 cache_state(state, cached_state);
 791                 merge_state(tree, state);
 792                 if (last_end == (u64)-1)
 793                         goto out;
 794
 795                 start = last_end + 1;
 796                 next_node = rb_next(&state->rb_node);
 797                 if (next_node && start < end && prealloc && !need_resched()) {
 798                         state = rb_entry(next_node, struct extent_state,
 799                                          rb_node);
 800                         if (state->start == start)
 801                                 goto hit_next;
 802                 }
 803                 goto search_again;
 804         }
 805
 806         /*
 807          *     | ---- desired range ---- |
 808          * | state |
 809          *   or
 810          * | ------------- state -------------- |
 811          *
 812          * We need to split the extent we found, and may flip bits on
 813          * second half.
 814          *
 815          * If the extent we found extends past our
 816          * range, we just split and search again.  It'll get split
 817          * again the next time though.
 818          *
 819          * If the extent we found is inside our range, we set the
 820          * desired bit on it.
 821          */
 822         if (state->start < start) {
 823                 if (state->state & exclusive_bits) {
 824                         *failed_start = start;
 825                         err = -EEXIST;
 826                         goto out;
 827                 }
 828
 829                 prealloc = alloc_extent_state_atomic(prealloc);
 830                 BUG_ON(!prealloc);
 831                 err = split_state(tree, state, prealloc, start);
 832                 if (err)
 833                         extent_io_tree_panic(tree, err);
 834
 835                 prealloc = NULL;
 836                 if (err)
 837                         goto out;
 838                 if (state->end <= end) {
 839                         set_state_bits(tree, state, &bits);
 840                         cache_state(state, cached_state);
 841                         merge_state(tree, state);
 842                         if (last_end == (u64)-1)
 843                                 goto out;
 844                         start = last_end + 1;
 845                 }
 846                 goto search_again;
 847         }
 848         /*
 849          * | ---- desired range ---- |
 850          *     | state | or               | state |
 851          *
 852          * There's a hole, we need to insert something in it and
 853          * ignore the extent we found.
 854          */
 855         if (state->start > start) {
 856                 u64 this_end;
 857                 if (end < last_start)
 858                         this_end = end;
 859                 else
 860                         this_end = last_start - 1;
 861
 862                 prealloc = alloc_extent_state_atomic(prealloc);
 863                 BUG_ON(!prealloc);
 864
 865                 /*
 866                  * Avoid to free 'prealloc' if it can be merged with
 867                  * the later extent.
 868                  */
 869                 err = insert_state(tree, prealloc, start, this_end,
 870                                    &bits);
 871                 if (err)
 872                         extent_io_tree_panic(tree, err);
 873
 874                 cache_state(prealloc, cached_state);
 875                 prealloc = NULL;
 876                 start = this_end + 1;
 877                 goto search_again;
 878         }
 879         /*
 880          * | ---- desired range ---- |
 881          *                        | state |
 882          * We need to split the extent, and set the bit
 883          * on the first half
 884          */
 885         if (state->start <= end && state->end > end) {
 886                 if (state->state & exclusive_bits) {
 887                         *failed_start = start;
 888                         err = -EEXIST;
 889                         goto out;
 890                 }
 891
 892                 prealloc = alloc_extent_state_atomic(prealloc);
 893                 BUG_ON(!prealloc);
 894                 err = split_state(tree, state, prealloc, end + 1);
 895                 if (err)
 896                         extent_io_tree_panic(tree, err);
 897
 898                 set_state_bits(tree, prealloc, &bits);
 899                 cache_state(prealloc, cached_state);
 900                 merge_state(tree, prealloc);
 901                 prealloc = NULL;
 902                 goto out;
 903         }
 904
 905         goto search_again;
 906
 907 out:
 908         spin_unlock(&tree->lock);
 909         if (prealloc)
 910                 free_extent_state(prealloc);
 911
 912         return err;
 913
 914 search_again:
 915         if (start > end)
 916                 goto out;
 917         spin_unlock(&tree->lock);
 918         if (mask & __GFP_WAIT)
 919                 cond_resched();
 920         goto again;
 921 }
 922
 923 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits,
 924                    u64 *failed_start, struct extent_state **cached_state,
 925                    gfp_t mask)
 926 {
 927         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
 928                                 cached_state, mask);
 929 }
 930
 931
 932 /**
 933  * convert_extent - convert all bits in a given range from one bit to another
 934  * @tree:       the io tree to search
 935  * @start:      the start offset in bytes
 936  * @end:        the end offset in bytes (inclusive)
 937  * @bits:       the bits to set in this range
 938  * @clear_bits: the bits to clear in this range
 939  * @mask:       the allocation mask
 940  *
 941  * This will go through and set bits for the given range.  If any states exist
 942  * already in this range they are set with the given bit and cleared of the
 943  * clear_bits.  This is only meant to be used by things that are mergeable, ie
 944  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
 945  * boundary bits like LOCK.
 946  */
 947 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 948                        int bits, int clear_bits, gfp_t mask)
 949 {
 950         struct extent_state *state;
 951         struct extent_state *prealloc = NULL;
 952         struct rb_node *node;
 953         int err = 0;
 954         u64 last_start;
 955         u64 last_end;
 956
 957 again:
 958         if (!prealloc && (mask & __GFP_WAIT)) {
 959                 prealloc = alloc_extent_state(mask);
 960                 if (!prealloc)
 961                         return -ENOMEM;
 962         }
 963
 964         spin_lock(&tree->lock);
 965         /*
 966          * this search will find all the extents that end after
 967          * our range starts.
 968          */
 969         node = tree_search(tree, start);
 970         if (!node) {
 971                 prealloc = alloc_extent_state_atomic(prealloc);
 972                 if (!prealloc) {
 973                         err = -ENOMEM;
 974                         goto out;
 975                 }
 976                 err = insert_state(tree, prealloc, start, end, &bits);
 977                 prealloc = NULL;
 978                 if (err)
 979                         extent_io_tree_panic(tree, err);
 980                 goto out;
 981         }
 982         state = rb_entry(node, struct extent_state, rb_node);
 983 hit_next:
 984         last_start = state->start;
 985         last_end = state->end;
 986
 987         /*
 988          * | ---- desired range ---- |
 989          * | state |
 990          *
 991          * Just lock what we found and keep going
 992          */
 993         if (state->start == start && state->end <= end) {
 994                 struct rb_node *next_node;
 995
 996                 set_state_bits(tree, state, &bits);
 997                 clear_state_bit(tree, state, &clear_bits, 0);
 998                 if (last_end == (u64)-1)
 999                         goto out;
1000
1001                 start = last_end + 1;
1002                 next_node = rb_next(&state->rb_node);
1003                 if (next_node && start < end && prealloc && !need_resched()) {
1004                         state = rb_entry(next_node, struct extent_state,
1005                                          rb_node);
1006                         if (state->start == start)
1007                                 goto hit_next;
1008                 }
1009                 goto search_again;
1010         }
1011
1012         /*
1013          *     | ---- desired range ---- |
1014          * | state |
1015          *   or
1016          * | ------------- state -------------- |
1017          *
1018          * We need to split the extent we found, and may flip bits on
1019          * second half.
1020          *
1021          * If the extent we found extends past our
1022          * range, we just split and search again.  It'll get split
1023          * again the next time though.
1024          *
1025          * If the extent we found is inside our range, we set the
1026          * desired bit on it.
1027          */
1028         if (state->start < start) {
1029                 prealloc = alloc_extent_state_atomic(prealloc);
1030                 if (!prealloc) {
1031                         err = -ENOMEM;
1032                         goto out;
1033                 }
1034                 err = split_state(tree, state, prealloc, start);
1035                 if (err)
1036                         extent_io_tree_panic(tree, err);
1037                 prealloc = NULL;
1038                 if (err)
1039                         goto out;
1040                 if (state->end <= end) {
1041                         set_state_bits(tree, state, &bits);
1042                         clear_state_bit(tree, state, &clear_bits, 0);
1043                         if (last_end == (u64)-1)
1044                                 goto out;
1045                         start = last_end + 1;
1046                 }
1047                 goto search_again;
1048         }
1049         /*
1050          * | ---- desired range ---- |
1051          *     | state | or               | state |
1052          *
1053          * There's a hole, we need to insert something in it and
1054          * ignore the extent we found.
1055          */
1056         if (state->start > start) {
1057                 u64 this_end;
1058                 if (end < last_start)
1059                         this_end = end;
1060                 else
1061                         this_end = last_start - 1;
1062
1063                 prealloc = alloc_extent_state_atomic(prealloc);
1064                 if (!prealloc) {
1065                         err = -ENOMEM;
1066                         goto out;
1067                 }
1068
1069                 /*
1070                  * Avoid to free 'prealloc' if it can be merged with
1071                  * the later extent.
1072                  */
1073                 err = insert_state(tree, prealloc, start, this_end,
1074                                    &bits);
1075                 if (err)
1076                         extent_io_tree_panic(tree, err);
1077                 prealloc = NULL;
1078                 start = this_end + 1;
1079                 goto search_again;
1080         }
1081         /*
1082          * | ---- desired range ---- |
1083          *                        | state |
1084          * We need to split the extent, and set the bit
1085          * on the first half
1086          */
1087         if (state->start <= end && state->end > end) {
1088                 prealloc = alloc_extent_state_atomic(prealloc);
1089                 if (!prealloc) {
1090                         err = -ENOMEM;
1091                         goto out;
1092                 }
1093
1094                 err = split_state(tree, state, prealloc, end + 1);
1095                 if (err)
1096                         extent_io_tree_panic(tree, err);
1097
1098                 set_state_bits(tree, prealloc, &bits);
1099                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1100                 prealloc = NULL;
1101                 goto out;
1102         }
1103
1104         goto search_again;
1105
1106 out:
1107         spin_unlock(&tree->lock);
1108         if (prealloc)
1109                 free_extent_state(prealloc);
1110
1111         return err;
1112
1113 search_again:
1114         if (start > end)
1115                 goto out;
1116         spin_unlock(&tree->lock);
1117         if (mask & __GFP_WAIT)
1118                 cond_resched();
1119         goto again;
1120 }
1121
1122 /* wrappers around set/clear extent bit */
1123 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1124                      gfp_t mask)
1125 {
1126         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1127                               NULL, mask);
1128 }
1129
1130 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1131                     int bits, gfp_t mask)
1132 {
1133         return set_extent_bit(tree, start, end, bits, NULL,
1134                               NULL, mask);
1135 }
1136
1137 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1138                       int bits, gfp_t mask)
1139 {
1140         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1141 }
1142
1143 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1144                         struct extent_state **cached_state, gfp_t mask)
1145 {
1146         return set_extent_bit(tree, start, end,
1147                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1148                               NULL, cached_state, mask);
1149 }
1150
1151 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1152                        gfp_t mask)
1153 {
1154         return clear_extent_bit(tree, start, end,
1155                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1156                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1157 }
1158
1159 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1160                      gfp_t mask)
1161 {
1162         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1163                               NULL, mask);
1164 }
1165
1166 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1167                         struct extent_state **cached_state, gfp_t mask)
1168 {
1169         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
1170                               cached_state, mask);
1171 }
1172
1173 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
1174                                  u64 end, struct extent_state **cached_state,
1175                                  gfp_t mask)
1176 {
1177         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1178                                 cached_state, mask);
1179 }
1180
1181 /*
1182  * either insert or lock state struct between start and end use mask to tell
1183  * us if waiting is desired.
1184  */
1185 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1186                      int bits, struct extent_state **cached_state)
1187 {
1188         int err;
1189         u64 failed_start;
1190         while (1) {
1191                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1192                                        EXTENT_LOCKED, &failed_start,
1193                                        cached_state, GFP_NOFS);
1194                 if (err == -EEXIST) {
1195                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1196                         start = failed_start;
1197                 } else
1198                         break;
1199                 WARN_ON(start > end);
1200         }
1201         return err;
1202 }
1203
1204 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1205 {
1206         return lock_extent_bits(tree, start, end, 0, NULL);
1207 }
1208
1209 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1210 {
1211         int err;
1212         u64 failed_start;
1213
1214         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1215                                &failed_start, NULL, GFP_NOFS);
1216         if (err == -EEXIST) {
1217                 if (failed_start > start)
1218                         clear_extent_bit(tree, start, failed_start - 1,
1219                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1220                 return 0;
1221         }
1222         return 1;
1223 }
1224
1225 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1226                          struct extent_state **cached, gfp_t mask)
1227 {
1228         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1229                                 mask);
1230 }
1231
1232 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1233 {
1234         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1235                                 GFP_NOFS);
1236 }
1237
1238 /*
1239  * helper function to set both pages and extents in the tree writeback
1240  */
1241 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1242 {
1243         unsigned long index = start >> PAGE_CACHE_SHIFT;
1244         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1245         struct page *page;
1246
1247         while (index <= end_index) {
1248                 page = find_get_page(tree->mapping, index);
1249                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1250                 set_page_writeback(page);
1251                 page_cache_release(page);
1252                 index++;
1253         }
1254         return 0;
1255 }
1256
1257 /* find the first state struct with 'bits' set after 'start', and
1258  * return it.  tree->lock must be held.  NULL will returned if
1259  * nothing was found after 'start'
1260  */
1261 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1262                                                  u64 start, int bits)
1263 {
1264         struct rb_node *node;
1265         struct extent_state *state;
1266
1267         /*
1268          * this search will find all the extents that end after
1269          * our range starts.
1270          */
1271         node = tree_search(tree, start);
1272         if (!node)
1273                 goto out;
1274
1275         while (1) {
1276                 state = rb_entry(node, struct extent_state, rb_node);
1277                 if (state->end >= start && (state->state & bits))
1278                         return state;
1279
1280                 node = rb_next(node);
1281                 if (!node)
1282                         break;
1283         }
1284 out:
1285         return NULL;
1286 }
1287
1288 /*
1289  * find the first offset in the io tree with 'bits' set. zero is
1290  * returned if we find something, and *start_ret and *end_ret are
1291  * set to reflect the state struct that was found.
1292  *
1293  * If nothing was found, 1 is returned, < 0 on error
1294  */
1295 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1296                           u64 *start_ret, u64 *end_ret, int bits)
1297 {
1298         struct extent_state *state;
1299         int ret = 1;
1300
1301         spin_lock(&tree->lock);
1302         state = find_first_extent_bit_state(tree, start, bits);
1303         if (state) {
1304                 *start_ret = state->start;
1305                 *end_ret = state->end;
1306                 ret = 0;
1307         }
1308         spin_unlock(&tree->lock);
1309         return ret;
1310 }
1311
1312 /*
1313  * find a contiguous range of bytes in the file marked as delalloc, not
1314  * more than 'max_bytes'.  start and end are used to return the range,
1315  *
1316  * 1 is returned if we find something, 0 if nothing was in the tree
1317  */
1318 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1319                                         u64 *start, u64 *end, u64 max_bytes,
1320                                         struct extent_state **cached_state)
1321 {
1322         struct rb_node *node;
1323         struct extent_state *state;
1324         u64 cur_start = *start;
1325         u64 found = 0;
1326         u64 total_bytes = 0;
1327
1328         spin_lock(&tree->lock);
1329
1330         /*
1331          * this search will find all the extents that end after
1332          * our range starts.
1333          */
1334         node = tree_search(tree, cur_start);
1335         if (!node) {
1336                 if (!found)
1337                         *end = (u64)-1;
1338                 goto out;
1339         }
1340
1341         while (1) {
1342                 state = rb_entry(node, struct extent_state, rb_node);
1343                 if (found && (state->start != cur_start ||
1344                               (state->state & EXTENT_BOUNDARY))) {
1345                         goto out;
1346                 }
1347                 if (!(state->state & EXTENT_DELALLOC)) {
1348                         if (!found)
1349                                 *end = state->end;
1350                         goto out;
1351                 }
1352                 if (!found) {
1353                         *start = state->start;
1354                         *cached_state = state;
1355                         atomic_inc(&state->refs);
1356                 }
1357                 found++;
1358                 *end = state->end;
1359                 cur_start = state->end + 1;
1360                 node = rb_next(node);
1361                 if (!node)
1362                         break;
1363                 total_bytes += state->end - state->start + 1;
1364                 if (total_bytes >= max_bytes)
1365                         break;
1366         }
1367 out:
1368         spin_unlock(&tree->lock);
1369         return found;
1370 }
1371
1372 static noinline void __unlock_for_delalloc(struct inode *inode,
1373                                            struct page *locked_page,
1374                                            u64 start, u64 end)
1375 {
1376         int ret;
1377         struct page *pages[16];
1378         unsigned long index = start >> PAGE_CACHE_SHIFT;
1379         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1380         unsigned long nr_pages = end_index - index + 1;
1381         int i;
1382
1383         if (index == locked_page->index && end_index == index)
1384                 return;
1385
1386         while (nr_pages > 0) {
1387                 ret = find_get_pages_contig(inode->i_mapping, index,
1388                                      min_t(unsigned long, nr_pages,
1389                                      ARRAY_SIZE(pages)), pages);
1390                 for (i = 0; i < ret; i++) {
1391                         if (pages[i] != locked_page)
1392                                 unlock_page(pages[i]);
1393                         page_cache_release(pages[i]);
1394                 }
1395                 nr_pages -= ret;
1396                 index += ret;
1397                 cond_resched();
1398         }
1399 }
1400
1401 static noinline int lock_delalloc_pages(struct inode *inode,
1402                                         struct page *locked_page,
1403                                         u64 delalloc_start,
1404                                         u64 delalloc_end)
1405 {
1406         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1407         unsigned long start_index = index;
1408         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1409         unsigned long pages_locked = 0;
1410         struct page *pages[16];
1411         unsigned long nrpages;
1412         int ret;
1413         int i;
1414
1415         /* the caller is responsible for locking the start index */
1416         if (index == locked_page->index && index == end_index)
1417                 return 0;
1418
1419         /* skip the page at the start index */
1420         nrpages = end_index - index + 1;
1421         while (nrpages > 0) {
1422                 ret = find_get_pages_contig(inode->i_mapping, index,
1423                                      min_t(unsigned long,
1424                                      nrpages, ARRAY_SIZE(pages)), pages);
1425                 if (ret == 0) {
1426                         ret = -EAGAIN;
1427                         goto done;
1428                 }
1429                 /* now we have an array of pages, lock them all */
1430                 for (i = 0; i < ret; i++) {
1431                         /*
1432                          * the caller is taking responsibility for
1433                          * locked_page
1434                          */
1435                         if (pages[i] != locked_page) {
1436                                 lock_page(pages[i]);
1437                                 if (!PageDirty(pages[i]) ||
1438                                     pages[i]->mapping != inode->i_mapping) {
1439                                         ret = -EAGAIN;
1440                                         unlock_page(pages[i]);
1441                                         page_cache_release(pages[i]);
1442                                         goto done;
1443                                 }
1444                         }
1445                         page_cache_release(pages[i]);
1446                         pages_locked++;
1447                 }
1448                 nrpages -= ret;
1449                 index += ret;
1450                 cond_resched();
1451         }
1452         ret = 0;
1453 done:
1454         if (ret && pages_locked) {
1455                 __unlock_for_delalloc(inode, locked_page,
1456                               delalloc_start,
1457                               ((u64)(start_index + pages_locked - 1)) <<
1458                               PAGE_CACHE_SHIFT);
1459         }
1460         return ret;
1461 }
1462
1463 /*
1464  * find a contiguous range of bytes in the file marked as delalloc, not
1465  * more than 'max_bytes'.  start and end are used to return the range,
1466  *
1467  * 1 is returned if we find something, 0 if nothing was in the tree
1468  */
1469 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1470                                              struct extent_io_tree *tree,
1471                                              struct page *locked_page,
1472                                              u64 *start, u64 *end,
1473                                              u64 max_bytes)
1474 {
1475         u64 delalloc_start;
1476         u64 delalloc_end;
1477         u64 found;
1478         struct extent_state *cached_state = NULL;
1479         int ret;
1480         int loops = 0;
1481
1482 again:
1483         /* step one, find a bunch of delalloc bytes starting at start */
1484         delalloc_start = *start;
1485         delalloc_end = 0;
1486         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1487                                     max_bytes, &cached_state);
1488         if (!found || delalloc_end <= *start) {
1489                 *start = delalloc_start;
1490                 *end = delalloc_end;
1491                 free_extent_state(cached_state);
1492                 return found;
1493         }
1494
1495         /*
1496          * start comes from the offset of locked_page.  We have to lock
1497          * pages in order, so we can't process delalloc bytes before
1498          * locked_page
1499          */
1500         if (delalloc_start < *start)
1501                 delalloc_start = *start;
1502
1503         /*
1504          * make sure to limit the number of pages we try to lock down
1505          * if we're looping.
1506          */
1507         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1508                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1509
1510         /* step two, lock all the pages after the page that has start */
1511         ret = lock_delalloc_pages(inode, locked_page,
1512                                   delalloc_start, delalloc_end);
1513         if (ret == -EAGAIN) {
1514                 /* some of the pages are gone, lets avoid looping by
1515                  * shortening the size of the delalloc range we're searching
1516                  */
1517                 free_extent_state(cached_state);
1518                 if (!loops) {
1519                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1520                         max_bytes = PAGE_CACHE_SIZE - offset;
1521                         loops = 1;
1522                         goto again;
1523                 } else {
1524                         found = 0;
1525                         goto out_failed;
1526                 }
1527         }
1528         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1529
1530         /* step three, lock the state bits for the whole range */
1531         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1532
1533         /* then test to make sure it is all still delalloc */
1534         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1535                              EXTENT_DELALLOC, 1, cached_state);
1536         if (!ret) {
1537                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1538                                      &cached_state, GFP_NOFS);
1539                 __unlock_for_delalloc(inode, locked_page,
1540                               delalloc_start, delalloc_end);
1541                 cond_resched();
1542                 goto again;
1543         }
1544         free_extent_state(cached_state);
1545         *start = delalloc_start;
1546         *end = delalloc_end;
1547 out_failed:
1548         return found;
1549 }
1550
1551 int extent_clear_unlock_delalloc(struct inode *inode,
1552                                 struct extent_io_tree *tree,
1553                                 u64 start, u64 end, struct page *locked_page,
1554                                 unsigned long op)
1555 {
1556         int ret;
1557         struct page *pages[16];
1558         unsigned long index = start >> PAGE_CACHE_SHIFT;
1559         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1560         unsigned long nr_pages = end_index - index + 1;
1561         int i;
1562         int clear_bits = 0;
1563
1564         if (op & EXTENT_CLEAR_UNLOCK)
1565                 clear_bits |= EXTENT_LOCKED;
1566         if (op & EXTENT_CLEAR_DIRTY)
1567                 clear_bits |= EXTENT_DIRTY;
1568
1569         if (op & EXTENT_CLEAR_DELALLOC)
1570                 clear_bits |= EXTENT_DELALLOC;
1571
1572         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1573         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1574                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1575                     EXTENT_SET_PRIVATE2)))
1576                 return 0;
1577
1578         while (nr_pages > 0) {
1579                 ret = find_get_pages_contig(inode->i_mapping, index,
1580                                      min_t(unsigned long,
1581                                      nr_pages, ARRAY_SIZE(pages)), pages);
1582                 for (i = 0; i < ret; i++) {
1583
1584                         if (op & EXTENT_SET_PRIVATE2)
1585                                 SetPagePrivate2(pages[i]);
1586
1587                         if (pages[i] == locked_page) {
1588                                 page_cache_release(pages[i]);
1589                                 continue;
1590                         }
1591                         if (op & EXTENT_CLEAR_DIRTY)
1592                                 clear_page_dirty_for_io(pages[i]);
1593                         if (op & EXTENT_SET_WRITEBACK)
1594                                 set_page_writeback(pages[i]);
1595                         if (op & EXTENT_END_WRITEBACK)
1596                                 end_page_writeback(pages[i]);
1597                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1598                                 unlock_page(pages[i]);
1599                         page_cache_release(pages[i]);
1600                 }
1601                 nr_pages -= ret;
1602                 index += ret;
1603                 cond_resched();
1604         }
1605         return 0;
1606 }
1607
1608 /*
1609  * count the number of bytes in the tree that have a given bit(s)
1610  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1611  * cached.  The total number found is returned.
1612  */
1613 u64 count_range_bits(struct extent_io_tree *tree,
1614                      u64 *start, u64 search_end, u64 max_bytes,
1615                      unsigned long bits, int contig)
1616 {
1617         struct rb_node *node;
1618         struct extent_state *state;
1619         u64 cur_start = *start;
1620         u64 total_bytes = 0;
1621         u64 last = 0;
1622         int found = 0;
1623
1624         if (search_end <= cur_start) {
1625                 WARN_ON(1);
1626                 return 0;
1627         }
1628
1629         spin_lock(&tree->lock);
1630         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1631                 total_bytes = tree->dirty_bytes;
1632                 goto out;
1633         }
1634         /*
1635          * this search will find all the extents that end after
1636          * our range starts.
1637          */
1638         node = tree_search(tree, cur_start);
1639         if (!node)
1640                 goto out;
1641
1642         while (1) {
1643                 state = rb_entry(node, struct extent_state, rb_node);
1644                 if (state->start > search_end)
1645                         break;
1646                 if (contig && found && state->start > last + 1)
1647                         break;
1648                 if (state->end >= cur_start && (state->state & bits) == bits) {
1649                         total_bytes += min(search_end, state->end) + 1 -
1650                                        max(cur_start, state->start);
1651                         if (total_bytes >= max_bytes)
1652                                 break;
1653                         if (!found) {
1654                                 *start = max(cur_start, state->start);
1655                                 found = 1;
1656                         }
1657                         last = state->end;
1658                 } else if (contig && found) {
1659                         break;
1660                 }
1661                 node = rb_next(node);
1662                 if (!node)
1663                         break;
1664         }
1665 out:
1666         spin_unlock(&tree->lock);
1667         return total_bytes;
1668 }
1669
1670 /*
1671  * set the private field for a given byte offset in the tree.  If there isn't
1672  * an extent_state there already, this does nothing.
1673  */
1674 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1675 {
1676         struct rb_node *node;
1677         struct extent_state *state;
1678         int ret = 0;
1679
1680         spin_lock(&tree->lock);
1681         /*
1682          * this search will find all the extents that end after
1683          * our range starts.
1684          */
1685         node = tree_search(tree, start);
1686         if (!node) {
1687                 ret = -ENOENT;
1688                 goto out;
1689         }
1690         state = rb_entry(node, struct extent_state, rb_node);
1691         if (state->start != start) {
1692                 ret = -ENOENT;
1693                 goto out;
1694         }
1695         state->private = private;
1696 out:
1697         spin_unlock(&tree->lock);
1698         return ret;
1699 }
1700
1701 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1702 {
1703         struct rb_node *node;
1704         struct extent_state *state;
1705         int ret = 0;
1706
1707         spin_lock(&tree->lock);
1708         /*
1709          * this search will find all the extents that end after
1710          * our range starts.
1711          */
1712         node = tree_search(tree, start);
1713         if (!node) {
1714                 ret = -ENOENT;
1715                 goto out;
1716         }
1717         state = rb_entry(node, struct extent_state, rb_node);
1718         if (state->start != start) {
1719                 ret = -ENOENT;
1720                 goto out;
1721         }
1722         *private = state->private;
1723 out:
1724         spin_unlock(&tree->lock);
1725         return ret;
1726 }
1727
1728 /*
1729  * searches a range in the state tree for a given mask.
1730  * If 'filled' == 1, this returns 1 only if every extent in the tree
1731  * has the bits set.  Otherwise, 1 is returned if any bit in the
1732  * range is found set.
1733  */
1734 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1735                    int bits, int filled, struct extent_state *cached)
1736 {
1737         struct extent_state *state = NULL;
1738         struct rb_node *node;
1739         int bitset = 0;
1740
1741         spin_lock(&tree->lock);
1742         if (cached && cached->tree && cached->start <= start &&
1743             cached->end > start)
1744                 node = &cached->rb_node;
1745         else
1746                 node = tree_search(tree, start);
1747         while (node && start <= end) {
1748                 state = rb_entry(node, struct extent_state, rb_node);
1749
1750                 if (filled && state->start > start) {
1751                         bitset = 0;
1752                         break;
1753                 }
1754
1755                 if (state->start > end)
1756                         break;
1757
1758                 if (state->state & bits) {
1759                         bitset = 1;
1760                         if (!filled)
1761                                 break;
1762                 } else if (filled) {
1763                         bitset = 0;
1764                         break;
1765                 }
1766
1767                 if (state->end == (u64)-1)
1768                         break;
1769
1770                 start = state->end + 1;
1771                 if (start > end)
1772                         break;
1773                 node = rb_next(node);
1774                 if (!node) {
1775                         if (filled)
1776                                 bitset = 0;
1777                         break;
1778                 }
1779         }
1780         spin_unlock(&tree->lock);
1781         return bitset;
1782 }
1783
1784 /*
1785  * helper function to set a given page up to date if all the
1786  * extents in the tree for that page are up to date
1787  */
1788 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1789 {
1790         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1791         u64 end = start + PAGE_CACHE_SIZE - 1;
1792         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1793                 SetPageUptodate(page);
1794 }
1795
1796 /*
1797  * helper function to unlock a page if all the extents in the tree
1798  * for that page are unlocked
1799  */
1800 static void check_page_locked(struct extent_io_tree *tree, struct page *page)
1801 {
1802         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1803         u64 end = start + PAGE_CACHE_SIZE - 1;
1804         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1805                 unlock_page(page);
1806 }
1807
1808 /*
1809  * helper function to end page writeback if all the extents
1810  * in the tree for that page are done with writeback
1811  */
1812 static void check_page_writeback(struct extent_io_tree *tree,
1813                                  struct page *page)
1814 {
1815         end_page_writeback(page);
1816 }
1817
1818 /*
1819  * When IO fails, either with EIO or csum verification fails, we
1820  * try other mirrors that might have a good copy of the data.  This
1821  * io_failure_record is used to record state as we go through all the
1822  * mirrors.  If another mirror has good data, the page is set up to date
1823  * and things continue.  If a good mirror can't be found, the original
1824  * bio end_io callback is called to indicate things have failed.
1825  */
1826 struct io_failure_record {
1827         struct page *page;
1828         u64 start;
1829         u64 len;
1830         u64 logical;
1831         unsigned long bio_flags;
1832         int this_mirror;
1833         int failed_mirror;
1834         int in_validation;
1835 };
1836
1837 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1838                                 int did_repair)
1839 {
1840         int ret;
1841         int err = 0;
1842         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1843
1844         set_state_private(failure_tree, rec->start, 0);
1845         ret = clear_extent_bits(failure_tree, rec->start,
1846                                 rec->start + rec->len - 1,
1847                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1848         if (ret)
1849                 err = ret;
1850
1851         if (did_repair) {
1852                 ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1853                                         rec->start + rec->len - 1,
1854                                         EXTENT_DAMAGED, GFP_NOFS);
1855                 if (ret && !err)
1856                         err = ret;
1857         }
1858
1859         kfree(rec);
1860         return err;
1861 }
1862
1863 static void repair_io_failure_callback(struct bio *bio, int err)
1864 {
1865         complete(bio->bi_private);
1866 }
1867
1868 /*
1869  * this bypasses the standard btrfs submit functions deliberately, as
1870  * the standard behavior is to write all copies in a raid setup. here we only
1871  * want to write the one bad copy. so we do the mapping for ourselves and issue
1872  * submit_bio directly.
1873  * to avoid any synchonization issues, wait for the data after writing, which
1874  * actually prevents the read that triggered the error from finishing.
1875  * currently, there can be no more than two copies of every data bit. thus,
1876  * exactly one rewrite is required.
1877  */
1878 int repair_io_failure(struct btrfs_mapping_tree *map_tree, u64 start,
1879                         u64 length, u64 logical, struct page *page,
1880                         int mirror_num)
1881 {
1882         struct bio *bio;
1883         struct btrfs_device *dev;
1884         DECLARE_COMPLETION_ONSTACK(compl);
1885         u64 map_length = 0;
1886         u64 sector;
1887         struct btrfs_bio *bbio = NULL;
1888         int ret;
1889
1890         BUG_ON(!mirror_num);
1891
1892         bio = bio_alloc(GFP_NOFS, 1);
1893         if (!bio)
1894                 return -EIO;
1895         bio->bi_private = &compl;
1896         bio->bi_end_io = repair_io_failure_callback;
1897         bio->bi_size = 0;
1898         map_length = length;
1899
1900         ret = btrfs_map_block(map_tree, WRITE, logical,
1901                               &map_length, &bbio, mirror_num);
1902         if (ret) {
1903                 bio_put(bio);
1904                 return -EIO;
1905         }
1906         BUG_ON(mirror_num != bbio->mirror_num);
1907         sector = bbio->stripes[mirror_num-1].physical >> 9;
1908         bio->bi_sector = sector;
1909         dev = bbio->stripes[mirror_num-1].dev;
1910         kfree(bbio);
1911         if (!dev || !dev->bdev || !dev->writeable) {
1912                 bio_put(bio);
1913                 return -EIO;
1914         }
1915         bio->bi_bdev = dev->bdev;
1916         bio_add_page(bio, page, length, start-page_offset(page));
1917         btrfsic_submit_bio(WRITE_SYNC, bio);
1918         wait_for_completion(&compl);
1919
1920         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
1921                 /* try to remap that extent elsewhere? */
1922                 bio_put(bio);
1923                 return -EIO;
1924         }
1925
1926         printk(KERN_INFO "btrfs read error corrected: ino %lu off %llu (dev %s "
1927                         "sector %llu)\n", page->mapping->host->i_ino, start,
1928                         dev->name, sector);
1929
1930         bio_put(bio);
1931         return 0;
1932 }
1933
1934 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
1935                          int mirror_num)
1936 {
1937         struct btrfs_mapping_tree *map_tree = &root->fs_info->mapping_tree;
1938         u64 start = eb->start;
1939         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
1940         int ret;
1941
1942         for (i = 0; i < num_pages; i++) {
1943                 struct page *p = extent_buffer_page(eb, i);
1944                 ret = repair_io_failure(map_tree, start, PAGE_CACHE_SIZE,
1945                                         start, p, mirror_num);
1946                 if (ret)
1947                         break;
1948                 start += PAGE_CACHE_SIZE;
1949         }
1950
1951         return ret;
1952 }
1953
1954 /*
1955  * each time an IO finishes, we do a fast check in the IO failure tree
1956  * to see if we need to process or clean up an io_failure_record
1957  */
1958 static int clean_io_failure(u64 start, struct page *page)
1959 {
1960         u64 private;
1961         u64 private_failure;
1962         struct io_failure_record *failrec;
1963         struct btrfs_mapping_tree *map_tree;
1964         struct extent_state *state;
1965         int num_copies;
1966         int did_repair = 0;
1967         int ret;
1968         struct inode *inode = page->mapping->host;
1969
1970         private = 0;
1971         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
1972                                 (u64)-1, 1, EXTENT_DIRTY, 0);
1973         if (!ret)
1974                 return 0;
1975
1976         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
1977                                 &private_failure);
1978         if (ret)
1979                 return 0;
1980
1981         failrec = (struct io_failure_record *)(unsigned long) private_failure;
1982         BUG_ON(!failrec->this_mirror);
1983
1984         if (failrec->in_validation) {
1985                 /* there was no real error, just free the record */
1986                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
1987                          failrec->start);
1988                 did_repair = 1;
1989                 goto out;
1990         }
1991
1992         spin_lock(&BTRFS_I(inode)->io_tree.lock);
1993         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
1994                                             failrec->start,
1995                                             EXTENT_LOCKED);
1996         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
1997
1998         if (state && state->start == failrec->start) {
1999                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
2000                 num_copies = btrfs_num_copies(map_tree, failrec->logical,
2001                                                 failrec->len);
2002                 if (num_copies > 1)  {
2003                         ret = repair_io_failure(map_tree, start, failrec->len,
2004                                                 failrec->logical, page,
2005                                                 failrec->failed_mirror);
2006                         did_repair = !ret;
2007                 }
2008         }
2009
2010 out:
2011         if (!ret)
2012                 ret = free_io_failure(inode, failrec, did_repair);
2013
2014         return ret;
2015 }
2016
2017 /*
2018  * this is a generic handler for readpage errors (default
2019  * readpage_io_failed_hook). if other copies exist, read those and write back
2020  * good data to the failed position. does not investigate in remapping the
2021  * failed extent elsewhere, hoping the device will be smart enough to do this as
2022  * needed
2023  */
2024
2025 static int bio_readpage_error(struct bio *failed_bio, struct page *page,
2026                                 u64 start, u64 end, int failed_mirror,
2027                                 struct extent_state *state)
2028 {
2029         struct io_failure_record *failrec = NULL;
2030         u64 private;
2031         struct extent_map *em;
2032         struct inode *inode = page->mapping->host;
2033         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2034         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2035         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2036         struct bio *bio;
2037         int num_copies;
2038         int ret;
2039         int read_mode;
2040         u64 logical;
2041
2042         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2043
2044         ret = get_state_private(failure_tree, start, &private);
2045         if (ret) {
2046                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2047                 if (!failrec)
2048                         return -ENOMEM;
2049                 failrec->start = start;
2050                 failrec->len = end - start + 1;
2051                 failrec->this_mirror = 0;
2052                 failrec->bio_flags = 0;
2053                 failrec->in_validation = 0;
2054
2055                 read_lock(&em_tree->lock);
2056                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2057                 if (!em) {
2058                         read_unlock(&em_tree->lock);
2059                         kfree(failrec);
2060                         return -EIO;
2061                 }
2062
2063                 if (em->start > start || em->start + em->len < start) {
2064                         free_extent_map(em);
2065                         em = NULL;
2066                 }
2067                 read_unlock(&em_tree->lock);
2068
2069                 if (!em || IS_ERR(em)) {
2070                         kfree(failrec);
2071                         return -EIO;
2072                 }
2073                 logical = start - em->start;
2074                 logical = em->block_start + logical;
2075                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2076                         logical = em->block_start;
2077                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2078                         extent_set_compress_type(&failrec->bio_flags,
2079                                                  em->compress_type);
2080                 }
2081                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2082                          "len=%llu\n", logical, start, failrec->len);
2083                 failrec->logical = logical;
2084                 free_extent_map(em);
2085
2086                 /* set the bits in the private failure tree */
2087                 ret = set_extent_bits(failure_tree, start, end,
2088                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2089                 if (ret >= 0)
2090                         ret = set_state_private(failure_tree, start,
2091                                                 (u64)(unsigned long)failrec);
2092                 /* set the bits in the inode's tree */
2093                 if (ret >= 0)
2094                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2095                                                 GFP_NOFS);
2096                 if (ret < 0) {
2097                         kfree(failrec);
2098                         return ret;
2099                 }
2100         } else {
2101                 failrec = (struct io_failure_record *)(unsigned long)private;
2102                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2103                          "start=%llu, len=%llu, validation=%d\n",
2104                          failrec->logical, failrec->start, failrec->len,
2105                          failrec->in_validation);
2106                 /*
2107                  * when data can be on disk more than twice, add to failrec here
2108                  * (e.g. with a list for failed_mirror) to make
2109                  * clean_io_failure() clean all those errors at once.
2110                  */
2111         }
2112         num_copies = btrfs_num_copies(
2113                               &BTRFS_I(inode)->root->fs_info->mapping_tree,
2114                               failrec->logical, failrec->len);
2115         if (num_copies == 1) {
2116                 /*
2117                  * we only have a single copy of the data, so don't bother with
2118                  * all the retry and error correction code that follows. no
2119                  * matter what the error is, it is very likely to persist.
2120                  */
2121                 pr_debug("bio_readpage_error: cannot repair, num_copies == 1. "
2122                          "state=%p, num_copies=%d, next_mirror %d, "
2123                          "failed_mirror %d\n", state, num_copies,
2124                          failrec->this_mirror, failed_mirror);
2125                 free_io_failure(inode, failrec, 0);
2126                 return -EIO;
2127         }
2128
2129         if (!state) {
2130                 spin_lock(&tree->lock);
2131                 state = find_first_extent_bit_state(tree, failrec->start,
2132                                                     EXTENT_LOCKED);
2133                 if (state && state->start != failrec->start)
2134                         state = NULL;
2135                 spin_unlock(&tree->lock);
2136         }
2137
2138         /*
2139          * there are two premises:
2140          *      a) deliver good data to the caller
2141          *      b) correct the bad sectors on disk
2142          */
2143         if (failed_bio->bi_vcnt > 1) {
2144                 /*
2145                  * to fulfill b), we need to know the exact failing sectors, as
2146                  * we don't want to rewrite any more than the failed ones. thus,
2147                  * we need separate read requests for the failed bio
2148                  *
2149                  * if the following BUG_ON triggers, our validation request got
2150                  * merged. we need separate requests for our algorithm to work.
2151                  */
2152                 BUG_ON(failrec->in_validation);
2153                 failrec->in_validation = 1;
2154                 failrec->this_mirror = failed_mirror;
2155                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2156         } else {
2157                 /*
2158                  * we're ready to fulfill a) and b) alongside. get a good copy
2159                  * of the failed sector and if we succeed, we have setup
2160                  * everything for repair_io_failure to do the rest for us.
2161                  */
2162                 if (failrec->in_validation) {
2163                         BUG_ON(failrec->this_mirror != failed_mirror);
2164                         failrec->in_validation = 0;
2165                         failrec->this_mirror = 0;
2166                 }
2167                 failrec->failed_mirror = failed_mirror;
2168                 failrec->this_mirror++;
2169                 if (failrec->this_mirror == failed_mirror)
2170                         failrec->this_mirror++;
2171                 read_mode = READ_SYNC;
2172         }
2173
2174         if (!state || failrec->this_mirror > num_copies) {
2175                 pr_debug("bio_readpage_error: (fail) state=%p, num_copies=%d, "
2176                          "next_mirror %d, failed_mirror %d\n", state,
2177                          num_copies, failrec->this_mirror, failed_mirror);
2178                 free_io_failure(inode, failrec, 0);
2179                 return -EIO;
2180         }
2181
2182         bio = bio_alloc(GFP_NOFS, 1);
2183         bio->bi_private = state;
2184         bio->bi_end_io = failed_bio->bi_end_io;
2185         bio->bi_sector = failrec->logical >> 9;
2186         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2187         bio->bi_size = 0;
2188
2189         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2190
2191         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2192                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2193                  failrec->this_mirror, num_copies, failrec->in_validation);
2194
2195         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2196                                          failrec->this_mirror,
2197                                          failrec->bio_flags, 0);
2198         return ret;
2199 }
2200
2201 /* lots and lots of room for performance fixes in the end_bio funcs */
2202
2203 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2204 {
2205         int uptodate = (err == 0);
2206         struct extent_io_tree *tree;
2207         int ret;
2208
2209         tree = &BTRFS_I(page->mapping->host)->io_tree;
2210
2211         if (tree->ops && tree->ops->writepage_end_io_hook) {
2212                 ret = tree->ops->writepage_end_io_hook(page, start,
2213                                                end, NULL, uptodate);
2214                 if (ret)
2215                         uptodate = 0;
2216         }
2217
2218         if (!uptodate && tree->ops &&
2219             tree->ops->writepage_io_failed_hook) {
2220                 ret = tree->ops->writepage_io_failed_hook(NULL, page,
2221                                                  start, end, NULL);
2222                 /* Writeback already completed */
2223                 if (ret == 0)
2224                         return 1;
2225         }
2226
2227         if (!uptodate) {
2228                 clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
2229                 ClearPageUptodate(page);
2230                 SetPageError(page);
2231         }
2232         return 0;
2233 }
2234
2235 /*
2236  * after a writepage IO is done, we need to:
2237  * clear the uptodate bits on error
2238  * clear the writeback bits in the extent tree for this IO
2239  * end_page_writeback if the page has no more pending IO
2240  *
2241  * Scheduling is not allowed, so the extent state tree is expected
2242  * to have one and only one object corresponding to this IO.
2243  */
2244 static void end_bio_extent_writepage(struct bio *bio, int err)
2245 {
2246         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2247         struct extent_io_tree *tree;
2248         u64 start;
2249         u64 end;
2250         int whole_page;
2251
2252         do {
2253                 struct page *page = bvec->bv_page;
2254                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2255
2256                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2257                          bvec->bv_offset;
2258                 end = start + bvec->bv_len - 1;
2259
2260                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2261                         whole_page = 1;
2262                 else
2263                         whole_page = 0;
2264
2265                 if (--bvec >= bio->bi_io_vec)
2266                         prefetchw(&bvec->bv_page->flags);
2267
2268                 if (end_extent_writepage(page, err, start, end))
2269                         continue;
2270
2271                 if (whole_page)
2272                         end_page_writeback(page);
2273                 else
2274                         check_page_writeback(tree, page);
2275         } while (bvec >= bio->bi_io_vec);
2276
2277         bio_put(bio);
2278 }
2279
2280 /*
2281  * after a readpage IO is done, we need to:
2282  * clear the uptodate bits on error
2283  * set the uptodate bits if things worked
2284  * set the page up to date if all extents in the tree are uptodate
2285  * clear the lock bit in the extent tree
2286  * unlock the page if there are no other extents locked for it
2287  *
2288  * Scheduling is not allowed, so the extent state tree is expected
2289  * to have one and only one object corresponding to this IO.
2290  */
2291 static void end_bio_extent_readpage(struct bio *bio, int err)
2292 {
2293         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2294         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2295         struct bio_vec *bvec = bio->bi_io_vec;
2296         struct extent_io_tree *tree;
2297         u64 start;
2298         u64 end;
2299         int whole_page;
2300         int failed_mirror;
2301         int ret;
2302
2303         if (err)
2304                 uptodate = 0;
2305
2306         do {
2307                 struct page *page = bvec->bv_page;
2308                 struct extent_state *cached = NULL;
2309                 struct extent_state *state;
2310
2311                 pr_debug("end_bio_extent_readpage: bi_vcnt=%d, idx=%d, err=%d, "
2312                          "mirror=%ld\n", bio->bi_vcnt, bio->bi_idx, err,
2313                          (long int)bio->bi_bdev);
2314                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2315
2316                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2317                         bvec->bv_offset;
2318                 end = start + bvec->bv_len - 1;
2319
2320                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2321                         whole_page = 1;
2322                 else
2323                         whole_page = 0;
2324
2325                 if (++bvec <= bvec_end)
2326                         prefetchw(&bvec->bv_page->flags);
2327
2328                 spin_lock(&tree->lock);
2329                 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
2330                 if (state && state->start == start) {
2331                         /*
2332                          * take a reference on the state, unlock will drop
2333                          * the ref
2334                          */
2335                         cache_state(state, &cached);
2336                 }
2337                 spin_unlock(&tree->lock);
2338
2339                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
2340                         ret = tree->ops->readpage_end_io_hook(page, start, end,
2341                                                               state);
2342                         if (ret)
2343                                 uptodate = 0;
2344                         else
2345                                 clean_io_failure(start, page);
2346                 }
2347
2348                 if (!uptodate)
2349                         failed_mirror = (int)(unsigned long)bio->bi_bdev;
2350
2351                 if (!uptodate && tree->ops && tree->ops->readpage_io_failed_hook) {
2352                         ret = tree->ops->readpage_io_failed_hook(page, failed_mirror);
2353                         if (!ret && !err &&
2354                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2355                                 uptodate = 1;
2356                 } else if (!uptodate) {
2357                         /*
2358                          * The generic bio_readpage_error handles errors the
2359                          * following way: If possible, new read requests are
2360                          * created and submitted and will end up in
2361                          * end_bio_extent_readpage as well (if we're lucky, not
2362                          * in the !uptodate case). In that case it returns 0 and
2363                          * we just go on with the next page in our bio. If it
2364                          * can't handle the error it will return -EIO and we
2365                          * remain responsible for that page.
2366                          */
2367                         ret = bio_readpage_error(bio, page, start, end,
2368                                                         failed_mirror, NULL);
2369                         if (ret == 0) {
2370                                 uptodate =
2371                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2372                                 if (err)
2373                                         uptodate = 0;
2374                                 uncache_state(&cached);
2375                                 continue;
2376                         }
2377                 }
2378
2379                 if (uptodate && tree->track_uptodate) {
2380                         set_extent_uptodate(tree, start, end, &cached,
2381                                             GFP_ATOMIC);
2382                 }
2383                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2384
2385                 if (whole_page) {
2386                         if (uptodate) {
2387                                 SetPageUptodate(page);
2388                         } else {
2389                                 ClearPageUptodate(page);
2390                                 SetPageError(page);
2391                         }
2392                         unlock_page(page);
2393                 } else {
2394                         if (uptodate) {
2395                                 check_page_uptodate(tree, page);
2396                         } else {
2397                                 ClearPageUptodate(page);
2398                                 SetPageError(page);
2399                         }
2400                         check_page_locked(tree, page);
2401                 }
2402         } while (bvec <= bvec_end);
2403
2404         bio_put(bio);
2405 }
2406
2407 struct bio *
2408 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2409                 gfp_t gfp_flags)
2410 {
2411         struct bio *bio;
2412
2413         bio = bio_alloc(gfp_flags, nr_vecs);
2414
2415         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2416                 while (!bio && (nr_vecs /= 2))
2417                         bio = bio_alloc(gfp_flags, nr_vecs);
2418         }
2419
2420         if (bio) {
2421                 bio->bi_size = 0;
2422                 bio->bi_bdev = bdev;
2423                 bio->bi_sector = first_sector;
2424         }
2425         return bio;
2426 }
2427
2428 /*
2429  * Since writes are async, they will only return -ENOMEM.
2430  * Reads can return the full range of I/O error conditions.
2431  */
2432 static int __must_check submit_one_bio(int rw, struct bio *bio,
2433                                        int mirror_num, unsigned long bio_flags)
2434 {
2435         int ret = 0;
2436         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2437         struct page *page = bvec->bv_page;
2438         struct extent_io_tree *tree = bio->bi_private;
2439         u64 start;
2440
2441         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
2442
2443         bio->bi_private = NULL;
2444
2445         bio_get(bio);
2446
2447         if (tree->ops && tree->ops->submit_bio_hook)
2448                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2449                                            mirror_num, bio_flags, start);
2450         else
2451                 btrfsic_submit_bio(rw, bio);
2452
2453         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2454                 ret = -EOPNOTSUPP;
2455         bio_put(bio);
2456         return ret;
2457 }
2458
2459 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2460                      unsigned long offset, size_t size, struct bio *bio,
2461                      unsigned long bio_flags)
2462 {
2463         int ret = 0;
2464         if (tree->ops && tree->ops->merge_bio_hook)
2465                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2466                                                 bio_flags);
2467         BUG_ON(ret < 0);
2468         return ret;
2469
2470 }
2471
2472 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2473                               struct page *page, sector_t sector,
2474                               size_t size, unsigned long offset,
2475                               struct block_device *bdev,
2476                               struct bio **bio_ret,
2477                               unsigned long max_pages,
2478                               bio_end_io_t end_io_func,
2479                               int mirror_num,
2480                               unsigned long prev_bio_flags,
2481                               unsigned long bio_flags)
2482 {
2483         int ret = 0;
2484         struct bio *bio;
2485         int nr;
2486         int contig = 0;
2487         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2488         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2489         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2490
2491         if (bio_ret && *bio_ret) {
2492                 bio = *bio_ret;
2493                 if (old_compressed)
2494                         contig = bio->bi_sector == sector;
2495                 else
2496                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
2497                                 sector;
2498
2499                 if (prev_bio_flags != bio_flags || !contig ||
2500                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2501                     bio_add_page(bio, page, page_size, offset) < page_size) {
2502                         ret = submit_one_bio(rw, bio, mirror_num,
2503                                              prev_bio_flags);
2504                         if (ret < 0)
2505                                 return ret;
2506                         bio = NULL;
2507                 } else {
2508                         return 0;
2509                 }
2510         }
2511         if (this_compressed)
2512                 nr = BIO_MAX_PAGES;
2513         else
2514                 nr = bio_get_nr_vecs(bdev);
2515
2516         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2517         if (!bio)
2518                 return -ENOMEM;
2519
2520         bio_add_page(bio, page, page_size, offset);
2521         bio->bi_end_io = end_io_func;
2522         bio->bi_private = tree;
2523
2524         if (bio_ret)
2525                 *bio_ret = bio;
2526         else
2527                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2528
2529         return ret;
2530 }
2531
2532 void attach_extent_buffer_page(struct extent_buffer *eb, struct page *page)
2533 {
2534         if (!PagePrivate(page)) {
2535                 SetPagePrivate(page);
2536                 page_cache_get(page);
2537                 set_page_private(page, (unsigned long)eb);
2538         } else {
2539                 WARN_ON(page->private != (unsigned long)eb);
2540         }
2541 }
2542
2543 void set_page_extent_mapped(struct page *page)
2544 {
2545         if (!PagePrivate(page)) {
2546                 SetPagePrivate(page);
2547                 page_cache_get(page);
2548                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2549         }
2550 }
2551
2552 /*
2553  * basic readpage implementation.  Locked extent state structs are inserted
2554  * into the tree that are removed when the IO is done (by the end_io
2555  * handlers)
2556  * XXX JDM: This needs looking at to ensure proper page locking
2557  */
2558 static int __extent_read_full_page(struct extent_io_tree *tree,
2559                                    struct page *page,
2560                                    get_extent_t *get_extent,
2561                                    struct bio **bio, int mirror_num,
2562                                    unsigned long *bio_flags)
2563 {
2564         struct inode *inode = page->mapping->host;
2565         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2566         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2567         u64 end;
2568         u64 cur = start;
2569         u64 extent_offset;
2570         u64 last_byte = i_size_read(inode);
2571         u64 block_start;
2572         u64 cur_end;
2573         sector_t sector;
2574         struct extent_map *em;
2575         struct block_device *bdev;
2576         struct btrfs_ordered_extent *ordered;
2577         int ret;
2578         int nr = 0;
2579         size_t pg_offset = 0;
2580         size_t iosize;
2581         size_t disk_io_size;
2582         size_t blocksize = inode->i_sb->s_blocksize;
2583         unsigned long this_bio_flag = 0;
2584
2585         set_page_extent_mapped(page);
2586
2587         if (!PageUptodate(page)) {
2588                 if (cleancache_get_page(page) == 0) {
2589                         BUG_ON(blocksize != PAGE_SIZE);
2590                         goto out;
2591                 }
2592         }
2593
2594         end = page_end;
2595         while (1) {
2596                 lock_extent(tree, start, end);
2597                 ordered = btrfs_lookup_ordered_extent(inode, start);
2598                 if (!ordered)
2599                         break;
2600                 unlock_extent(tree, start, end);
2601                 btrfs_start_ordered_extent(inode, ordered, 1);
2602                 btrfs_put_ordered_extent(ordered);
2603         }
2604
2605         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2606                 char *userpage;
2607                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2608
2609                 if (zero_offset) {
2610                         iosize = PAGE_CACHE_SIZE - zero_offset;
2611                         userpage = kmap_atomic(page);
2612                         memset(userpage + zero_offset, 0, iosize);
2613                         flush_dcache_page(page);
2614                         kunmap_atomic(userpage);
2615                 }
2616         }
2617         while (cur <= end) {
2618                 if (cur >= last_byte) {
2619                         char *userpage;
2620                         struct extent_state *cached = NULL;
2621
2622                         iosize = PAGE_CACHE_SIZE - pg_offset;
2623                         userpage = kmap_atomic(page);
2624                         memset(userpage + pg_offset, 0, iosize);
2625                         flush_dcache_page(page);
2626                         kunmap_atomic(userpage);
2627                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2628                                             &cached, GFP_NOFS);
2629                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2630                                              &cached, GFP_NOFS);
2631                         break;
2632                 }
2633                 em = get_extent(inode, page, pg_offset, cur,
2634                                 end - cur + 1, 0);
2635                 if (IS_ERR_OR_NULL(em)) {
2636                         SetPageError(page);
2637                         unlock_extent(tree, cur, end);
2638                         break;
2639                 }
2640                 extent_offset = cur - em->start;
2641                 BUG_ON(extent_map_end(em) <= cur);
2642                 BUG_ON(end < cur);
2643
2644                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2645                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2646                         extent_set_compress_type(&this_bio_flag,
2647                                                  em->compress_type);
2648                 }
2649
2650                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2651                 cur_end = min(extent_map_end(em) - 1, end);
2652                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2653                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2654                         disk_io_size = em->block_len;
2655                         sector = em->block_start >> 9;
2656                 } else {
2657                         sector = (em->block_start + extent_offset) >> 9;
2658                         disk_io_size = iosize;
2659                 }
2660                 bdev = em->bdev;
2661                 block_start = em->block_start;
2662                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2663                         block_start = EXTENT_MAP_HOLE;
2664                 free_extent_map(em);
2665                 em = NULL;
2666
2667                 /* we've found a hole, just zero and go on */
2668                 if (block_start == EXTENT_MAP_HOLE) {
2669                         char *userpage;
2670                         struct extent_state *cached = NULL;
2671
2672                         userpage = kmap_atomic(page);
2673                         memset(userpage + pg_offset, 0, iosize);
2674                         flush_dcache_page(page);
2675                         kunmap_atomic(userpage);
2676
2677                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2678                                             &cached, GFP_NOFS);
2679                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2680                                              &cached, GFP_NOFS);
2681                         cur = cur + iosize;
2682                         pg_offset += iosize;
2683                         continue;
2684                 }
2685                 /* the get_extent function already copied into the page */
2686                 if (test_range_bit(tree, cur, cur_end,
2687                                    EXTENT_UPTODATE, 1, NULL)) {
2688                         check_page_uptodate(tree, page);
2689                         unlock_extent(tree, cur, cur + iosize - 1);
2690                         cur = cur + iosize;
2691                         pg_offset += iosize;
2692                         continue;
2693                 }
2694                 /* we have an inline extent but it didn't get marked up
2695                  * to date.  Error out
2696                  */
2697                 if (block_start == EXTENT_MAP_INLINE) {
2698                         SetPageError(page);
2699                         unlock_extent(tree, cur, cur + iosize - 1);
2700                         cur = cur + iosize;
2701                         pg_offset += iosize;
2702                         continue;
2703                 }
2704
2705                 ret = 0;
2706                 if (tree->ops && tree->ops->readpage_io_hook) {
2707                         ret = tree->ops->readpage_io_hook(page, cur,
2708                                                           cur + iosize - 1);
2709                 }
2710                 if (!ret) {
2711                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2712                         pnr -= page->index;
2713                         ret = submit_extent_page(READ, tree, page,
2714                                          sector, disk_io_size, pg_offset,
2715                                          bdev, bio, pnr,
2716                                          end_bio_extent_readpage, mirror_num,
2717                                          *bio_flags,
2718                                          this_bio_flag);
2719                         BUG_ON(ret == -ENOMEM);
2720                         nr++;
2721                         *bio_flags = this_bio_flag;
2722                 }
2723                 if (ret)
2724                         SetPageError(page);
2725                 cur = cur + iosize;
2726                 pg_offset += iosize;
2727         }
2728 out:
2729         if (!nr) {
2730                 if (!PageError(page))
2731                         SetPageUptodate(page);
2732                 unlock_page(page);
2733         }
2734         return 0;
2735 }
2736
2737 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2738                             get_extent_t *get_extent, int mirror_num)
2739 {
2740         struct bio *bio = NULL;
2741         unsigned long bio_flags = 0;
2742         int ret;
2743
2744         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
2745                                       &bio_flags);
2746         if (bio)
2747                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
2748         return ret;
2749 }
2750
2751 static noinline void update_nr_written(struct page *page,
2752                                       struct writeback_control *wbc,
2753                                       unsigned long nr_written)
2754 {
2755         wbc->nr_to_write -= nr_written;
2756         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2757             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2758                 page->mapping->writeback_index = page->index + nr_written;
2759 }
2760
2761 /*
2762  * the writepage semantics are similar to regular writepage.  extent
2763  * records are inserted to lock ranges in the tree, and as dirty areas
2764  * are found, they are marked writeback.  Then the lock bits are removed
2765  * and the end_io handler clears the writeback ranges
2766  */
2767 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2768                               void *data)
2769 {
2770         struct inode *inode = page->mapping->host;
2771         struct extent_page_data *epd = data;
2772         struct extent_io_tree *tree = epd->tree;
2773         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2774         u64 delalloc_start;
2775         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2776         u64 end;
2777         u64 cur = start;
2778         u64 extent_offset;
2779         u64 last_byte = i_size_read(inode);
2780         u64 block_start;
2781         u64 iosize;
2782         sector_t sector;
2783         struct extent_state *cached_state = NULL;
2784         struct extent_map *em;
2785         struct block_device *bdev;
2786         int ret;
2787         int nr = 0;
2788         size_t pg_offset = 0;
2789         size_t blocksize;
2790         loff_t i_size = i_size_read(inode);
2791         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2792         u64 nr_delalloc;
2793         u64 delalloc_end;
2794         int page_started;
2795         int compressed;
2796         int write_flags;
2797         unsigned long nr_written = 0;
2798         bool fill_delalloc = true;
2799
2800         if (wbc->sync_mode == WB_SYNC_ALL)
2801                 write_flags = WRITE_SYNC;
2802         else
2803                 write_flags = WRITE;
2804
2805         trace___extent_writepage(page, inode, wbc);
2806
2807         WARN_ON(!PageLocked(page));
2808
2809         ClearPageError(page);
2810
2811         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2812         if (page->index > end_index ||
2813            (page->index == end_index && !pg_offset)) {
2814                 page->mapping->a_ops->invalidatepage(page, 0);
2815                 unlock_page(page);
2816                 return 0;
2817         }
2818
2819         if (page->index == end_index) {
2820                 char *userpage;
2821
2822                 userpage = kmap_atomic(page);
2823                 memset(userpage + pg_offset, 0,
2824                        PAGE_CACHE_SIZE - pg_offset);
2825                 kunmap_atomic(userpage);
2826                 flush_dcache_page(page);
2827         }
2828         pg_offset = 0;
2829
2830         set_page_extent_mapped(page);
2831
2832         if (!tree->ops || !tree->ops->fill_delalloc)
2833                 fill_delalloc = false;
2834
2835         delalloc_start = start;
2836         delalloc_end = 0;
2837         page_started = 0;
2838         if (!epd->extent_locked && fill_delalloc) {
2839                 u64 delalloc_to_write = 0;
2840                 /*
2841                  * make sure the wbc mapping index is at least updated
2842                  * to this page.
2843                  */
2844                 update_nr_written(page, wbc, 0);
2845
2846                 while (delalloc_end < page_end) {
2847                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2848                                                        page,
2849                                                        &delalloc_start,
2850                                                        &delalloc_end,
2851                                                        128 * 1024 * 1024);
2852                         if (nr_delalloc == 0) {
2853                                 delalloc_start = delalloc_end + 1;
2854                                 continue;
2855                         }
2856                         ret = tree->ops->fill_delalloc(inode, page,
2857                                                        delalloc_start,
2858                                                        delalloc_end,
2859                                                        &page_started,
2860                                                        &nr_written);
2861                         /* File system has been set read-only */
2862                         if (ret) {
2863                                 SetPageError(page);
2864                                 goto done;
2865                         }
2866                         /*
2867                          * delalloc_end is already one less than the total
2868                          * length, so we don't subtract one from
2869                          * PAGE_CACHE_SIZE
2870                          */
2871                         delalloc_to_write += (delalloc_end - delalloc_start +
2872                                               PAGE_CACHE_SIZE) >>
2873                                               PAGE_CACHE_SHIFT;
2874                         delalloc_start = delalloc_end + 1;
2875                 }
2876                 if (wbc->nr_to_write < delalloc_to_write) {
2877                         int thresh = 8192;
2878
2879                         if (delalloc_to_write < thresh * 2)
2880                                 thresh = delalloc_to_write;
2881                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2882                                                  thresh);
2883                 }
2884
2885                 /* did the fill delalloc function already unlock and start
2886                  * the IO?
2887                  */
2888                 if (page_started) {
2889                         ret = 0;
2890                         /*
2891                          * we've unlocked the page, so we can't update
2892                          * the mapping's writeback index, just update
2893                          * nr_to_write.
2894                          */
2895                         wbc->nr_to_write -= nr_written;
2896                         goto done_unlocked;
2897                 }
2898         }
2899         if (tree->ops && tree->ops->writepage_start_hook) {
2900                 ret = tree->ops->writepage_start_hook(page, start,
2901                                                       page_end);
2902                 if (ret) {
2903                         /* Fixup worker will requeue */
2904                         if (ret == -EBUSY)
2905                                 wbc->pages_skipped++;
2906                         else
2907                                 redirty_page_for_writepage(wbc, page);
2908                         update_nr_written(page, wbc, nr_written);
2909                         unlock_page(page);
2910                         ret = 0;
2911                         goto done_unlocked;
2912                 }
2913         }
2914
2915         /*
2916          * we don't want to touch the inode after unlocking the page,
2917          * so we update the mapping writeback index now
2918          */
2919         update_nr_written(page, wbc, nr_written + 1);
2920
2921         end = page_end;
2922         if (last_byte <= start) {
2923                 if (tree->ops && tree->ops->writepage_end_io_hook)
2924                         tree->ops->writepage_end_io_hook(page, start,
2925                                                          page_end, NULL, 1);
2926                 goto done;
2927         }
2928
2929         blocksize = inode->i_sb->s_blocksize;
2930
2931         while (cur <= end) {
2932                 if (cur >= last_byte) {
2933                         if (tree->ops && tree->ops->writepage_end_io_hook)
2934                                 tree->ops->writepage_end_io_hook(page, cur,
2935                                                          page_end, NULL, 1);
2936                         break;
2937                 }
2938                 em = epd->get_extent(inode, page, pg_offset, cur,
2939                                      end - cur + 1, 1);
2940                 if (IS_ERR_OR_NULL(em)) {
2941                         SetPageError(page);
2942                         break;
2943                 }
2944
2945                 extent_offset = cur - em->start;
2946                 BUG_ON(extent_map_end(em) <= cur);
2947                 BUG_ON(end < cur);
2948                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2949                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2950                 sector = (em->block_start + extent_offset) >> 9;
2951                 bdev = em->bdev;
2952                 block_start = em->block_start;
2953                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2954                 free_extent_map(em);
2955                 em = NULL;
2956
2957                 /*
2958                  * compressed and inline extents are written through other
2959                  * paths in the FS
2960                  */
2961                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2962                     block_start == EXTENT_MAP_INLINE) {
2963                         /*
2964                          * end_io notification does not happen here for
2965                          * compressed extents
2966                          */
2967                         if (!compressed && tree->ops &&
2968                             tree->ops->writepage_end_io_hook)
2969                                 tree->ops->writepage_end_io_hook(page, cur,
2970                                                          cur + iosize - 1,
2971                                                          NULL, 1);
2972                         else if (compressed) {
2973                                 /* we don't want to end_page_writeback on
2974                                  * a compressed extent.  this happens
2975                                  * elsewhere
2976                                  */
2977                                 nr++;
2978                         }
2979
2980                         cur += iosize;
2981                         pg_offset += iosize;
2982                         continue;
2983                 }
2984                 /* leave this out until we have a page_mkwrite call */
2985                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2986                                    EXTENT_DIRTY, 0, NULL)) {
2987                         cur = cur + iosize;
2988                         pg_offset += iosize;
2989                         continue;
2990                 }
2991
2992                 if (tree->ops && tree->ops->writepage_io_hook) {
2993                         ret = tree->ops->writepage_io_hook(page, cur,
2994                                                 cur + iosize - 1);
2995                 } else {
2996                         ret = 0;
2997                 }
2998                 if (ret) {
2999                         SetPageError(page);
3000                 } else {
3001                         unsigned long max_nr = end_index + 1;
3002
3003                         set_range_writeback(tree, cur, cur + iosize - 1);
3004                         if (!PageWriteback(page)) {
3005                                 printk(KERN_ERR "btrfs warning page %lu not "
3006                                        "writeback, cur %llu end %llu\n",
3007                                        page->index, (unsigned long long)cur,
3008                                        (unsigned long long)end);
3009                         }
3010
3011                         ret = submit_extent_page(write_flags, tree, page,
3012                                                  sector, iosize, pg_offset,
3013                                                  bdev, &epd->bio, max_nr,
3014                                                  end_bio_extent_writepage,
3015                                                  0, 0, 0);
3016                         if (ret)
3017                                 SetPageError(page);
3018                 }
3019                 cur = cur + iosize;
3020                 pg_offset += iosize;
3021                 nr++;
3022         }
3023 done:
3024         if (nr == 0) {
3025                 /* make sure the mapping tag for page dirty gets cleared */
3026                 set_page_writeback(page);
3027                 end_page_writeback(page);
3028         }
3029         unlock_page(page);
3030
3031 done_unlocked:
3032
3033         /* drop our reference on any cached states */
3034         free_extent_state(cached_state);
3035         return 0;
3036 }
3037
3038 static int eb_wait(void *word)
3039 {
3040         io_schedule();
3041         return 0;
3042 }
3043
3044 static void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3045 {
3046         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3047                     TASK_UNINTERRUPTIBLE);
3048 }
3049
3050 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3051                                      struct btrfs_fs_info *fs_info,
3052                                      struct extent_page_data *epd)
3053 {
3054         unsigned long i, num_pages;
3055         int flush = 0;
3056         int ret = 0;
3057
3058         if (!btrfs_try_tree_write_lock(eb)) {
3059                 flush = 1;
3060                 flush_write_bio(epd);
3061                 btrfs_tree_lock(eb);
3062         }
3063
3064         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3065                 btrfs_tree_unlock(eb);
3066                 if (!epd->sync_io)
3067                         return 0;
3068                 if (!flush) {
3069                         flush_write_bio(epd);
3070                         flush = 1;
3071                 }
3072                 while (1) {
3073                         wait_on_extent_buffer_writeback(eb);
3074                         btrfs_tree_lock(eb);
3075                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3076                                 break;
3077                         btrfs_tree_unlock(eb);
3078                 }
3079         }
3080
3081         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3082                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3083                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3084                 spin_lock(&fs_info->delalloc_lock);
3085                 if (fs_info->dirty_metadata_bytes >= eb->len)
3086                         fs_info->dirty_metadata_bytes -= eb->len;
3087                 else
3088                         WARN_ON(1);
3089                 spin_unlock(&fs_info->delalloc_lock);
3090                 ret = 1;
3091         }
3092
3093         btrfs_tree_unlock(eb);
3094
3095         if (!ret)
3096                 return ret;
3097
3098         num_pages = num_extent_pages(eb->start, eb->len);
3099         for (i = 0; i < num_pages; i++) {
3100                 struct page *p = extent_buffer_page(eb, i);
3101
3102                 if (!trylock_page(p)) {
3103                         if (!flush) {
3104                                 flush_write_bio(epd);
3105                                 flush = 1;
3106                         }
3107                         lock_page(p);
3108                 }
3109         }
3110
3111         return ret;
3112 }
3113
3114 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3115 {
3116         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3117         smp_mb__after_clear_bit();
3118         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3119 }
3120
3121 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3122 {
3123         int uptodate = err == 0;
3124         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
3125         struct extent_buffer *eb;
3126         int done;
3127
3128         do {
3129                 struct page *page = bvec->bv_page;
3130
3131                 bvec--;
3132                 eb = (struct extent_buffer *)page->private;
3133                 BUG_ON(!eb);
3134                 done = atomic_dec_and_test(&eb->io_pages);
3135
3136                 if (!uptodate || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3137                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3138                         ClearPageUptodate(page);
3139                         SetPageError(page);
3140                 }
3141
3142                 end_page_writeback(page);
3143
3144                 if (!done)
3145                         continue;
3146
3147                 end_extent_buffer_writeback(eb);
3148         } while (bvec >= bio->bi_io_vec);
3149
3150         bio_put(bio);
3151
3152 }
3153
3154 static int write_one_eb(struct extent_buffer *eb,
3155                         struct btrfs_fs_info *fs_info,
3156                         struct writeback_control *wbc,
3157                         struct extent_page_data *epd)
3158 {
3159         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3160         u64 offset = eb->start;
3161         unsigned long i, num_pages;
3162         int rw = (epd->sync_io ? WRITE_SYNC : WRITE);
3163         int ret;
3164
3165         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3166         num_pages = num_extent_pages(eb->start, eb->len);
3167         atomic_set(&eb->io_pages, num_pages);
3168         for (i = 0; i < num_pages; i++) {
3169                 struct page *p = extent_buffer_page(eb, i);
3170
3171                 clear_page_dirty_for_io(p);
3172                 set_page_writeback(p);
3173                 ret = submit_extent_page(rw, eb->tree, p, offset >> 9,
3174                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3175                                          -1, end_bio_extent_buffer_writepage,
3176                                          0, 0, 0);
3177                 if (ret) {
3178                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3179                         SetPageError(p);
3180                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3181                                 end_extent_buffer_writeback(eb);
3182                         ret = -EIO;
3183                         break;
3184                 }
3185                 offset += PAGE_CACHE_SIZE;
3186                 update_nr_written(p, wbc, 1);
3187                 unlock_page(p);
3188         }
3189
3190         if (unlikely(ret)) {
3191                 for (; i < num_pages; i++) {
3192                         struct page *p = extent_buffer_page(eb, i);
3193                         unlock_page(p);
3194                 }
3195         }
3196
3197         return ret;
3198 }
3199
3200 int btree_write_cache_pages(struct address_space *mapping,
3201                                    struct writeback_control *wbc)
3202 {
3203         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3204         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3205         struct extent_buffer *eb, *prev_eb = NULL;
3206         struct extent_page_data epd = {
3207                 .bio = NULL,
3208                 .tree = tree,
3209                 .extent_locked = 0,
3210                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3211         };
3212         int ret = 0;
3213         int done = 0;
3214         int nr_to_write_done = 0;
3215         struct pagevec pvec;
3216         int nr_pages;
3217         pgoff_t index;
3218         pgoff_t end;            /* Inclusive */
3219         int scanned = 0;
3220         int tag;
3221
3222         pagevec_init(&pvec, 0);
3223         if (wbc->range_cyclic) {
3224                 index = mapping->writeback_index; /* Start from prev offset */
3225                 end = -1;
3226         } else {
3227                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3228                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3229                 scanned = 1;
3230         }
3231         if (wbc->sync_mode == WB_SYNC_ALL)
3232                 tag = PAGECACHE_TAG_TOWRITE;
3233         else
3234                 tag = PAGECACHE_TAG_DIRTY;
3235 retry:
3236         if (wbc->sync_mode == WB_SYNC_ALL)
3237                 tag_pages_for_writeback(mapping, index, end);
3238         while (!done && !nr_to_write_done && (index <= end) &&
3239                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3240                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3241                 unsigned i;
3242
3243                 scanned = 1;
3244                 for (i = 0; i < nr_pages; i++) {
3245                         struct page *page = pvec.pages[i];
3246
3247                         if (!PagePrivate(page))
3248                                 continue;
3249
3250                         if (!wbc->range_cyclic && page->index > end) {
3251                                 done = 1;
3252                                 break;
3253                         }
3254
3255                         eb = (struct extent_buffer *)page->private;
3256                         if (!eb) {
3257                                 WARN_ON(1);
3258                                 continue;
3259                         }
3260
3261                         if (eb == prev_eb)
3262                                 continue;
3263
3264                         if (!atomic_inc_not_zero(&eb->refs)) {
3265                                 WARN_ON(1);
3266                                 continue;
3267                         }
3268
3269                         prev_eb = eb;
3270                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3271                         if (!ret) {
3272                                 free_extent_buffer(eb);
3273                                 continue;
3274                         }
3275
3276                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3277                         if (ret) {
3278                                 done = 1;
3279                                 free_extent_buffer(eb);
3280                                 break;
3281                         }
3282                         free_extent_buffer(eb);
3283
3284                         /*
3285                          * the filesystem may choose to bump up nr_to_write.
3286                          * We have to make sure to honor the new nr_to_write
3287                          * at any time
3288                          */
3289                         nr_to_write_done = wbc->nr_to_write <= 0;
3290                 }
3291                 pagevec_release(&pvec);
3292                 cond_resched();
3293         }
3294         if (!scanned && !done) {
3295                 /*
3296                  * We hit the last page and there is more work to be done: wrap
3297                  * back to the start of the file
3298                  */
3299                 scanned = 1;
3300                 index = 0;
3301                 goto retry;
3302         }
3303         flush_write_bio(&epd);
3304         return ret;
3305 }
3306
3307 /**
3308  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3309  * @mapping: address space structure to write
3310  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3311  * @writepage: function called for each page
3312  * @data: data passed to writepage function
3313  *
3314  * If a page is already under I/O, write_cache_pages() skips it, even
3315  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3316  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3317  * and msync() need to guarantee that all the data which was dirty at the time
3318  * the call was made get new I/O started against them.  If wbc->sync_mode is
3319  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3320  * existing IO to complete.
3321  */
3322 static int extent_write_cache_pages(struct extent_io_tree *tree,
3323                              struct address_space *mapping,
3324                              struct writeback_control *wbc,
3325                              writepage_t writepage, void *data,
3326                              void (*flush_fn)(void *))
3327 {
3328         int ret = 0;
3329         int done = 0;
3330         int nr_to_write_done = 0;
3331         struct pagevec pvec;
3332         int nr_pages;
3333         pgoff_t index;
3334         pgoff_t end;            /* Inclusive */
3335         int scanned = 0;
3336         int tag;
3337
3338         pagevec_init(&pvec, 0);
3339         if (wbc->range_cyclic) {
3340                 index = mapping->writeback_index; /* Start from prev offset */
3341                 end = -1;
3342         } else {
3343                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3344                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3345                 scanned = 1;
3346         }
3347         if (wbc->sync_mode == WB_SYNC_ALL)
3348                 tag = PAGECACHE_TAG_TOWRITE;
3349         else
3350                 tag = PAGECACHE_TAG_DIRTY;
3351 retry:
3352         if (wbc->sync_mode == WB_SYNC_ALL)
3353                 tag_pages_for_writeback(mapping, index, end);
3354         while (!done && !nr_to_write_done && (index <= end) &&
3355                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3356                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3357                 unsigned i;
3358
3359                 scanned = 1;
3360                 for (i = 0; i < nr_pages; i++) {
3361                         struct page *page = pvec.pages[i];
3362
3363                         /*
3364                          * At this point we hold neither mapping->tree_lock nor
3365                          * lock on the page itself: the page may be truncated or
3366                          * invalidated (changing page->mapping to NULL), or even
3367                          * swizzled back from swapper_space to tmpfs file
3368                          * mapping
3369                          */
3370                         if (tree->ops &&
3371                             tree->ops->write_cache_pages_lock_hook) {
3372                                 tree->ops->write_cache_pages_lock_hook(page,
3373                                                                data, flush_fn);
3374                         } else {
3375                                 if (!trylock_page(page)) {
3376                                         flush_fn(data);
3377                                         lock_page(page);
3378                                 }
3379                         }
3380
3381                         if (unlikely(page->mapping != mapping)) {
3382                                 unlock_page(page);
3383                                 continue;
3384                         }
3385
3386                         if (!wbc->range_cyclic && page->index > end) {
3387                                 done = 1;
3388                                 unlock_page(page);
3389                                 continue;
3390                         }
3391
3392                         if (wbc->sync_mode != WB_SYNC_NONE) {
3393                                 if (PageWriteback(page))
3394                                         flush_fn(data);
3395                                 wait_on_page_writeback(page);
3396                         }
3397
3398                         if (PageWriteback(page) ||
3399                             !clear_page_dirty_for_io(page)) {
3400                                 unlock_page(page);
3401                                 continue;
3402                         }
3403
3404                         ret = (*writepage)(page, wbc, data);
3405
3406                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3407                                 unlock_page(page);
3408                                 ret = 0;
3409                         }
3410                         if (ret)
3411                                 done = 1;
3412
3413                         /*
3414                          * the filesystem may choose to bump up nr_to_write.
3415                          * We have to make sure to honor the new nr_to_write
3416                          * at any time
3417                          */
3418                         nr_to_write_done = wbc->nr_to_write <= 0;
3419                 }
3420                 pagevec_release(&pvec);
3421                 cond_resched();
3422         }
3423         if (!scanned && !done) {
3424                 /*
3425                  * We hit the last page and there is more work to be done: wrap
3426                  * back to the start of the file
3427                  */
3428                 scanned = 1;
3429                 index = 0;
3430                 goto retry;
3431         }
3432         return ret;
3433 }
3434
3435 static void flush_epd_write_bio(struct extent_page_data *epd)
3436 {
3437         if (epd->bio) {
3438                 int rw = WRITE;
3439                 int ret;
3440
3441                 if (epd->sync_io)
3442                         rw = WRITE_SYNC;
3443
3444                 ret = submit_one_bio(rw, epd->bio, 0, 0);
3445                 BUG_ON(ret < 0); /* -ENOMEM */
3446                 epd->bio = NULL;
3447         }
3448 }
3449
3450 static noinline void flush_write_bio(void *data)
3451 {
3452         struct extent_page_data *epd = data;
3453         flush_epd_write_bio(epd);
3454 }
3455
3456 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3457                           get_extent_t *get_extent,
3458                           struct writeback_control *wbc)
3459 {
3460         int ret;
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_writepage(page, wbc, &epd);
3470
3471         flush_epd_write_bio(&epd);
3472         return ret;
3473 }
3474
3475 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3476                               u64 start, u64 end, get_extent_t *get_extent,
3477                               int mode)
3478 {
3479         int ret = 0;
3480         struct address_space *mapping = inode->i_mapping;
3481         struct page *page;
3482         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3483                 PAGE_CACHE_SHIFT;
3484
3485         struct extent_page_data epd = {
3486                 .bio = NULL,
3487                 .tree = tree,
3488                 .get_extent = get_extent,
3489                 .extent_locked = 1,
3490                 .sync_io = mode == WB_SYNC_ALL,
3491         };
3492         struct writeback_control wbc_writepages = {
3493                 .sync_mode      = mode,
3494                 .nr_to_write    = nr_pages * 2,
3495                 .range_start    = start,
3496                 .range_end      = end + 1,
3497         };
3498
3499         while (start <= end) {
3500                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3501                 if (clear_page_dirty_for_io(page))
3502                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3503                 else {
3504                         if (tree->ops && tree->ops->writepage_end_io_hook)
3505                                 tree->ops->writepage_end_io_hook(page, start,
3506                                                  start + PAGE_CACHE_SIZE - 1,
3507                                                  NULL, 1);
3508                         unlock_page(page);
3509                 }
3510                 page_cache_release(page);
3511                 start += PAGE_CACHE_SIZE;
3512         }
3513
3514         flush_epd_write_bio(&epd);
3515         return ret;
3516 }
3517
3518 int extent_writepages(struct extent_io_tree *tree,
3519                       struct address_space *mapping,
3520                       get_extent_t *get_extent,
3521                       struct writeback_control *wbc)
3522 {
3523         int ret = 0;
3524         struct extent_page_data epd = {
3525                 .bio = NULL,
3526                 .tree = tree,
3527                 .get_extent = get_extent,
3528                 .extent_locked = 0,
3529                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3530         };
3531
3532         ret = extent_write_cache_pages(tree, mapping, wbc,
3533                                        __extent_writepage, &epd,
3534                                        flush_write_bio);
3535         flush_epd_write_bio(&epd);
3536         return ret;
3537 }
3538
3539 int extent_readpages(struct extent_io_tree *tree,
3540                      struct address_space *mapping,
3541                      struct list_head *pages, unsigned nr_pages,
3542                      get_extent_t get_extent)
3543 {
3544         struct bio *bio = NULL;
3545         unsigned page_idx;
3546         unsigned long bio_flags = 0;
3547
3548         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3549                 struct page *page = list_entry(pages->prev, struct page, lru);
3550
3551                 prefetchw(&page->flags);
3552                 list_del(&page->lru);
3553                 if (!add_to_page_cache_lru(page, mapping,
3554                                         page->index, GFP_NOFS)) {
3555                         __extent_read_full_page(tree, page, get_extent,
3556                                                 &bio, 0, &bio_flags);
3557                 }
3558                 page_cache_release(page);
3559         }
3560         BUG_ON(!list_empty(pages));
3561         if (bio)
3562                 return submit_one_bio(READ, bio, 0, bio_flags);
3563         return 0;
3564 }
3565
3566 /*
3567  * basic invalidatepage code, this waits on any locked or writeback
3568  * ranges corresponding to the page, and then deletes any extent state
3569  * records from the tree
3570  */
3571 int extent_invalidatepage(struct extent_io_tree *tree,
3572                           struct page *page, unsigned long offset)
3573 {
3574         struct extent_state *cached_state = NULL;
3575         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
3576         u64 end = start + PAGE_CACHE_SIZE - 1;
3577         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3578
3579         start += (offset + blocksize - 1) & ~(blocksize - 1);
3580         if (start > end)
3581                 return 0;
3582
3583         lock_extent_bits(tree, start, end, 0, &cached_state);
3584         wait_on_page_writeback(page);
3585         clear_extent_bit(tree, start, end,
3586                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3587                          EXTENT_DO_ACCOUNTING,
3588                          1, 1, &cached_state, GFP_NOFS);
3589         return 0;
3590 }
3591
3592 /*
3593  * a helper for releasepage, this tests for areas of the page that
3594  * are locked or under IO and drops the related state bits if it is safe
3595  * to drop the page.
3596  */
3597 int try_release_extent_state(struct extent_map_tree *map,
3598                              struct extent_io_tree *tree, struct page *page,
3599                              gfp_t mask)
3600 {
3601         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3602         u64 end = start + PAGE_CACHE_SIZE - 1;
3603         int ret = 1;
3604
3605         if (test_range_bit(tree, start, end,
3606                            EXTENT_IOBITS, 0, NULL))
3607                 ret = 0;
3608         else {
3609                 if ((mask & GFP_NOFS) == GFP_NOFS)
3610                         mask = GFP_NOFS;
3611                 /*
3612                  * at this point we can safely clear everything except the
3613                  * locked bit and the nodatasum bit
3614                  */
3615                 ret = clear_extent_bit(tree, start, end,
3616                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3617                                  0, 0, NULL, mask);
3618
3619                 /* if clear_extent_bit failed for enomem reasons,
3620                  * we can't allow the release to continue.
3621                  */
3622                 if (ret < 0)
3623                         ret = 0;
3624                 else
3625                         ret = 1;
3626         }
3627         return ret;
3628 }
3629
3630 /*
3631  * a helper for releasepage.  As long as there are no locked extents
3632  * in the range corresponding to the page, both state records and extent
3633  * map records are removed
3634  */
3635 int try_release_extent_mapping(struct extent_map_tree *map,
3636                                struct extent_io_tree *tree, struct page *page,
3637                                gfp_t mask)
3638 {
3639         struct extent_map *em;
3640         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3641         u64 end = start + PAGE_CACHE_SIZE - 1;
3642
3643         if ((mask & __GFP_WAIT) &&
3644             page->mapping->host->i_size > 16 * 1024 * 1024) {
3645                 u64 len;
3646                 while (start <= end) {
3647                         len = end - start + 1;
3648                         write_lock(&map->lock);
3649                         em = lookup_extent_mapping(map, start, len);
3650                         if (!em) {
3651                                 write_unlock(&map->lock);
3652                                 break;
3653                         }
3654                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3655                             em->start != start) {
3656                                 write_unlock(&map->lock);
3657                                 free_extent_map(em);
3658                                 break;
3659                         }
3660                         if (!test_range_bit(tree, em->start,
3661                                             extent_map_end(em) - 1,
3662                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
3663                                             0, NULL)) {
3664                                 remove_extent_mapping(map, em);
3665                                 /* once for the rb tree */
3666                                 free_extent_map(em);
3667                         }
3668                         start = extent_map_end(em);
3669                         write_unlock(&map->lock);
3670
3671                         /* once for us */
3672                         free_extent_map(em);
3673                 }
3674         }
3675         return try_release_extent_state(map, tree, page, mask);
3676 }
3677
3678 /*
3679  * helper function for fiemap, which doesn't want to see any holes.
3680  * This maps until we find something past 'last'
3681  */
3682 static struct extent_map *get_extent_skip_holes(struct inode *inode,
3683                                                 u64 offset,
3684                                                 u64 last,
3685                                                 get_extent_t *get_extent)
3686 {
3687         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
3688         struct extent_map *em;
3689         u64 len;
3690
3691         if (offset >= last)
3692                 return NULL;
3693
3694         while(1) {
3695                 len = last - offset;
3696                 if (len == 0)
3697                         break;
3698                 len = (len + sectorsize - 1) & ~(sectorsize - 1);
3699                 em = get_extent(inode, NULL, 0, offset, len, 0);
3700                 if (IS_ERR_OR_NULL(em))
3701                         return em;
3702
3703                 /* if this isn't a hole return it */
3704                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
3705                     em->block_start != EXTENT_MAP_HOLE) {
3706                         return em;
3707                 }
3708
3709                 /* this is a hole, advance to the next extent */
3710                 offset = extent_map_end(em);
3711                 free_extent_map(em);
3712                 if (offset >= last)
3713                         break;
3714         }
3715         return NULL;
3716 }
3717
3718 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
3719                 __u64 start, __u64 len, get_extent_t *get_extent)
3720 {
3721         int ret = 0;
3722         u64 off = start;
3723         u64 max = start + len;
3724         u32 flags = 0;
3725         u32 found_type;
3726         u64 last;
3727         u64 last_for_get_extent = 0;
3728         u64 disko = 0;
3729         u64 isize = i_size_read(inode);
3730         struct btrfs_key found_key;
3731         struct extent_map *em = NULL;
3732         struct extent_state *cached_state = NULL;
3733         struct btrfs_path *path;
3734         struct btrfs_file_extent_item *item;
3735         int end = 0;
3736         u64 em_start = 0;
3737         u64 em_len = 0;
3738         u64 em_end = 0;
3739         unsigned long emflags;
3740
3741         if (len == 0)
3742                 return -EINVAL;
3743
3744         path = btrfs_alloc_path();
3745         if (!path)
3746                 return -ENOMEM;
3747         path->leave_spinning = 1;
3748
3749         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
3750         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
3751
3752         /*
3753          * lookup the last file extent.  We're not using i_size here
3754          * because there might be preallocation past i_size
3755          */
3756         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3757                                        path, btrfs_ino(inode), -1, 0);
3758         if (ret < 0) {
3759                 btrfs_free_path(path);
3760                 return ret;
3761         }
3762         WARN_ON(!ret);
3763         path->slots[0]--;
3764         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3765                               struct btrfs_file_extent_item);
3766         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3767         found_type = btrfs_key_type(&found_key);
3768
3769         /* No extents, but there might be delalloc bits */
3770         if (found_key.objectid != btrfs_ino(inode) ||
3771             found_type != BTRFS_EXTENT_DATA_KEY) {
3772                 /* have to trust i_size as the end */
3773                 last = (u64)-1;
3774                 last_for_get_extent = isize;
3775         } else {
3776                 /*
3777                  * remember the start of the last extent.  There are a
3778                  * bunch of different factors that go into the length of the
3779                  * extent, so its much less complex to remember where it started
3780                  */
3781                 last = found_key.offset;
3782                 last_for_get_extent = last + 1;
3783         }
3784         btrfs_free_path(path);
3785
3786         /*
3787          * we might have some extents allocated but more delalloc past those
3788          * extents.  so, we trust isize unless the start of the last extent is
3789          * beyond isize
3790          */
3791         if (last < isize) {
3792                 last = (u64)-1;
3793                 last_for_get_extent = isize;
3794         }
3795
3796         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3797                          &cached_state);
3798
3799         em = get_extent_skip_holes(inode, start, last_for_get_extent,
3800                                    get_extent);
3801         if (!em)
3802                 goto out;
3803         if (IS_ERR(em)) {
3804                 ret = PTR_ERR(em);
3805                 goto out;
3806         }
3807
3808         while (!end) {
3809                 u64 offset_in_extent;
3810
3811                 /* break if the extent we found is outside the range */
3812                 if (em->start >= max || extent_map_end(em) < off)
3813                         break;
3814
3815                 /*
3816                  * get_extent may return an extent that starts before our
3817                  * requested range.  We have to make sure the ranges
3818                  * we return to fiemap always move forward and don't
3819                  * overlap, so adjust the offsets here
3820                  */
3821                 em_start = max(em->start, off);
3822
3823                 /*
3824                  * record the offset from the start of the extent
3825                  * for adjusting the disk offset below
3826                  */
3827                 offset_in_extent = em_start - em->start;
3828                 em_end = extent_map_end(em);
3829                 em_len = em_end - em_start;
3830                 emflags = em->flags;
3831                 disko = 0;
3832                 flags = 0;
3833
3834                 /*
3835                  * bump off for our next call to get_extent
3836                  */
3837                 off = extent_map_end(em);
3838                 if (off >= max)
3839                         end = 1;
3840
3841                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3842                         end = 1;
3843                         flags |= FIEMAP_EXTENT_LAST;
3844                 } else if (em->block_start == EXTENT_MAP_INLINE) {
3845                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
3846                                   FIEMAP_EXTENT_NOT_ALIGNED);
3847                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3848                         flags |= (FIEMAP_EXTENT_DELALLOC |
3849                                   FIEMAP_EXTENT_UNKNOWN);
3850                 } else {
3851                         disko = em->block_start + offset_in_extent;
3852                 }
3853                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3854                         flags |= FIEMAP_EXTENT_ENCODED;
3855
3856                 free_extent_map(em);
3857                 em = NULL;
3858                 if ((em_start >= last) || em_len == (u64)-1 ||
3859                    (last == (u64)-1 && isize <= em_end)) {
3860                         flags |= FIEMAP_EXTENT_LAST;
3861                         end = 1;
3862                 }
3863
3864                 /* now scan forward to see if this is really the last extent. */
3865                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
3866                                            get_extent);
3867                 if (IS_ERR(em)) {
3868                         ret = PTR_ERR(em);
3869                         goto out;
3870                 }
3871                 if (!em) {
3872                         flags |= FIEMAP_EXTENT_LAST;
3873                         end = 1;
3874                 }
3875                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3876                                               em_len, flags);
3877                 if (ret)
3878                         goto out_free;
3879         }
3880 out_free:
3881         free_extent_map(em);
3882 out:
3883         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3884                              &cached_state, GFP_NOFS);
3885         return ret;
3886 }
3887
3888 inline struct page *extent_buffer_page(struct extent_buffer *eb,
3889                                               unsigned long i)
3890 {
3891         return eb->pages[i];
3892 }
3893
3894 inline unsigned long num_extent_pages(u64 start, u64 len)
3895 {
3896         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3897                 (start >> PAGE_CACHE_SHIFT);
3898 }
3899
3900 static void __free_extent_buffer(struct extent_buffer *eb)
3901 {
3902 #if LEAK_DEBUG
3903         unsigned long flags;
3904         spin_lock_irqsave(&leak_lock, flags);
3905         list_del(&eb->leak_list);
3906         spin_unlock_irqrestore(&leak_lock, flags);
3907 #endif
3908         if (eb->pages && eb->pages != eb->inline_pages)
3909                 kfree(eb->pages);
3910         kmem_cache_free(extent_buffer_cache, eb);
3911 }
3912
3913 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3914                                                    u64 start,
3915                                                    unsigned long len,
3916                                                    gfp_t mask)
3917 {
3918         struct extent_buffer *eb = NULL;
3919 #if LEAK_DEBUG
3920         unsigned long flags;
3921 #endif
3922
3923         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3924         if (eb == NULL)
3925                 return NULL;
3926         eb->start = start;
3927         eb->len = len;
3928         eb->tree = tree;
3929         rwlock_init(&eb->lock);
3930         atomic_set(&eb->write_locks, 0);
3931         atomic_set(&eb->read_locks, 0);
3932         atomic_set(&eb->blocking_readers, 0);
3933         atomic_set(&eb->blocking_writers, 0);
3934         atomic_set(&eb->spinning_readers, 0);
3935         atomic_set(&eb->spinning_writers, 0);
3936         eb->lock_nested = 0;
3937         init_waitqueue_head(&eb->write_lock_wq);
3938         init_waitqueue_head(&eb->read_lock_wq);
3939
3940 #if LEAK_DEBUG
3941         spin_lock_irqsave(&leak_lock, flags);
3942         list_add(&eb->leak_list, &buffers);
3943         spin_unlock_irqrestore(&leak_lock, flags);
3944 #endif
3945         spin_lock_init(&eb->refs_lock);
3946         atomic_set(&eb->refs, 1);
3947         atomic_set(&eb->io_pages, 0);
3948
3949         if (len > MAX_INLINE_EXTENT_BUFFER_SIZE) {
3950                 struct page **pages;
3951                 int num_pages = (len + PAGE_CACHE_SIZE - 1) >>
3952                         PAGE_CACHE_SHIFT;
3953                 pages = kzalloc(num_pages, mask);
3954                 if (!pages) {
3955                         __free_extent_buffer(eb);
3956                         return NULL;
3957                 }
3958                 eb->pages = pages;
3959         } else {
3960                 eb->pages = eb->inline_pages;
3961         }
3962
3963         return eb;
3964 }
3965
3966 static int extent_buffer_under_io(struct extent_buffer *eb)
3967 {
3968         return (atomic_read(&eb->io_pages) ||
3969                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
3970                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3971 }
3972
3973 /*
3974  * Helper for releasing extent buffer page.
3975  */
3976 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3977                                                 unsigned long start_idx)
3978 {
3979         unsigned long index;
3980         struct page *page;
3981
3982         BUG_ON(extent_buffer_under_io(eb));
3983
3984         index = num_extent_pages(eb->start, eb->len);
3985         if (start_idx >= index)
3986                 return;
3987
3988         do {
3989                 index--;
3990                 page = extent_buffer_page(eb, index);
3991                 if (page) {
3992                         spin_lock(&page->mapping->private_lock);
3993                         /*
3994                          * We do this since we'll remove the pages after we've
3995                          * removed the eb from the radix tree, so we could race
3996                          * and have this page now attached to the new eb.  So
3997                          * only clear page_private if it's still connected to
3998                          * this eb.
3999                          */
4000                         if (PagePrivate(page) &&
4001                             page->private == (unsigned long)eb) {
4002                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4003                                 BUG_ON(PageDirty(page));
4004                                 BUG_ON(PageWriteback(page));
4005                                 /*
4006                                  * We need to make sure we haven't be attached
4007                                  * to a new eb.
4008                                  */
4009                                 ClearPagePrivate(page);
4010                                 set_page_private(page, 0);
4011                                 /* One for the page private */
4012                                 page_cache_release(page);
4013                         }
4014                         spin_unlock(&page->mapping->private_lock);
4015
4016                         /* One for when we alloced the page */
4017                         page_cache_release(page);
4018                 }
4019         } while (index != start_idx);
4020 }
4021
4022 /*
4023  * Helper for releasing the extent buffer.
4024  */
4025 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4026 {
4027         btrfs_release_extent_buffer_page(eb, 0);
4028         __free_extent_buffer(eb);
4029 }
4030
4031 static void check_buffer_tree_ref(struct extent_buffer *eb)
4032 {
4033         /* the ref bit is tricky.  We have to make sure it is set
4034          * if we have the buffer dirty.   Otherwise the
4035          * code to free a buffer can end up dropping a dirty
4036          * page
4037          *
4038          * Once the ref bit is set, it won't go away while the
4039          * buffer is dirty or in writeback, and it also won't
4040          * go away while we have the reference count on the
4041          * eb bumped.
4042          *
4043          * We can't just set the ref bit without bumping the
4044          * ref on the eb because free_extent_buffer might
4045          * see the ref bit and try to clear it.  If this happens
4046          * free_extent_buffer might end up dropping our original
4047          * ref by mistake and freeing the page before we are able
4048          * to add one more ref.
4049          *
4050          * So bump the ref count first, then set the bit.  If someone
4051          * beat us to it, drop the ref we added.
4052          */
4053         if (!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4054                 atomic_inc(&eb->refs);
4055                 if (test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4056                         atomic_dec(&eb->refs);
4057         }
4058 }
4059
4060 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4061 {
4062         unsigned long num_pages, i;
4063
4064         check_buffer_tree_ref(eb);
4065
4066         num_pages = num_extent_pages(eb->start, eb->len);
4067         for (i = 0; i < num_pages; i++) {
4068                 struct page *p = extent_buffer_page(eb, i);
4069                 mark_page_accessed(p);
4070         }
4071 }
4072
4073 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
4074                                           u64 start, unsigned long len)
4075 {
4076         unsigned long num_pages = num_extent_pages(start, len);
4077         unsigned long i;
4078         unsigned long index = start >> PAGE_CACHE_SHIFT;
4079         struct extent_buffer *eb;
4080         struct extent_buffer *exists = NULL;
4081         struct page *p;
4082         struct address_space *mapping = tree->mapping;
4083         int uptodate = 1;
4084         int ret;
4085
4086         rcu_read_lock();
4087         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4088         if (eb && atomic_inc_not_zero(&eb->refs)) {
4089                 rcu_read_unlock();
4090                 mark_extent_buffer_accessed(eb);
4091                 return eb;
4092         }
4093         rcu_read_unlock();
4094
4095         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
4096         if (!eb)
4097                 return NULL;
4098
4099         for (i = 0; i < num_pages; i++, index++) {
4100                 p = find_or_create_page(mapping, index, GFP_NOFS);
4101                 if (!p) {
4102                         WARN_ON(1);
4103                         goto free_eb;
4104                 }
4105
4106                 spin_lock(&mapping->private_lock);
4107                 if (PagePrivate(p)) {
4108                         /*
4109                          * We could have already allocated an eb for this page
4110                          * and attached one so lets see if we can get a ref on
4111                          * the existing eb, and if we can we know it's good and
4112                          * we can just return that one, else we know we can just
4113                          * overwrite page->private.
4114                          */
4115                         exists = (struct extent_buffer *)p->private;
4116                         if (atomic_inc_not_zero(&exists->refs)) {
4117                                 spin_unlock(&mapping->private_lock);
4118                                 unlock_page(p);
4119                                 mark_extent_buffer_accessed(exists);
4120                                 goto free_eb;
4121                         }
4122
4123                         /*
4124                          * Do this so attach doesn't complain and we need to
4125                          * drop the ref the old guy had.
4126                          */
4127                         ClearPagePrivate(p);
4128                         WARN_ON(PageDirty(p));
4129                         page_cache_release(p);
4130                 }
4131                 attach_extent_buffer_page(eb, p);
4132                 spin_unlock(&mapping->private_lock);
4133                 WARN_ON(PageDirty(p));
4134                 mark_page_accessed(p);
4135                 eb->pages[i] = p;
4136                 if (!PageUptodate(p))
4137                         uptodate = 0;
4138
4139                 /*
4140                  * see below about how we avoid a nasty race with release page
4141                  * and why we unlock later
4142                  */
4143         }
4144         if (uptodate)
4145                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4146 again:
4147         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4148         if (ret)
4149                 goto free_eb;
4150
4151         spin_lock(&tree->buffer_lock);
4152         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
4153         if (ret == -EEXIST) {
4154                 exists = radix_tree_lookup(&tree->buffer,
4155                                                 start >> PAGE_CACHE_SHIFT);
4156                 if (!atomic_inc_not_zero(&exists->refs)) {
4157                         spin_unlock(&tree->buffer_lock);
4158                         radix_tree_preload_end();
4159                         exists = NULL;
4160                         goto again;
4161                 }
4162                 spin_unlock(&tree->buffer_lock);
4163                 radix_tree_preload_end();
4164                 mark_extent_buffer_accessed(exists);
4165                 goto free_eb;
4166         }
4167         /* add one reference for the tree */
4168         spin_lock(&eb->refs_lock);
4169         check_buffer_tree_ref(eb);
4170         spin_unlock(&eb->refs_lock);
4171         spin_unlock(&tree->buffer_lock);
4172         radix_tree_preload_end();
4173
4174         /*
4175          * there is a race where release page may have
4176          * tried to find this extent buffer in the radix
4177          * but failed.  It will tell the VM it is safe to
4178          * reclaim the, and it will clear the page private bit.
4179          * We must make sure to set the page private bit properly
4180          * after the extent buffer is in the radix tree so
4181          * it doesn't get lost
4182          */
4183         SetPageChecked(eb->pages[0]);
4184         for (i = 1; i < num_pages; i++) {
4185                 p = extent_buffer_page(eb, i);
4186                 ClearPageChecked(p);
4187                 unlock_page(p);
4188         }
4189         unlock_page(eb->pages[0]);
4190         return eb;
4191
4192 free_eb:
4193         for (i = 0; i < num_pages; i++) {
4194                 if (eb->pages[i])
4195                         unlock_page(eb->pages[i]);
4196         }
4197
4198         if (!atomic_dec_and_test(&eb->refs))
4199                 return exists;
4200         btrfs_release_extent_buffer(eb);
4201         return exists;
4202 }
4203
4204 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
4205                                          u64 start, unsigned long len)
4206 {
4207         struct extent_buffer *eb;
4208
4209         rcu_read_lock();
4210         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4211         if (eb && atomic_inc_not_zero(&eb->refs)) {
4212                 rcu_read_unlock();
4213                 mark_extent_buffer_accessed(eb);
4214                 return eb;
4215         }
4216         rcu_read_unlock();
4217
4218         return NULL;
4219 }
4220
4221 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4222 {
4223         struct extent_buffer *eb =
4224                         container_of(head, struct extent_buffer, rcu_head);
4225
4226         __free_extent_buffer(eb);
4227 }
4228
4229 /* Expects to have eb->eb_lock already held */
4230 static void release_extent_buffer(struct extent_buffer *eb, gfp_t mask)
4231 {
4232         WARN_ON(atomic_read(&eb->refs) == 0);
4233         if (atomic_dec_and_test(&eb->refs)) {
4234                 struct extent_io_tree *tree = eb->tree;
4235
4236                 spin_unlock(&eb->refs_lock);
4237
4238                 spin_lock(&tree->buffer_lock);
4239                 radix_tree_delete(&tree->buffer,
4240                                   eb->start >> PAGE_CACHE_SHIFT);
4241                 spin_unlock(&tree->buffer_lock);
4242
4243                 /* Should be safe to release our pages at this point */
4244                 btrfs_release_extent_buffer_page(eb, 0);
4245
4246                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4247                 return;
4248         }
4249         spin_unlock(&eb->refs_lock);
4250 }
4251
4252 void free_extent_buffer(struct extent_buffer *eb)
4253 {
4254         if (!eb)
4255                 return;
4256
4257         spin_lock(&eb->refs_lock);
4258         if (atomic_read(&eb->refs) == 2 &&
4259             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4260             !extent_buffer_under_io(eb) &&
4261             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4262                 atomic_dec(&eb->refs);
4263
4264         /*
4265          * I know this is terrible, but it's temporary until we stop tracking
4266          * the uptodate bits and such for the extent buffers.
4267          */
4268         release_extent_buffer(eb, GFP_ATOMIC);
4269 }
4270
4271 void free_extent_buffer_stale(struct extent_buffer *eb)
4272 {
4273         if (!eb)
4274                 return;
4275
4276         spin_lock(&eb->refs_lock);
4277         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4278
4279         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4280             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4281                 atomic_dec(&eb->refs);
4282         release_extent_buffer(eb, GFP_NOFS);
4283 }
4284
4285 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4286 {
4287         unsigned long i;
4288         unsigned long num_pages;
4289         struct page *page;
4290
4291         num_pages = num_extent_pages(eb->start, eb->len);
4292
4293         for (i = 0; i < num_pages; i++) {
4294                 page = extent_buffer_page(eb, i);
4295                 if (!PageDirty(page))
4296                         continue;
4297
4298                 lock_page(page);
4299                 WARN_ON(!PagePrivate(page));
4300
4301                 clear_page_dirty_for_io(page);
4302                 spin_lock_irq(&page->mapping->tree_lock);
4303                 if (!PageDirty(page)) {
4304                         radix_tree_tag_clear(&page->mapping->page_tree,
4305                                                 page_index(page),
4306                                                 PAGECACHE_TAG_DIRTY);
4307                 }
4308                 spin_unlock_irq(&page->mapping->tree_lock);
4309                 ClearPageError(page);
4310                 unlock_page(page);
4311         }
4312         WARN_ON(atomic_read(&eb->refs) == 0);
4313 }
4314
4315 int set_extent_buffer_dirty(struct extent_buffer *eb)
4316 {
4317         unsigned long i;
4318         unsigned long num_pages;
4319         int was_dirty = 0;
4320
4321         check_buffer_tree_ref(eb);
4322
4323         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4324
4325         num_pages = num_extent_pages(eb->start, eb->len);
4326         WARN_ON(atomic_read(&eb->refs) == 0);
4327         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4328
4329         for (i = 0; i < num_pages; i++)
4330                 set_page_dirty(extent_buffer_page(eb, i));
4331         return was_dirty;
4332 }
4333
4334 static int range_straddles_pages(u64 start, u64 len)
4335 {
4336         if (len < PAGE_CACHE_SIZE)
4337                 return 1;
4338         if (start & (PAGE_CACHE_SIZE - 1))
4339                 return 1;
4340         if ((start + len) & (PAGE_CACHE_SIZE - 1))
4341                 return 1;
4342         return 0;
4343 }
4344
4345 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4346 {
4347         unsigned long i;
4348         struct page *page;
4349         unsigned long num_pages;
4350
4351         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4352         num_pages = num_extent_pages(eb->start, eb->len);
4353         for (i = 0; i < num_pages; i++) {
4354                 page = extent_buffer_page(eb, i);
4355                 if (page)
4356                         ClearPageUptodate(page);
4357         }
4358         return 0;
4359 }
4360
4361 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4362 {
4363         unsigned long i;
4364         struct page *page;
4365         unsigned long num_pages;
4366
4367         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4368         num_pages = num_extent_pages(eb->start, eb->len);
4369         for (i = 0; i < num_pages; i++) {
4370                 page = extent_buffer_page(eb, i);
4371                 SetPageUptodate(page);
4372         }
4373         return 0;
4374 }
4375
4376 int extent_range_uptodate(struct extent_io_tree *tree,
4377                           u64 start, u64 end)
4378 {
4379         struct page *page;
4380         int ret;
4381         int pg_uptodate = 1;
4382         int uptodate;
4383         unsigned long index;
4384
4385         if (range_straddles_pages(start, end - start + 1)) {
4386                 ret = test_range_bit(tree, start, end,
4387                                      EXTENT_UPTODATE, 1, NULL);
4388                 if (ret)
4389                         return 1;
4390         }
4391         while (start <= end) {
4392                 index = start >> PAGE_CACHE_SHIFT;
4393                 page = find_get_page(tree->mapping, index);
4394                 if (!page)
4395                         return 1;
4396                 uptodate = PageUptodate(page);
4397                 page_cache_release(page);
4398                 if (!uptodate) {
4399                         pg_uptodate = 0;
4400                         break;
4401                 }
4402                 start += PAGE_CACHE_SIZE;
4403         }
4404         return pg_uptodate;
4405 }
4406
4407 int extent_buffer_uptodate(struct extent_buffer *eb)
4408 {
4409         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4410 }
4411
4412 int read_extent_buffer_pages(struct extent_io_tree *tree,
4413                              struct extent_buffer *eb, u64 start, int wait,
4414                              get_extent_t *get_extent, int mirror_num)
4415 {
4416         unsigned long i;
4417         unsigned long start_i;
4418         struct page *page;
4419         int err;
4420         int ret = 0;
4421         int locked_pages = 0;
4422         int all_uptodate = 1;
4423         unsigned long num_pages;
4424         unsigned long num_reads = 0;
4425         struct bio *bio = NULL;
4426         unsigned long bio_flags = 0;
4427
4428         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4429                 return 0;
4430
4431         if (start) {
4432                 WARN_ON(start < eb->start);
4433                 start_i = (start >> PAGE_CACHE_SHIFT) -
4434                         (eb->start >> PAGE_CACHE_SHIFT);
4435         } else {
4436                 start_i = 0;
4437         }
4438
4439         num_pages = num_extent_pages(eb->start, eb->len);
4440         for (i = start_i; i < num_pages; i++) {
4441                 page = extent_buffer_page(eb, i);
4442                 if (wait == WAIT_NONE) {
4443                         if (!trylock_page(page))
4444                                 goto unlock_exit;
4445                 } else {
4446                         lock_page(page);
4447                 }
4448                 locked_pages++;
4449                 if (!PageUptodate(page)) {
4450                         num_reads++;
4451                         all_uptodate = 0;
4452                 }
4453         }
4454         if (all_uptodate) {
4455                 if (start_i == 0)
4456                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4457                 goto unlock_exit;
4458         }
4459
4460         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4461         eb->failed_mirror = 0;
4462         atomic_set(&eb->io_pages, num_reads);
4463         for (i = start_i; i < num_pages; i++) {
4464                 page = extent_buffer_page(eb, i);
4465                 if (!PageUptodate(page)) {
4466                         ClearPageError(page);
4467                         err = __extent_read_full_page(tree, page,
4468                                                       get_extent, &bio,
4469                                                       mirror_num, &bio_flags);
4470                         if (err)
4471                                 ret = err;
4472                 } else {
4473                         unlock_page(page);
4474                 }
4475         }
4476
4477         if (bio) {
4478                 err = submit_one_bio(READ, bio, mirror_num, bio_flags);
4479                 if (err)
4480                         return err;
4481         }
4482
4483         if (ret || wait != WAIT_COMPLETE)
4484                 return ret;
4485
4486         for (i = start_i; i < num_pages; i++) {
4487                 page = extent_buffer_page(eb, i);
4488                 wait_on_page_locked(page);
4489                 if (!PageUptodate(page))
4490                         ret = -EIO;
4491         }
4492
4493         return ret;
4494
4495 unlock_exit:
4496         i = start_i;
4497         while (locked_pages > 0) {
4498                 page = extent_buffer_page(eb, i);
4499                 i++;
4500                 unlock_page(page);
4501                 locked_pages--;
4502         }
4503         return ret;
4504 }
4505
4506 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4507                         unsigned long start,
4508                         unsigned long len)
4509 {
4510         size_t cur;
4511         size_t offset;
4512         struct page *page;
4513         char *kaddr;
4514         char *dst = (char *)dstv;
4515         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4516         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4517
4518         WARN_ON(start > eb->len);
4519         WARN_ON(start + len > eb->start + eb->len);
4520
4521         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4522
4523         while (len > 0) {
4524                 page = extent_buffer_page(eb, i);
4525
4526                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4527                 kaddr = page_address(page);
4528                 memcpy(dst, kaddr + offset, cur);
4529
4530                 dst += cur;
4531                 len -= cur;
4532                 offset = 0;
4533                 i++;
4534         }
4535 }
4536
4537 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4538                                unsigned long min_len, char **map,
4539                                unsigned long *map_start,
4540                                unsigned long *map_len)
4541 {
4542         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4543         char *kaddr;
4544         struct page *p;
4545         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4546         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4547         unsigned long end_i = (start_offset + start + min_len - 1) >>
4548                 PAGE_CACHE_SHIFT;
4549
4550         if (i != end_i)
4551                 return -EINVAL;
4552
4553         if (i == 0) {
4554                 offset = start_offset;
4555                 *map_start = 0;
4556         } else {
4557                 offset = 0;
4558                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4559         }
4560
4561         if (start + min_len > eb->len) {
4562                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4563                        "wanted %lu %lu\n", (unsigned long long)eb->start,
4564                        eb->len, start, min_len);
4565                 WARN_ON(1);
4566                 return -EINVAL;
4567         }
4568
4569         p = extent_buffer_page(eb, i);
4570         kaddr = page_address(p);
4571         *map = kaddr + offset;
4572         *map_len = PAGE_CACHE_SIZE - offset;
4573         return 0;
4574 }
4575
4576 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4577                           unsigned long start,
4578                           unsigned long len)
4579 {
4580         size_t cur;
4581         size_t offset;
4582         struct page *page;
4583         char *kaddr;
4584         char *ptr = (char *)ptrv;
4585         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4586         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4587         int ret = 0;
4588
4589         WARN_ON(start > eb->len);
4590         WARN_ON(start + len > eb->start + eb->len);
4591
4592         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4593
4594         while (len > 0) {
4595                 page = extent_buffer_page(eb, i);
4596
4597                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4598
4599                 kaddr = page_address(page);
4600                 ret = memcmp(ptr, kaddr + offset, cur);
4601                 if (ret)
4602                         break;
4603
4604                 ptr += cur;
4605                 len -= cur;
4606                 offset = 0;
4607                 i++;
4608         }
4609         return ret;
4610 }
4611
4612 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4613                          unsigned long start, unsigned long len)
4614 {
4615         size_t cur;
4616         size_t offset;
4617         struct page *page;
4618         char *kaddr;
4619         char *src = (char *)srcv;
4620         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4621         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4622
4623         WARN_ON(start > eb->len);
4624         WARN_ON(start + len > eb->start + eb->len);
4625
4626         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4627
4628         while (len > 0) {
4629                 page = extent_buffer_page(eb, i);
4630                 WARN_ON(!PageUptodate(page));
4631
4632                 cur = min(len, PAGE_CACHE_SIZE - offset);
4633                 kaddr = page_address(page);
4634                 memcpy(kaddr + offset, src, cur);
4635
4636                 src += cur;
4637                 len -= cur;
4638                 offset = 0;
4639                 i++;
4640         }
4641 }
4642
4643 void memset_extent_buffer(struct extent_buffer *eb, char c,
4644                           unsigned long start, unsigned long len)
4645 {
4646         size_t cur;
4647         size_t offset;
4648         struct page *page;
4649         char *kaddr;
4650         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4651         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4652
4653         WARN_ON(start > eb->len);
4654         WARN_ON(start + len > eb->start + eb->len);
4655
4656         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4657
4658         while (len > 0) {
4659                 page = extent_buffer_page(eb, i);
4660                 WARN_ON(!PageUptodate(page));
4661
4662                 cur = min(len, PAGE_CACHE_SIZE - offset);
4663                 kaddr = page_address(page);
4664                 memset(kaddr + offset, c, cur);
4665
4666                 len -= cur;
4667                 offset = 0;
4668                 i++;
4669         }
4670 }
4671
4672 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
4673                         unsigned long dst_offset, unsigned long src_offset,
4674                         unsigned long len)
4675 {
4676         u64 dst_len = dst->len;
4677         size_t cur;
4678         size_t offset;
4679         struct page *page;
4680         char *kaddr;
4681         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4682         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4683
4684         WARN_ON(src->len != dst_len);
4685
4686         offset = (start_offset + dst_offset) &
4687                 ((unsigned long)PAGE_CACHE_SIZE - 1);
4688
4689         while (len > 0) {
4690                 page = extent_buffer_page(dst, i);
4691                 WARN_ON(!PageUptodate(page));
4692
4693                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
4694
4695                 kaddr = page_address(page);
4696                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4697
4698                 src_offset += cur;
4699                 len -= cur;
4700                 offset = 0;
4701                 i++;
4702         }
4703 }
4704
4705 static void move_pages(struct page *dst_page, struct page *src_page,
4706                        unsigned long dst_off, unsigned long src_off,
4707                        unsigned long len)
4708 {
4709         char *dst_kaddr = page_address(dst_page);
4710         if (dst_page == src_page) {
4711                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
4712         } else {
4713                 char *src_kaddr = page_address(src_page);
4714                 char *p = dst_kaddr + dst_off + len;
4715                 char *s = src_kaddr + src_off + len;
4716
4717                 while (len--)
4718                         *--p = *--s;
4719         }
4720 }
4721
4722 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4723 {
4724         unsigned long distance = (src > dst) ? src - dst : dst - src;
4725         return distance < len;
4726 }
4727
4728 static void copy_pages(struct page *dst_page, struct page *src_page,
4729                        unsigned long dst_off, unsigned long src_off,
4730                        unsigned long len)
4731 {
4732         char *dst_kaddr = page_address(dst_page);
4733         char *src_kaddr;
4734         int must_memmove = 0;
4735
4736         if (dst_page != src_page) {
4737                 src_kaddr = page_address(src_page);
4738         } else {
4739                 src_kaddr = dst_kaddr;
4740                 if (areas_overlap(src_off, dst_off, len))
4741                         must_memmove = 1;
4742         }
4743
4744         if (must_memmove)
4745                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
4746         else
4747                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
4748 }
4749
4750 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4751                            unsigned long src_offset, unsigned long len)
4752 {
4753         size_t cur;
4754         size_t dst_off_in_page;
4755         size_t src_off_in_page;
4756         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4757         unsigned long dst_i;
4758         unsigned long src_i;
4759
4760         if (src_offset + len > dst->len) {
4761                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4762                        "len %lu dst len %lu\n", src_offset, len, dst->len);
4763                 BUG_ON(1);
4764         }
4765         if (dst_offset + len > dst->len) {
4766                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4767                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
4768                 BUG_ON(1);
4769         }
4770
4771         while (len > 0) {
4772                 dst_off_in_page = (start_offset + dst_offset) &
4773                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4774                 src_off_in_page = (start_offset + src_offset) &
4775                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4776
4777                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4778                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
4779
4780                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
4781                                                src_off_in_page));
4782                 cur = min_t(unsigned long, cur,
4783                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
4784
4785                 copy_pages(extent_buffer_page(dst, dst_i),
4786                            extent_buffer_page(dst, src_i),
4787                            dst_off_in_page, src_off_in_page, cur);
4788
4789                 src_offset += cur;
4790                 dst_offset += cur;
4791                 len -= cur;
4792         }
4793 }
4794
4795 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4796                            unsigned long src_offset, unsigned long len)
4797 {
4798         size_t cur;
4799         size_t dst_off_in_page;
4800         size_t src_off_in_page;
4801         unsigned long dst_end = dst_offset + len - 1;
4802         unsigned long src_end = src_offset + len - 1;
4803         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4804         unsigned long dst_i;
4805         unsigned long src_i;
4806
4807         if (src_offset + len > dst->len) {
4808                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4809                        "len %lu len %lu\n", src_offset, len, dst->len);
4810                 BUG_ON(1);
4811         }
4812         if (dst_offset + len > dst->len) {
4813                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4814                        "len %lu len %lu\n", dst_offset, len, dst->len);
4815                 BUG_ON(1);
4816         }
4817         if (dst_offset < src_offset) {
4818                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4819                 return;
4820         }
4821         while (len > 0) {
4822                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4823                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4824
4825                 dst_off_in_page = (start_offset + dst_end) &
4826                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4827                 src_off_in_page = (start_offset + src_end) &
4828                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4829
4830                 cur = min_t(unsigned long, len, src_off_in_page + 1);
4831                 cur = min(cur, dst_off_in_page + 1);
4832                 move_pages(extent_buffer_page(dst, dst_i),
4833                            extent_buffer_page(dst, src_i),
4834                            dst_off_in_page - cur + 1,
4835                            src_off_in_page - cur + 1, cur);
4836
4837                 dst_end -= cur;
4838                 src_end -= cur;
4839                 len -= cur;
4840         }
4841 }
4842
4843 int try_release_extent_buffer(struct page *page, gfp_t mask)
4844 {
4845         struct extent_buffer *eb;
4846
4847         /*
4848          * We need to make sure noboody is attaching this page to an eb right
4849          * now.
4850          */
4851         spin_lock(&page->mapping->private_lock);
4852         if (!PagePrivate(page)) {
4853                 spin_unlock(&page->mapping->private_lock);
4854                 return 1;
4855         }
4856
4857         eb = (struct extent_buffer *)page->private;
4858         BUG_ON(!eb);
4859
4860         /*
4861          * This is a little awful but should be ok, we need to make sure that
4862          * the eb doesn't disappear out from under us while we're looking at
4863          * this page.
4864          */
4865         spin_lock(&eb->refs_lock);
4866         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4867                 spin_unlock(&eb->refs_lock);
4868                 spin_unlock(&page->mapping->private_lock);
4869                 return 0;
4870         }
4871         spin_unlock(&page->mapping->private_lock);
4872
4873         if ((mask & GFP_NOFS) == GFP_NOFS)
4874                 mask = GFP_NOFS;
4875
4876         /*
4877          * If tree ref isn't set then we know the ref on this eb is a real ref,
4878          * so just return, this page will likely be freed soon anyway.
4879          */
4880         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4881                 spin_unlock(&eb->refs_lock);
4882                 return 0;
4883         }
4884         release_extent_buffer(eb, mask);
4885
4886         return 1;
4887 }