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 #include "rcu-string.h"
  24
  25 static struct kmem_cache *extent_state_cache;
  26 static struct kmem_cache *extent_buffer_cache;
  27
  28 static LIST_HEAD(buffers);
  29 static LIST_HEAD(states);
  30
  31 #define LEAK_DEBUG 0
  32 #if LEAK_DEBUG
  33 static DEFINE_SPINLOCK(leak_lock);
  34 #endif
  35
  36 #define BUFFER_LRU_MAX 64
  37
  38 struct tree_entry {
  39         u64 start;
  40         u64 end;
  41         struct rb_node rb_node;
  42 };
  43
  44 struct extent_page_data {
  45         struct bio *bio;
  46         struct extent_io_tree *tree;
  47         get_extent_t *get_extent;
  48
  49         /* tells writepage not to lock the state bits for this range
  50          * it still does the unlocking
  51          */
  52         unsigned int extent_locked:1;
  53
  54         /* tells the submit_bio code to use a WRITE_SYNC */
  55         unsigned int sync_io:1;
  56 };
  57
  58 static noinline void flush_write_bio(void *data);
  59 static inline struct btrfs_fs_info *
  60 tree_fs_info(struct extent_io_tree *tree)
  61 {
  62         return btrfs_sb(tree->mapping->host->i_sb);
  63 }
  64
  65 int __init extent_io_init(void)
  66 {
  67         extent_state_cache = kmem_cache_create("extent_state",
  68                         sizeof(struct extent_state), 0,
  69                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  70         if (!extent_state_cache)
  71                 return -ENOMEM;
  72
  73         extent_buffer_cache = kmem_cache_create("extent_buffers",
  74                         sizeof(struct extent_buffer), 0,
  75                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  76         if (!extent_buffer_cache)
  77                 goto free_state_cache;
  78         return 0;
  79
  80 free_state_cache:
  81         kmem_cache_destroy(extent_state_cache);
  82         return -ENOMEM;
  83 }
  84
  85 void extent_io_exit(void)
  86 {
  87         struct extent_state *state;
  88         struct extent_buffer *eb;
  89
  90         while (!list_empty(&states)) {
  91                 state = list_entry(states.next, struct extent_state, leak_list);
  92                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  93                        "state %lu in tree %p refs %d\n",
  94                        (unsigned long long)state->start,
  95                        (unsigned long long)state->end,
  96                        state->state, state->tree, atomic_read(&state->refs));
  97                 list_del(&state->leak_list);
  98                 kmem_cache_free(extent_state_cache, state);
  99
 100         }
 101
 102         while (!list_empty(&buffers)) {
 103                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
 104                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
 105                        "refs %d\n", (unsigned long long)eb->start,
 106                        eb->len, atomic_read(&eb->refs));
 107                 list_del(&eb->leak_list);
 108                 kmem_cache_free(extent_buffer_cache, eb);
 109         }
 110         if (extent_state_cache)
 111                 kmem_cache_destroy(extent_state_cache);
 112         if (extent_buffer_cache)
 113                 kmem_cache_destroy(extent_buffer_cache);
 114 }
 115
 116 void extent_io_tree_init(struct extent_io_tree *tree,
 117                          struct address_space *mapping)
 118 {
 119         tree->state = RB_ROOT;
 120         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 121         tree->ops = NULL;
 122         tree->dirty_bytes = 0;
 123         spin_lock_init(&tree->lock);
 124         spin_lock_init(&tree->buffer_lock);
 125         tree->mapping = mapping;
 126 }
 127
 128 static struct extent_state *alloc_extent_state(gfp_t mask)
 129 {
 130         struct extent_state *state;
 131 #if LEAK_DEBUG
 132         unsigned long flags;
 133 #endif
 134
 135         state = kmem_cache_alloc(extent_state_cache, mask);
 136         if (!state)
 137                 return state;
 138         state->state = 0;
 139         state->private = 0;
 140         state->tree = NULL;
 141 #if LEAK_DEBUG
 142         spin_lock_irqsave(&leak_lock, flags);
 143         list_add(&state->leak_list, &states);
 144         spin_unlock_irqrestore(&leak_lock, flags);
 145 #endif
 146         atomic_set(&state->refs, 1);
 147         init_waitqueue_head(&state->wq);
 148         trace_alloc_extent_state(state, mask, _RET_IP_);
 149         return state;
 150 }
 151
 152 void free_extent_state(struct extent_state *state)
 153 {
 154         if (!state)
 155                 return;
 156         if (atomic_dec_and_test(&state->refs)) {
 157 #if LEAK_DEBUG
 158                 unsigned long flags;
 159 #endif
 160                 WARN_ON(state->tree);
 161 #if LEAK_DEBUG
 162                 spin_lock_irqsave(&leak_lock, flags);
 163                 list_del(&state->leak_list);
 164                 spin_unlock_irqrestore(&leak_lock, flags);
 165 #endif
 166                 trace_free_extent_state(state, _RET_IP_);
 167                 kmem_cache_free(extent_state_cache, state);
 168         }
 169 }
 170
 171 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 172                                    struct rb_node *node)
 173 {
 174         struct rb_node **p = &root->rb_node;
 175         struct rb_node *parent = NULL;
 176         struct tree_entry *entry;
 177
 178         while (*p) {
 179                 parent = *p;
 180                 entry = rb_entry(parent, struct tree_entry, rb_node);
 181
 182                 if (offset < entry->start)
 183                         p = &(*p)->rb_left;
 184                 else if (offset > entry->end)
 185                         p = &(*p)->rb_right;
 186                 else
 187                         return parent;
 188         }
 189
 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 static struct extent_state *next_state(struct extent_state *state)
 406 {
 407         struct rb_node *next = rb_next(&state->rb_node);
 408         if (next)
 409                 return rb_entry(next, struct extent_state, rb_node);
 410         else
 411                 return NULL;
 412 }
 413
 414 /*
 415  * utility function to clear some bits in an extent state struct.
 416  * it will optionally wake up any one waiting on this state (wake == 1).
 417  *
 418  * If no bits are set on the state struct after clearing things, the
 419  * struct is freed and removed from the tree
 420  */
 421 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 422                                             struct extent_state *state,
 423                                             int *bits, int wake)
 424 {
 425         struct extent_state *next;
 426         int bits_to_clear = *bits & ~EXTENT_CTLBITS;
 427
 428         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 429                 u64 range = state->end - state->start + 1;
 430                 WARN_ON(range > tree->dirty_bytes);
 431                 tree->dirty_bytes -= range;
 432         }
 433         clear_state_cb(tree, state, bits);
 434         state->state &= ~bits_to_clear;
 435         if (wake)
 436                 wake_up(&state->wq);
 437         if (state->state == 0) {
 438                 next = next_state(state);
 439                 if (state->tree) {
 440                         rb_erase(&state->rb_node, &tree->state);
 441                         state->tree = NULL;
 442                         free_extent_state(state);
 443                 } else {
 444                         WARN_ON(1);
 445                 }
 446         } else {
 447                 merge_state(tree, state);
 448                 next = next_state(state);
 449         }
 450         return next;
 451 }
 452
 453 static struct extent_state *
 454 alloc_extent_state_atomic(struct extent_state *prealloc)
 455 {
 456         if (!prealloc)
 457                 prealloc = alloc_extent_state(GFP_ATOMIC);
 458
 459         return prealloc;
 460 }
 461
 462 void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 463 {
 464         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 465                     "Extent tree was modified by another "
 466                     "thread while locked.");
 467 }
 468
 469 /*
 470  * clear some bits on a range in the tree.  This may require splitting
 471  * or inserting elements in the tree, so the gfp mask is used to
 472  * indicate which allocations or sleeping are allowed.
 473  *
 474  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 475  * the given range from the tree regardless of state (ie for truncate).
 476  *
 477  * the range [start, end] is inclusive.
 478  *
 479  * This takes the tree lock, and returns 0 on success and < 0 on error.
 480  */
 481 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 482                      int bits, int wake, int delete,
 483                      struct extent_state **cached_state,
 484                      gfp_t mask)
 485 {
 486         struct extent_state *state;
 487         struct extent_state *cached;
 488         struct extent_state *prealloc = NULL;
 489         struct rb_node *node;
 490         u64 last_end;
 491         int err;
 492         int clear = 0;
 493
 494         if (delete)
 495                 bits |= ~EXTENT_CTLBITS;
 496         bits |= EXTENT_FIRST_DELALLOC;
 497
 498         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 499                 clear = 1;
 500 again:
 501         if (!prealloc && (mask & __GFP_WAIT)) {
 502                 prealloc = alloc_extent_state(mask);
 503                 if (!prealloc)
 504                         return -ENOMEM;
 505         }
 506
 507         spin_lock(&tree->lock);
 508         if (cached_state) {
 509                 cached = *cached_state;
 510
 511                 if (clear) {
 512                         *cached_state = NULL;
 513                         cached_state = NULL;
 514                 }
 515
 516                 if (cached && cached->tree && cached->start <= start &&
 517                     cached->end > start) {
 518                         if (clear)
 519                                 atomic_dec(&cached->refs);
 520                         state = cached;
 521                         goto hit_next;
 522                 }
 523                 if (clear)
 524                         free_extent_state(cached);
 525         }
 526         /*
 527          * this search will find the extents that end after
 528          * our range starts
 529          */
 530         node = tree_search(tree, start);
 531         if (!node)
 532                 goto out;
 533         state = rb_entry(node, struct extent_state, rb_node);
 534 hit_next:
 535         if (state->start > end)
 536                 goto out;
 537         WARN_ON(state->end < start);
 538         last_end = state->end;
 539
 540         /* the state doesn't have the wanted bits, go ahead */
 541         if (!(state->state & bits)) {
 542                 state = next_state(state);
 543                 goto next;
 544         }
 545
 546         /*
 547          *     | ---- desired range ---- |
 548          *  | state | or
 549          *  | ------------- state -------------- |
 550          *
 551          * We need to split the extent we found, and may flip
 552          * bits on second half.
 553          *
 554          * If the extent we found extends past our range, we
 555          * just split and search again.  It'll get split again
 556          * the next time though.
 557          *
 558          * If the extent we found is inside our range, we clear
 559          * the desired bit on it.
 560          */
 561
 562         if (state->start < start) {
 563                 prealloc = alloc_extent_state_atomic(prealloc);
 564                 BUG_ON(!prealloc);
 565                 err = split_state(tree, state, prealloc, start);
 566                 if (err)
 567                         extent_io_tree_panic(tree, err);
 568
 569                 prealloc = NULL;
 570                 if (err)
 571                         goto out;
 572                 if (state->end <= end) {
 573                         state = clear_state_bit(tree, state, &bits, wake);
 574                         goto next;
 575                 }
 576                 goto search_again;
 577         }
 578         /*
 579          * | ---- desired range ---- |
 580          *                        | state |
 581          * We need to split the extent, and clear the bit
 582          * on the first half
 583          */
 584         if (state->start <= end && state->end > end) {
 585                 prealloc = alloc_extent_state_atomic(prealloc);
 586                 BUG_ON(!prealloc);
 587                 err = split_state(tree, state, prealloc, end + 1);
 588                 if (err)
 589                         extent_io_tree_panic(tree, err);
 590
 591                 if (wake)
 592                         wake_up(&state->wq);
 593
 594                 clear_state_bit(tree, prealloc, &bits, wake);
 595
 596                 prealloc = NULL;
 597                 goto out;
 598         }
 599
 600         state = clear_state_bit(tree, state, &bits, wake);
 601 next:
 602         if (last_end == (u64)-1)
 603                 goto out;
 604         start = last_end + 1;
 605         if (start <= end && state && !need_resched())
 606                 goto hit_next;
 607         goto search_again;
 608
 609 out:
 610         spin_unlock(&tree->lock);
 611         if (prealloc)
 612                 free_extent_state(prealloc);
 613
 614         return 0;
 615
 616 search_again:
 617         if (start > end)
 618                 goto out;
 619         spin_unlock(&tree->lock);
 620         if (mask & __GFP_WAIT)
 621                 cond_resched();
 622         goto again;
 623 }
 624
 625 static void wait_on_state(struct extent_io_tree *tree,
 626                           struct extent_state *state)
 627                 __releases(tree->lock)
 628                 __acquires(tree->lock)
 629 {
 630         DEFINE_WAIT(wait);
 631         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 632         spin_unlock(&tree->lock);
 633         schedule();
 634         spin_lock(&tree->lock);
 635         finish_wait(&state->wq, &wait);
 636 }
 637
 638 /*
 639  * waits for one or more bits to clear on a range in the state tree.
 640  * The range [start, end] is inclusive.
 641  * The tree lock is taken by this function
 642  */
 643 void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 644 {
 645         struct extent_state *state;
 646         struct rb_node *node;
 647
 648         spin_lock(&tree->lock);
 649 again:
 650         while (1) {
 651                 /*
 652                  * this search will find all the extents that end after
 653                  * our range starts
 654                  */
 655                 node = tree_search(tree, start);
 656                 if (!node)
 657                         break;
 658
 659                 state = rb_entry(node, struct extent_state, rb_node);
 660
 661                 if (state->start > end)
 662                         goto out;
 663
 664                 if (state->state & bits) {
 665                         start = state->start;
 666                         atomic_inc(&state->refs);
 667                         wait_on_state(tree, state);
 668                         free_extent_state(state);
 669                         goto again;
 670                 }
 671                 start = state->end + 1;
 672
 673                 if (start > end)
 674                         break;
 675
 676                 cond_resched_lock(&tree->lock);
 677         }
 678 out:
 679         spin_unlock(&tree->lock);
 680 }
 681
 682 static void set_state_bits(struct extent_io_tree *tree,
 683                            struct extent_state *state,
 684                            int *bits)
 685 {
 686         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 687
 688         set_state_cb(tree, state, bits);
 689         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 690                 u64 range = state->end - state->start + 1;
 691                 tree->dirty_bytes += range;
 692         }
 693         state->state |= bits_to_set;
 694 }
 695
 696 static void cache_state(struct extent_state *state,
 697                         struct extent_state **cached_ptr)
 698 {
 699         if (cached_ptr && !(*cached_ptr)) {
 700                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 701                         *cached_ptr = state;
 702                         atomic_inc(&state->refs);
 703                 }
 704         }
 705 }
 706
 707 static void uncache_state(struct extent_state **cached_ptr)
 708 {
 709         if (cached_ptr && (*cached_ptr)) {
 710                 struct extent_state *state = *cached_ptr;
 711                 *cached_ptr = NULL;
 712                 free_extent_state(state);
 713         }
 714 }
 715
 716 /*
 717  * set some bits on a range in the tree.  This may require allocations or
 718  * sleeping, so the gfp mask is used to indicate what is allowed.
 719  *
 720  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 721  * part of the range already has the desired bits set.  The start of the
 722  * existing range is returned in failed_start in this case.
 723  *
 724  * [start, end] is inclusive This takes the tree lock.
 725  */
 726
 727 static int __must_check
 728 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 729                  int bits, int exclusive_bits, u64 *failed_start,
 730                  struct extent_state **cached_state, gfp_t mask)
 731 {
 732         struct extent_state *state;
 733         struct extent_state *prealloc = NULL;
 734         struct rb_node *node;
 735         int err = 0;
 736         u64 last_start;
 737         u64 last_end;
 738
 739         bits |= EXTENT_FIRST_DELALLOC;
 740 again:
 741         if (!prealloc && (mask & __GFP_WAIT)) {
 742                 prealloc = alloc_extent_state(mask);
 743                 BUG_ON(!prealloc);
 744         }
 745
 746         spin_lock(&tree->lock);
 747         if (cached_state && *cached_state) {
 748                 state = *cached_state;
 749                 if (state->start <= start && state->end > start &&
 750                     state->tree) {
 751                         node = &state->rb_node;
 752                         goto hit_next;
 753                 }
 754         }
 755         /*
 756          * this search will find all the extents that end after
 757          * our range starts.
 758          */
 759         node = tree_search(tree, start);
 760         if (!node) {
 761                 prealloc = alloc_extent_state_atomic(prealloc);
 762                 BUG_ON(!prealloc);
 763                 err = insert_state(tree, prealloc, start, end, &bits);
 764                 if (err)
 765                         extent_io_tree_panic(tree, err);
 766
 767                 prealloc = NULL;
 768                 goto out;
 769         }
 770         state = rb_entry(node, struct extent_state, rb_node);
 771 hit_next:
 772         last_start = state->start;
 773         last_end = state->end;
 774
 775         /*
 776          * | ---- desired range ---- |
 777          * | state |
 778          *
 779          * Just lock what we found and keep going
 780          */
 781         if (state->start == start && state->end <= end) {
 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                 cache_state(state, cached_state);
 790                 merge_state(tree, state);
 791                 if (last_end == (u64)-1)
 792                         goto out;
 793                 start = last_end + 1;
 794                 state = next_state(state);
 795                 if (start < end && state && state->start == start &&
 796                     !need_resched())
 797                         goto hit_next;
 798                 goto search_again;
 799         }
 800
 801         /*
 802          *     | ---- desired range ---- |
 803          * | state |
 804          *   or
 805          * | ------------- state -------------- |
 806          *
 807          * We need to split the extent we found, and may flip bits on
 808          * second half.
 809          *
 810          * If the extent we found extends past our
 811          * range, we just split and search again.  It'll get split
 812          * again the next time though.
 813          *
 814          * If the extent we found is inside our range, we set the
 815          * desired bit on it.
 816          */
 817         if (state->start < start) {
 818                 if (state->state & exclusive_bits) {
 819                         *failed_start = start;
 820                         err = -EEXIST;
 821                         goto out;
 822                 }
 823
 824                 prealloc = alloc_extent_state_atomic(prealloc);
 825                 BUG_ON(!prealloc);
 826                 err = split_state(tree, state, prealloc, start);
 827                 if (err)
 828                         extent_io_tree_panic(tree, err);
 829
 830                 prealloc = NULL;
 831                 if (err)
 832                         goto out;
 833                 if (state->end <= end) {
 834                         set_state_bits(tree, state, &bits);
 835                         cache_state(state, cached_state);
 836                         merge_state(tree, state);
 837                         if (last_end == (u64)-1)
 838                                 goto out;
 839                         start = last_end + 1;
 840                         state = next_state(state);
 841                         if (start < end && state && state->start == start &&
 842                             !need_resched())
 843                                 goto hit_next;
 844                 }
 845                 goto search_again;
 846         }
 847         /*
 848          * | ---- desired range ---- |
 849          *     | state | or               | state |
 850          *
 851          * There's a hole, we need to insert something in it and
 852          * ignore the extent we found.
 853          */
 854         if (state->start > start) {
 855                 u64 this_end;
 856                 if (end < last_start)
 857                         this_end = end;
 858                 else
 859                         this_end = last_start - 1;
 860
 861                 prealloc = alloc_extent_state_atomic(prealloc);
 862                 BUG_ON(!prealloc);
 863
 864                 /*
 865                  * Avoid to free 'prealloc' if it can be merged with
 866                  * the later extent.
 867                  */
 868                 err = insert_state(tree, prealloc, start, this_end,
 869                                    &bits);
 870                 if (err)
 871                         extent_io_tree_panic(tree, err);
 872
 873                 cache_state(prealloc, cached_state);
 874                 prealloc = NULL;
 875                 start = this_end + 1;
 876                 goto search_again;
 877         }
 878         /*
 879          * | ---- desired range ---- |
 880          *                        | state |
 881          * We need to split the extent, and set the bit
 882          * on the first half
 883          */
 884         if (state->start <= end && state->end > end) {
 885                 if (state->state & exclusive_bits) {
 886                         *failed_start = start;
 887                         err = -EEXIST;
 888                         goto out;
 889                 }
 890
 891                 prealloc = alloc_extent_state_atomic(prealloc);
 892                 BUG_ON(!prealloc);
 893                 err = split_state(tree, state, prealloc, end + 1);
 894                 if (err)
 895                         extent_io_tree_panic(tree, err);
 896
 897                 set_state_bits(tree, prealloc, &bits);
 898                 cache_state(prealloc, cached_state);
 899                 merge_state(tree, prealloc);
 900                 prealloc = NULL;
 901                 goto out;
 902         }
 903
 904         goto search_again;
 905
 906 out:
 907         spin_unlock(&tree->lock);
 908         if (prealloc)
 909                 free_extent_state(prealloc);
 910
 911         return err;
 912
 913 search_again:
 914         if (start > end)
 915                 goto out;
 916         spin_unlock(&tree->lock);
 917         if (mask & __GFP_WAIT)
 918                 cond_resched();
 919         goto again;
 920 }
 921
 922 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits,
 923                    u64 *failed_start, struct extent_state **cached_state,
 924                    gfp_t mask)
 925 {
 926         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
 927                                 cached_state, mask);
 928 }
 929
 930
 931 /**
 932  * convert_extent_bit - convert all bits in a given range from one bit to
 933  *                      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                 set_state_bits(tree, state, &bits);
 995                 state = clear_state_bit(tree, state, &clear_bits, 0);
 996                 if (last_end == (u64)-1)
 997                         goto out;
 998                 start = last_end + 1;
 999                 if (start < end && state && state->start == start &&
1000                     !need_resched())
1001                         goto hit_next;
1002                 goto search_again;
1003         }
1004
1005         /*
1006          *     | ---- desired range ---- |
1007          * | state |
1008          *   or
1009          * | ------------- state -------------- |
1010          *
1011          * We need to split the extent we found, and may flip bits on
1012          * second half.
1013          *
1014          * If the extent we found extends past our
1015          * range, we just split and search again.  It'll get split
1016          * again the next time though.
1017          *
1018          * If the extent we found is inside our range, we set the
1019          * desired bit on it.
1020          */
1021         if (state->start < start) {
1022                 prealloc = alloc_extent_state_atomic(prealloc);
1023                 if (!prealloc) {
1024                         err = -ENOMEM;
1025                         goto out;
1026                 }
1027                 err = split_state(tree, state, prealloc, start);
1028                 if (err)
1029                         extent_io_tree_panic(tree, err);
1030                 prealloc = NULL;
1031                 if (err)
1032                         goto out;
1033                 if (state->end <= end) {
1034                         set_state_bits(tree, state, &bits);
1035                         state = clear_state_bit(tree, state, &clear_bits, 0);
1036                         if (last_end == (u64)-1)
1037                                 goto out;
1038                         start = last_end + 1;
1039                         if (start < end && state && state->start == start &&
1040                             !need_resched())
1041                                 goto hit_next;
1042                 }
1043                 goto search_again;
1044         }
1045         /*
1046          * | ---- desired range ---- |
1047          *     | state | or               | state |
1048          *
1049          * There's a hole, we need to insert something in it and
1050          * ignore the extent we found.
1051          */
1052         if (state->start > start) {
1053                 u64 this_end;
1054                 if (end < last_start)
1055                         this_end = end;
1056                 else
1057                         this_end = last_start - 1;
1058
1059                 prealloc = alloc_extent_state_atomic(prealloc);
1060                 if (!prealloc) {
1061                         err = -ENOMEM;
1062                         goto out;
1063                 }
1064
1065                 /*
1066                  * Avoid to free 'prealloc' if it can be merged with
1067                  * the later extent.
1068                  */
1069                 err = insert_state(tree, prealloc, start, this_end,
1070                                    &bits);
1071                 if (err)
1072                         extent_io_tree_panic(tree, err);
1073                 prealloc = NULL;
1074                 start = this_end + 1;
1075                 goto search_again;
1076         }
1077         /*
1078          * | ---- desired range ---- |
1079          *                        | state |
1080          * We need to split the extent, and set the bit
1081          * on the first half
1082          */
1083         if (state->start <= end && state->end > end) {
1084                 prealloc = alloc_extent_state_atomic(prealloc);
1085                 if (!prealloc) {
1086                         err = -ENOMEM;
1087                         goto out;
1088                 }
1089
1090                 err = split_state(tree, state, prealloc, end + 1);
1091                 if (err)
1092                         extent_io_tree_panic(tree, err);
1093
1094                 set_state_bits(tree, prealloc, &bits);
1095                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1096                 prealloc = NULL;
1097                 goto out;
1098         }
1099
1100         goto search_again;
1101
1102 out:
1103         spin_unlock(&tree->lock);
1104         if (prealloc)
1105                 free_extent_state(prealloc);
1106
1107         return err;
1108
1109 search_again:
1110         if (start > end)
1111                 goto out;
1112         spin_unlock(&tree->lock);
1113         if (mask & __GFP_WAIT)
1114                 cond_resched();
1115         goto again;
1116 }
1117
1118 /* wrappers around set/clear extent bit */
1119 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1120                      gfp_t mask)
1121 {
1122         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1123                               NULL, mask);
1124 }
1125
1126 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1127                     int bits, gfp_t mask)
1128 {
1129         return set_extent_bit(tree, start, end, bits, NULL,
1130                               NULL, mask);
1131 }
1132
1133 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1134                       int bits, gfp_t mask)
1135 {
1136         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1137 }
1138
1139 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1140                         struct extent_state **cached_state, gfp_t mask)
1141 {
1142         return set_extent_bit(tree, start, end,
1143                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1144                               NULL, cached_state, mask);
1145 }
1146
1147 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1148                        gfp_t mask)
1149 {
1150         return clear_extent_bit(tree, start, end,
1151                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1152                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1153 }
1154
1155 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1156                      gfp_t mask)
1157 {
1158         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1159                               NULL, mask);
1160 }
1161
1162 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1163                         struct extent_state **cached_state, gfp_t mask)
1164 {
1165         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
1166                               cached_state, mask);
1167 }
1168
1169 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1170                           struct extent_state **cached_state, gfp_t mask)
1171 {
1172         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1173                                 cached_state, mask);
1174 }
1175
1176 /*
1177  * either insert or lock state struct between start and end use mask to tell
1178  * us if waiting is desired.
1179  */
1180 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1181                      int bits, struct extent_state **cached_state)
1182 {
1183         int err;
1184         u64 failed_start;
1185         while (1) {
1186                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1187                                        EXTENT_LOCKED, &failed_start,
1188                                        cached_state, GFP_NOFS);
1189                 if (err == -EEXIST) {
1190                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1191                         start = failed_start;
1192                 } else
1193                         break;
1194                 WARN_ON(start > end);
1195         }
1196         return err;
1197 }
1198
1199 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1200 {
1201         return lock_extent_bits(tree, start, end, 0, NULL);
1202 }
1203
1204 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1205 {
1206         int err;
1207         u64 failed_start;
1208
1209         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1210                                &failed_start, NULL, GFP_NOFS);
1211         if (err == -EEXIST) {
1212                 if (failed_start > start)
1213                         clear_extent_bit(tree, start, failed_start - 1,
1214                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1215                 return 0;
1216         }
1217         return 1;
1218 }
1219
1220 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1221                          struct extent_state **cached, gfp_t mask)
1222 {
1223         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1224                                 mask);
1225 }
1226
1227 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1228 {
1229         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1230                                 GFP_NOFS);
1231 }
1232
1233 /*
1234  * helper function to set both pages and extents in the tree writeback
1235  */
1236 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1237 {
1238         unsigned long index = start >> PAGE_CACHE_SHIFT;
1239         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1240         struct page *page;
1241
1242         while (index <= end_index) {
1243                 page = find_get_page(tree->mapping, index);
1244                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1245                 set_page_writeback(page);
1246                 page_cache_release(page);
1247                 index++;
1248         }
1249         return 0;
1250 }
1251
1252 /* find the first state struct with 'bits' set after 'start', and
1253  * return it.  tree->lock must be held.  NULL will returned if
1254  * nothing was found after 'start'
1255  */
1256 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1257                                                  u64 start, int bits)
1258 {
1259         struct rb_node *node;
1260         struct extent_state *state;
1261
1262         /*
1263          * this search will find all the extents that end after
1264          * our range starts.
1265          */
1266         node = tree_search(tree, start);
1267         if (!node)
1268                 goto out;
1269
1270         while (1) {
1271                 state = rb_entry(node, struct extent_state, rb_node);
1272                 if (state->end >= start && (state->state & bits))
1273                         return state;
1274
1275                 node = rb_next(node);
1276                 if (!node)
1277                         break;
1278         }
1279 out:
1280         return NULL;
1281 }
1282
1283 /*
1284  * find the first offset in the io tree with 'bits' set. zero is
1285  * returned if we find something, and *start_ret and *end_ret are
1286  * set to reflect the state struct that was found.
1287  *
1288  * If nothing was found, 1 is returned. If found something, return 0.
1289  */
1290 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1291                           u64 *start_ret, u64 *end_ret, int bits)
1292 {
1293         struct extent_state *state;
1294         int ret = 1;
1295
1296         spin_lock(&tree->lock);
1297         state = find_first_extent_bit_state(tree, start, bits);
1298         if (state) {
1299                 *start_ret = state->start;
1300                 *end_ret = state->end;
1301                 ret = 0;
1302         }
1303         spin_unlock(&tree->lock);
1304         return ret;
1305 }
1306
1307 /*
1308  * find a contiguous range of bytes in the file marked as delalloc, not
1309  * more than 'max_bytes'.  start and end are used to return the range,
1310  *
1311  * 1 is returned if we find something, 0 if nothing was in the tree
1312  */
1313 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1314                                         u64 *start, u64 *end, u64 max_bytes,
1315                                         struct extent_state **cached_state)
1316 {
1317         struct rb_node *node;
1318         struct extent_state *state;
1319         u64 cur_start = *start;
1320         u64 found = 0;
1321         u64 total_bytes = 0;
1322
1323         spin_lock(&tree->lock);
1324
1325         /*
1326          * this search will find all the extents that end after
1327          * our range starts.
1328          */
1329         node = tree_search(tree, cur_start);
1330         if (!node) {
1331                 if (!found)
1332                         *end = (u64)-1;
1333                 goto out;
1334         }
1335
1336         while (1) {
1337                 state = rb_entry(node, struct extent_state, rb_node);
1338                 if (found && (state->start != cur_start ||
1339                               (state->state & EXTENT_BOUNDARY))) {
1340                         goto out;
1341                 }
1342                 if (!(state->state & EXTENT_DELALLOC)) {
1343                         if (!found)
1344                                 *end = state->end;
1345                         goto out;
1346                 }
1347                 if (!found) {
1348                         *start = state->start;
1349                         *cached_state = state;
1350                         atomic_inc(&state->refs);
1351                 }
1352                 found++;
1353                 *end = state->end;
1354                 cur_start = state->end + 1;
1355                 node = rb_next(node);
1356                 if (!node)
1357                         break;
1358                 total_bytes += state->end - state->start + 1;
1359                 if (total_bytes >= max_bytes)
1360                         break;
1361         }
1362 out:
1363         spin_unlock(&tree->lock);
1364         return found;
1365 }
1366
1367 static noinline void __unlock_for_delalloc(struct inode *inode,
1368                                            struct page *locked_page,
1369                                            u64 start, u64 end)
1370 {
1371         int ret;
1372         struct page *pages[16];
1373         unsigned long index = start >> PAGE_CACHE_SHIFT;
1374         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1375         unsigned long nr_pages = end_index - index + 1;
1376         int i;
1377
1378         if (index == locked_page->index && end_index == index)
1379                 return;
1380
1381         while (nr_pages > 0) {
1382                 ret = find_get_pages_contig(inode->i_mapping, index,
1383                                      min_t(unsigned long, nr_pages,
1384                                      ARRAY_SIZE(pages)), pages);
1385                 for (i = 0; i < ret; i++) {
1386                         if (pages[i] != locked_page)
1387                                 unlock_page(pages[i]);
1388                         page_cache_release(pages[i]);
1389                 }
1390                 nr_pages -= ret;
1391                 index += ret;
1392                 cond_resched();
1393         }
1394 }
1395
1396 static noinline int lock_delalloc_pages(struct inode *inode,
1397                                         struct page *locked_page,
1398                                         u64 delalloc_start,
1399                                         u64 delalloc_end)
1400 {
1401         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1402         unsigned long start_index = index;
1403         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1404         unsigned long pages_locked = 0;
1405         struct page *pages[16];
1406         unsigned long nrpages;
1407         int ret;
1408         int i;
1409
1410         /* the caller is responsible for locking the start index */
1411         if (index == locked_page->index && index == end_index)
1412                 return 0;
1413
1414         /* skip the page at the start index */
1415         nrpages = end_index - index + 1;
1416         while (nrpages > 0) {
1417                 ret = find_get_pages_contig(inode->i_mapping, index,
1418                                      min_t(unsigned long,
1419                                      nrpages, ARRAY_SIZE(pages)), pages);
1420                 if (ret == 0) {
1421                         ret = -EAGAIN;
1422                         goto done;
1423                 }
1424                 /* now we have an array of pages, lock them all */
1425                 for (i = 0; i < ret; i++) {
1426                         /*
1427                          * the caller is taking responsibility for
1428                          * locked_page
1429                          */
1430                         if (pages[i] != locked_page) {
1431                                 lock_page(pages[i]);
1432                                 if (!PageDirty(pages[i]) ||
1433                                     pages[i]->mapping != inode->i_mapping) {
1434                                         ret = -EAGAIN;
1435                                         unlock_page(pages[i]);
1436                                         page_cache_release(pages[i]);
1437                                         goto done;
1438                                 }
1439                         }
1440                         page_cache_release(pages[i]);
1441                         pages_locked++;
1442                 }
1443                 nrpages -= ret;
1444                 index += ret;
1445                 cond_resched();
1446         }
1447         ret = 0;
1448 done:
1449         if (ret && pages_locked) {
1450                 __unlock_for_delalloc(inode, locked_page,
1451                               delalloc_start,
1452                               ((u64)(start_index + pages_locked - 1)) <<
1453                               PAGE_CACHE_SHIFT);
1454         }
1455         return ret;
1456 }
1457
1458 /*
1459  * find a contiguous range of bytes in the file marked as delalloc, not
1460  * more than 'max_bytes'.  start and end are used to return the range,
1461  *
1462  * 1 is returned if we find something, 0 if nothing was in the tree
1463  */
1464 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1465                                              struct extent_io_tree *tree,
1466                                              struct page *locked_page,
1467                                              u64 *start, u64 *end,
1468                                              u64 max_bytes)
1469 {
1470         u64 delalloc_start;
1471         u64 delalloc_end;
1472         u64 found;
1473         struct extent_state *cached_state = NULL;
1474         int ret;
1475         int loops = 0;
1476
1477 again:
1478         /* step one, find a bunch of delalloc bytes starting at start */
1479         delalloc_start = *start;
1480         delalloc_end = 0;
1481         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1482                                     max_bytes, &cached_state);
1483         if (!found || delalloc_end <= *start) {
1484                 *start = delalloc_start;
1485                 *end = delalloc_end;
1486                 free_extent_state(cached_state);
1487                 return found;
1488         }
1489
1490         /*
1491          * start comes from the offset of locked_page.  We have to lock
1492          * pages in order, so we can't process delalloc bytes before
1493          * locked_page
1494          */
1495         if (delalloc_start < *start)
1496                 delalloc_start = *start;
1497
1498         /*
1499          * make sure to limit the number of pages we try to lock down
1500          * if we're looping.
1501          */
1502         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1503                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1504
1505         /* step two, lock all the pages after the page that has start */
1506         ret = lock_delalloc_pages(inode, locked_page,
1507                                   delalloc_start, delalloc_end);
1508         if (ret == -EAGAIN) {
1509                 /* some of the pages are gone, lets avoid looping by
1510                  * shortening the size of the delalloc range we're searching
1511                  */
1512                 free_extent_state(cached_state);
1513                 if (!loops) {
1514                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1515                         max_bytes = PAGE_CACHE_SIZE - offset;
1516                         loops = 1;
1517                         goto again;
1518                 } else {
1519                         found = 0;
1520                         goto out_failed;
1521                 }
1522         }
1523         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1524
1525         /* step three, lock the state bits for the whole range */
1526         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1527
1528         /* then test to make sure it is all still delalloc */
1529         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1530                              EXTENT_DELALLOC, 1, cached_state);
1531         if (!ret) {
1532                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1533                                      &cached_state, GFP_NOFS);
1534                 __unlock_for_delalloc(inode, locked_page,
1535                               delalloc_start, delalloc_end);
1536                 cond_resched();
1537                 goto again;
1538         }
1539         free_extent_state(cached_state);
1540         *start = delalloc_start;
1541         *end = delalloc_end;
1542 out_failed:
1543         return found;
1544 }
1545
1546 int extent_clear_unlock_delalloc(struct inode *inode,
1547                                 struct extent_io_tree *tree,
1548                                 u64 start, u64 end, struct page *locked_page,
1549                                 unsigned long op)
1550 {
1551         int ret;
1552         struct page *pages[16];
1553         unsigned long index = start >> PAGE_CACHE_SHIFT;
1554         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1555         unsigned long nr_pages = end_index - index + 1;
1556         int i;
1557         int clear_bits = 0;
1558
1559         if (op & EXTENT_CLEAR_UNLOCK)
1560                 clear_bits |= EXTENT_LOCKED;
1561         if (op & EXTENT_CLEAR_DIRTY)
1562                 clear_bits |= EXTENT_DIRTY;
1563
1564         if (op & EXTENT_CLEAR_DELALLOC)
1565                 clear_bits |= EXTENT_DELALLOC;
1566
1567         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1568         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1569                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1570                     EXTENT_SET_PRIVATE2)))
1571                 return 0;
1572
1573         while (nr_pages > 0) {
1574                 ret = find_get_pages_contig(inode->i_mapping, index,
1575                                      min_t(unsigned long,
1576                                      nr_pages, ARRAY_SIZE(pages)), pages);
1577                 for (i = 0; i < ret; i++) {
1578
1579                         if (op & EXTENT_SET_PRIVATE2)
1580                                 SetPagePrivate2(pages[i]);
1581
1582                         if (pages[i] == locked_page) {
1583                                 page_cache_release(pages[i]);
1584                                 continue;
1585                         }
1586                         if (op & EXTENT_CLEAR_DIRTY)
1587                                 clear_page_dirty_for_io(pages[i]);
1588                         if (op & EXTENT_SET_WRITEBACK)
1589                                 set_page_writeback(pages[i]);
1590                         if (op & EXTENT_END_WRITEBACK)
1591                                 end_page_writeback(pages[i]);
1592                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1593                                 unlock_page(pages[i]);
1594                         page_cache_release(pages[i]);
1595                 }
1596                 nr_pages -= ret;
1597                 index += ret;
1598                 cond_resched();
1599         }
1600         return 0;
1601 }
1602
1603 /*
1604  * count the number of bytes in the tree that have a given bit(s)
1605  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1606  * cached.  The total number found is returned.
1607  */
1608 u64 count_range_bits(struct extent_io_tree *tree,
1609                      u64 *start, u64 search_end, u64 max_bytes,
1610                      unsigned long bits, int contig)
1611 {
1612         struct rb_node *node;
1613         struct extent_state *state;
1614         u64 cur_start = *start;
1615         u64 total_bytes = 0;
1616         u64 last = 0;
1617         int found = 0;
1618
1619         if (search_end <= cur_start) {
1620                 WARN_ON(1);
1621                 return 0;
1622         }
1623
1624         spin_lock(&tree->lock);
1625         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1626                 total_bytes = tree->dirty_bytes;
1627                 goto out;
1628         }
1629         /*
1630          * this search will find all the extents that end after
1631          * our range starts.
1632          */
1633         node = tree_search(tree, cur_start);
1634         if (!node)
1635                 goto out;
1636
1637         while (1) {
1638                 state = rb_entry(node, struct extent_state, rb_node);
1639                 if (state->start > search_end)
1640                         break;
1641                 if (contig && found && state->start > last + 1)
1642                         break;
1643                 if (state->end >= cur_start && (state->state & bits) == bits) {
1644                         total_bytes += min(search_end, state->end) + 1 -
1645                                        max(cur_start, state->start);
1646                         if (total_bytes >= max_bytes)
1647                                 break;
1648                         if (!found) {
1649                                 *start = max(cur_start, state->start);
1650                                 found = 1;
1651                         }
1652                         last = state->end;
1653                 } else if (contig && found) {
1654                         break;
1655                 }
1656                 node = rb_next(node);
1657                 if (!node)
1658                         break;
1659         }
1660 out:
1661         spin_unlock(&tree->lock);
1662         return total_bytes;
1663 }
1664
1665 /*
1666  * set the private field for a given byte offset in the tree.  If there isn't
1667  * an extent_state there already, this does nothing.
1668  */
1669 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1670 {
1671         struct rb_node *node;
1672         struct extent_state *state;
1673         int ret = 0;
1674
1675         spin_lock(&tree->lock);
1676         /*
1677          * this search will find all the extents that end after
1678          * our range starts.
1679          */
1680         node = tree_search(tree, start);
1681         if (!node) {
1682                 ret = -ENOENT;
1683                 goto out;
1684         }
1685         state = rb_entry(node, struct extent_state, rb_node);
1686         if (state->start != start) {
1687                 ret = -ENOENT;
1688                 goto out;
1689         }
1690         state->private = private;
1691 out:
1692         spin_unlock(&tree->lock);
1693         return ret;
1694 }
1695
1696 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1697 {
1698         struct rb_node *node;
1699         struct extent_state *state;
1700         int ret = 0;
1701
1702         spin_lock(&tree->lock);
1703         /*
1704          * this search will find all the extents that end after
1705          * our range starts.
1706          */
1707         node = tree_search(tree, start);
1708         if (!node) {
1709                 ret = -ENOENT;
1710                 goto out;
1711         }
1712         state = rb_entry(node, struct extent_state, rb_node);
1713         if (state->start != start) {
1714                 ret = -ENOENT;
1715                 goto out;
1716         }
1717         *private = state->private;
1718 out:
1719         spin_unlock(&tree->lock);
1720         return ret;
1721 }
1722
1723 /*
1724  * searches a range in the state tree for a given mask.
1725  * If 'filled' == 1, this returns 1 only if every extent in the tree
1726  * has the bits set.  Otherwise, 1 is returned if any bit in the
1727  * range is found set.
1728  */
1729 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1730                    int bits, int filled, struct extent_state *cached)
1731 {
1732         struct extent_state *state = NULL;
1733         struct rb_node *node;
1734         int bitset = 0;
1735
1736         spin_lock(&tree->lock);
1737         if (cached && cached->tree && cached->start <= start &&
1738             cached->end > start)
1739                 node = &cached->rb_node;
1740         else
1741                 node = tree_search(tree, start);
1742         while (node && start <= end) {
1743                 state = rb_entry(node, struct extent_state, rb_node);
1744
1745                 if (filled && state->start > start) {
1746                         bitset = 0;
1747                         break;
1748                 }
1749
1750                 if (state->start > end)
1751                         break;
1752
1753                 if (state->state & bits) {
1754                         bitset = 1;
1755                         if (!filled)
1756                                 break;
1757                 } else if (filled) {
1758                         bitset = 0;
1759                         break;
1760                 }
1761
1762                 if (state->end == (u64)-1)
1763                         break;
1764
1765                 start = state->end + 1;
1766                 if (start > end)
1767                         break;
1768                 node = rb_next(node);
1769                 if (!node) {
1770                         if (filled)
1771                                 bitset = 0;
1772                         break;
1773                 }
1774         }
1775         spin_unlock(&tree->lock);
1776         return bitset;
1777 }
1778
1779 /*
1780  * helper function to set a given page up to date if all the
1781  * extents in the tree for that page are up to date
1782  */
1783 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1784 {
1785         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1786         u64 end = start + PAGE_CACHE_SIZE - 1;
1787         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1788                 SetPageUptodate(page);
1789 }
1790
1791 /*
1792  * helper function to unlock a page if all the extents in the tree
1793  * for that page are unlocked
1794  */
1795 static void check_page_locked(struct extent_io_tree *tree, struct page *page)
1796 {
1797         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1798         u64 end = start + PAGE_CACHE_SIZE - 1;
1799         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1800                 unlock_page(page);
1801 }
1802
1803 /*
1804  * helper function to end page writeback if all the extents
1805  * in the tree for that page are done with writeback
1806  */
1807 static void check_page_writeback(struct extent_io_tree *tree,
1808                                  struct page *page)
1809 {
1810         end_page_writeback(page);
1811 }
1812
1813 /*
1814  * When IO fails, either with EIO or csum verification fails, we
1815  * try other mirrors that might have a good copy of the data.  This
1816  * io_failure_record is used to record state as we go through all the
1817  * mirrors.  If another mirror has good data, the page is set up to date
1818  * and things continue.  If a good mirror can't be found, the original
1819  * bio end_io callback is called to indicate things have failed.
1820  */
1821 struct io_failure_record {
1822         struct page *page;
1823         u64 start;
1824         u64 len;
1825         u64 logical;
1826         unsigned long bio_flags;
1827         int this_mirror;
1828         int failed_mirror;
1829         int in_validation;
1830 };
1831
1832 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1833                                 int did_repair)
1834 {
1835         int ret;
1836         int err = 0;
1837         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1838
1839         set_state_private(failure_tree, rec->start, 0);
1840         ret = clear_extent_bits(failure_tree, rec->start,
1841                                 rec->start + rec->len - 1,
1842                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1843         if (ret)
1844                 err = ret;
1845
1846         if (did_repair) {
1847                 ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1848                                         rec->start + rec->len - 1,
1849                                         EXTENT_DAMAGED, GFP_NOFS);
1850                 if (ret && !err)
1851                         err = ret;
1852         }
1853
1854         kfree(rec);
1855         return err;
1856 }
1857
1858 static void repair_io_failure_callback(struct bio *bio, int err)
1859 {
1860         complete(bio->bi_private);
1861 }
1862
1863 /*
1864  * this bypasses the standard btrfs submit functions deliberately, as
1865  * the standard behavior is to write all copies in a raid setup. here we only
1866  * want to write the one bad copy. so we do the mapping for ourselves and issue
1867  * submit_bio directly.
1868  * to avoid any synchonization issues, wait for the data after writing, which
1869  * actually prevents the read that triggered the error from finishing.
1870  * currently, there can be no more than two copies of every data bit. thus,
1871  * exactly one rewrite is required.
1872  */
1873 int repair_io_failure(struct btrfs_mapping_tree *map_tree, u64 start,
1874                         u64 length, u64 logical, struct page *page,
1875                         int mirror_num)
1876 {
1877         struct bio *bio;
1878         struct btrfs_device *dev;
1879         DECLARE_COMPLETION_ONSTACK(compl);
1880         u64 map_length = 0;
1881         u64 sector;
1882         struct btrfs_bio *bbio = NULL;
1883         int ret;
1884
1885         BUG_ON(!mirror_num);
1886
1887         bio = bio_alloc(GFP_NOFS, 1);
1888         if (!bio)
1889                 return -EIO;
1890         bio->bi_private = &compl;
1891         bio->bi_end_io = repair_io_failure_callback;
1892         bio->bi_size = 0;
1893         map_length = length;
1894
1895         ret = btrfs_map_block(map_tree, WRITE, logical,
1896                               &map_length, &bbio, mirror_num);
1897         if (ret) {
1898                 bio_put(bio);
1899                 return -EIO;
1900         }
1901         BUG_ON(mirror_num != bbio->mirror_num);
1902         sector = bbio->stripes[mirror_num-1].physical >> 9;
1903         bio->bi_sector = sector;
1904         dev = bbio->stripes[mirror_num-1].dev;
1905         kfree(bbio);
1906         if (!dev || !dev->bdev || !dev->writeable) {
1907                 bio_put(bio);
1908                 return -EIO;
1909         }
1910         bio->bi_bdev = dev->bdev;
1911         bio_add_page(bio, page, length, start-page_offset(page));
1912         btrfsic_submit_bio(WRITE_SYNC, bio);
1913         wait_for_completion(&compl);
1914
1915         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
1916                 /* try to remap that extent elsewhere? */
1917                 bio_put(bio);
1918                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
1919                 return -EIO;
1920         }
1921
1922         printk_ratelimited_in_rcu(KERN_INFO "btrfs read error corrected: ino %lu off %llu "
1923                       "(dev %s sector %llu)\n", page->mapping->host->i_ino,
1924                       start, rcu_str_deref(dev->name), sector);
1925
1926         bio_put(bio);
1927         return 0;
1928 }
1929
1930 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
1931                          int mirror_num)
1932 {
1933         struct btrfs_mapping_tree *map_tree = &root->fs_info->mapping_tree;
1934         u64 start = eb->start;
1935         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
1936         int ret = 0;
1937
1938         for (i = 0; i < num_pages; i++) {
1939                 struct page *p = extent_buffer_page(eb, i);
1940                 ret = repair_io_failure(map_tree, start, PAGE_CACHE_SIZE,
1941                                         start, p, mirror_num);
1942                 if (ret)
1943                         break;
1944                 start += PAGE_CACHE_SIZE;
1945         }
1946
1947         return ret;
1948 }
1949
1950 /*
1951  * each time an IO finishes, we do a fast check in the IO failure tree
1952  * to see if we need to process or clean up an io_failure_record
1953  */
1954 static int clean_io_failure(u64 start, struct page *page)
1955 {
1956         u64 private;
1957         u64 private_failure;
1958         struct io_failure_record *failrec;
1959         struct btrfs_mapping_tree *map_tree;
1960         struct extent_state *state;
1961         int num_copies;
1962         int did_repair = 0;
1963         int ret;
1964         struct inode *inode = page->mapping->host;
1965
1966         private = 0;
1967         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
1968                                 (u64)-1, 1, EXTENT_DIRTY, 0);
1969         if (!ret)
1970                 return 0;
1971
1972         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
1973                                 &private_failure);
1974         if (ret)
1975                 return 0;
1976
1977         failrec = (struct io_failure_record *)(unsigned long) private_failure;
1978         BUG_ON(!failrec->this_mirror);
1979
1980         if (failrec->in_validation) {
1981                 /* there was no real error, just free the record */
1982                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
1983                          failrec->start);
1984                 did_repair = 1;
1985                 goto out;
1986         }
1987
1988         spin_lock(&BTRFS_I(inode)->io_tree.lock);
1989         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
1990                                             failrec->start,
1991                                             EXTENT_LOCKED);
1992         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
1993
1994         if (state && state->start == failrec->start) {
1995                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
1996                 num_copies = btrfs_num_copies(map_tree, failrec->logical,
1997                                                 failrec->len);
1998                 if (num_copies > 1)  {
1999                         ret = repair_io_failure(map_tree, start, failrec->len,
2000                                                 failrec->logical, page,
2001                                                 failrec->failed_mirror);
2002                         did_repair = !ret;
2003                 }
2004         }
2005
2006 out:
2007         if (!ret)
2008                 ret = free_io_failure(inode, failrec, did_repair);
2009
2010         return ret;
2011 }
2012
2013 /*
2014  * this is a generic handler for readpage errors (default
2015  * readpage_io_failed_hook). if other copies exist, read those and write back
2016  * good data to the failed position. does not investigate in remapping the
2017  * failed extent elsewhere, hoping the device will be smart enough to do this as
2018  * needed
2019  */
2020
2021 static int bio_readpage_error(struct bio *failed_bio, struct page *page,
2022                                 u64 start, u64 end, int failed_mirror,
2023                                 struct extent_state *state)
2024 {
2025         struct io_failure_record *failrec = NULL;
2026         u64 private;
2027         struct extent_map *em;
2028         struct inode *inode = page->mapping->host;
2029         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2030         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2031         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2032         struct bio *bio;
2033         int num_copies;
2034         int ret;
2035         int read_mode;
2036         u64 logical;
2037
2038         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2039
2040         ret = get_state_private(failure_tree, start, &private);
2041         if (ret) {
2042                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2043                 if (!failrec)
2044                         return -ENOMEM;
2045                 failrec->start = start;
2046                 failrec->len = end - start + 1;
2047                 failrec->this_mirror = 0;
2048                 failrec->bio_flags = 0;
2049                 failrec->in_validation = 0;
2050
2051                 read_lock(&em_tree->lock);
2052                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2053                 if (!em) {
2054                         read_unlock(&em_tree->lock);
2055                         kfree(failrec);
2056                         return -EIO;
2057                 }
2058
2059                 if (em->start > start || em->start + em->len < start) {
2060                         free_extent_map(em);
2061                         em = NULL;
2062                 }
2063                 read_unlock(&em_tree->lock);
2064
2065                 if (!em || IS_ERR(em)) {
2066                         kfree(failrec);
2067                         return -EIO;
2068                 }
2069                 logical = start - em->start;
2070                 logical = em->block_start + logical;
2071                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2072                         logical = em->block_start;
2073                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2074                         extent_set_compress_type(&failrec->bio_flags,
2075                                                  em->compress_type);
2076                 }
2077                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2078                          "len=%llu\n", logical, start, failrec->len);
2079                 failrec->logical = logical;
2080                 free_extent_map(em);
2081
2082                 /* set the bits in the private failure tree */
2083                 ret = set_extent_bits(failure_tree, start, end,
2084                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2085                 if (ret >= 0)
2086                         ret = set_state_private(failure_tree, start,
2087                                                 (u64)(unsigned long)failrec);
2088                 /* set the bits in the inode's tree */
2089                 if (ret >= 0)
2090                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2091                                                 GFP_NOFS);
2092                 if (ret < 0) {
2093                         kfree(failrec);
2094                         return ret;
2095                 }
2096         } else {
2097                 failrec = (struct io_failure_record *)(unsigned long)private;
2098                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2099                          "start=%llu, len=%llu, validation=%d\n",
2100                          failrec->logical, failrec->start, failrec->len,
2101                          failrec->in_validation);
2102                 /*
2103                  * when data can be on disk more than twice, add to failrec here
2104                  * (e.g. with a list for failed_mirror) to make
2105                  * clean_io_failure() clean all those errors at once.
2106                  */
2107         }
2108         num_copies = btrfs_num_copies(
2109                               &BTRFS_I(inode)->root->fs_info->mapping_tree,
2110                               failrec->logical, failrec->len);
2111         if (num_copies == 1) {
2112                 /*
2113                  * we only have a single copy of the data, so don't bother with
2114                  * all the retry and error correction code that follows. no
2115                  * matter what the error is, it is very likely to persist.
2116                  */
2117                 pr_debug("bio_readpage_error: cannot repair, num_copies == 1. "
2118                          "state=%p, num_copies=%d, next_mirror %d, "
2119                          "failed_mirror %d\n", state, num_copies,
2120                          failrec->this_mirror, failed_mirror);
2121                 free_io_failure(inode, failrec, 0);
2122                 return -EIO;
2123         }
2124
2125         if (!state) {
2126                 spin_lock(&tree->lock);
2127                 state = find_first_extent_bit_state(tree, failrec->start,
2128                                                     EXTENT_LOCKED);
2129                 if (state && state->start != failrec->start)
2130                         state = NULL;
2131                 spin_unlock(&tree->lock);
2132         }
2133
2134         /*
2135          * there are two premises:
2136          *      a) deliver good data to the caller
2137          *      b) correct the bad sectors on disk
2138          */
2139         if (failed_bio->bi_vcnt > 1) {
2140                 /*
2141                  * to fulfill b), we need to know the exact failing sectors, as
2142                  * we don't want to rewrite any more than the failed ones. thus,
2143                  * we need separate read requests for the failed bio
2144                  *
2145                  * if the following BUG_ON triggers, our validation request got
2146                  * merged. we need separate requests for our algorithm to work.
2147                  */
2148                 BUG_ON(failrec->in_validation);
2149                 failrec->in_validation = 1;
2150                 failrec->this_mirror = failed_mirror;
2151                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2152         } else {
2153                 /*
2154                  * we're ready to fulfill a) and b) alongside. get a good copy
2155                  * of the failed sector and if we succeed, we have setup
2156                  * everything for repair_io_failure to do the rest for us.
2157                  */
2158                 if (failrec->in_validation) {
2159                         BUG_ON(failrec->this_mirror != failed_mirror);
2160                         failrec->in_validation = 0;
2161                         failrec->this_mirror = 0;
2162                 }
2163                 failrec->failed_mirror = failed_mirror;
2164                 failrec->this_mirror++;
2165                 if (failrec->this_mirror == failed_mirror)
2166                         failrec->this_mirror++;
2167                 read_mode = READ_SYNC;
2168         }
2169
2170         if (!state || failrec->this_mirror > num_copies) {
2171                 pr_debug("bio_readpage_error: (fail) state=%p, num_copies=%d, "
2172                          "next_mirror %d, failed_mirror %d\n", state,
2173                          num_copies, failrec->this_mirror, failed_mirror);
2174                 free_io_failure(inode, failrec, 0);
2175                 return -EIO;
2176         }
2177
2178         bio = bio_alloc(GFP_NOFS, 1);
2179         if (!bio) {
2180                 free_io_failure(inode, failrec, 0);
2181                 return -EIO;
2182         }
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) {
2219                 ClearPageUptodate(page);
2220                 SetPageError(page);
2221         }
2222         return 0;
2223 }
2224
2225 /*
2226  * after a writepage IO is done, we need to:
2227  * clear the uptodate bits on error
2228  * clear the writeback bits in the extent tree for this IO
2229  * end_page_writeback if the page has no more pending IO
2230  *
2231  * Scheduling is not allowed, so the extent state tree is expected
2232  * to have one and only one object corresponding to this IO.
2233  */
2234 static void end_bio_extent_writepage(struct bio *bio, int err)
2235 {
2236         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2237         struct extent_io_tree *tree;
2238         u64 start;
2239         u64 end;
2240         int whole_page;
2241
2242         do {
2243                 struct page *page = bvec->bv_page;
2244                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2245
2246                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2247                          bvec->bv_offset;
2248                 end = start + bvec->bv_len - 1;
2249
2250                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2251                         whole_page = 1;
2252                 else
2253                         whole_page = 0;
2254
2255                 if (--bvec >= bio->bi_io_vec)
2256                         prefetchw(&bvec->bv_page->flags);
2257
2258                 if (end_extent_writepage(page, err, start, end))
2259                         continue;
2260
2261                 if (whole_page)
2262                         end_page_writeback(page);
2263                 else
2264                         check_page_writeback(tree, page);
2265         } while (bvec >= bio->bi_io_vec);
2266
2267         bio_put(bio);
2268 }
2269
2270 /*
2271  * after a readpage IO is done, we need to:
2272  * clear the uptodate bits on error
2273  * set the uptodate bits if things worked
2274  * set the page up to date if all extents in the tree are uptodate
2275  * clear the lock bit in the extent tree
2276  * unlock the page if there are no other extents locked for it
2277  *
2278  * Scheduling is not allowed, so the extent state tree is expected
2279  * to have one and only one object corresponding to this IO.
2280  */
2281 static void end_bio_extent_readpage(struct bio *bio, int err)
2282 {
2283         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2284         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2285         struct bio_vec *bvec = bio->bi_io_vec;
2286         struct extent_io_tree *tree;
2287         u64 start;
2288         u64 end;
2289         int whole_page;
2290         int mirror;
2291         int ret;
2292
2293         if (err)
2294                 uptodate = 0;
2295
2296         do {
2297                 struct page *page = bvec->bv_page;
2298                 struct extent_state *cached = NULL;
2299                 struct extent_state *state;
2300
2301                 pr_debug("end_bio_extent_readpage: bi_vcnt=%d, idx=%d, err=%d, "
2302                          "mirror=%ld\n", bio->bi_vcnt, bio->bi_idx, err,
2303                          (long int)bio->bi_bdev);
2304                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2305
2306                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2307                         bvec->bv_offset;
2308                 end = start + bvec->bv_len - 1;
2309
2310                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2311                         whole_page = 1;
2312                 else
2313                         whole_page = 0;
2314
2315                 if (++bvec <= bvec_end)
2316                         prefetchw(&bvec->bv_page->flags);
2317
2318                 spin_lock(&tree->lock);
2319                 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
2320                 if (state && state->start == start) {
2321                         /*
2322                          * take a reference on the state, unlock will drop
2323                          * the ref
2324                          */
2325                         cache_state(state, &cached);
2326                 }
2327                 spin_unlock(&tree->lock);
2328
2329                 mirror = (int)(unsigned long)bio->bi_bdev;
2330                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
2331                         ret = tree->ops->readpage_end_io_hook(page, start, end,
2332                                                               state, mirror);
2333                         if (ret)
2334                                 uptodate = 0;
2335                         else
2336                                 clean_io_failure(start, page);
2337                 }
2338
2339                 if (!uptodate && tree->ops && tree->ops->readpage_io_failed_hook) {
2340                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2341                         if (!ret && !err &&
2342                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2343                                 uptodate = 1;
2344                 } else if (!uptodate) {
2345                         /*
2346                          * The generic bio_readpage_error handles errors the
2347                          * following way: If possible, new read requests are
2348                          * created and submitted and will end up in
2349                          * end_bio_extent_readpage as well (if we're lucky, not
2350                          * in the !uptodate case). In that case it returns 0 and
2351                          * we just go on with the next page in our bio. If it
2352                          * can't handle the error it will return -EIO and we
2353                          * remain responsible for that page.
2354                          */
2355                         ret = bio_readpage_error(bio, page, start, end, mirror, NULL);
2356                         if (ret == 0) {
2357                                 uptodate =
2358                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2359                                 if (err)
2360                                         uptodate = 0;
2361                                 uncache_state(&cached);
2362                                 continue;
2363                         }
2364                 }
2365
2366                 if (uptodate && tree->track_uptodate) {
2367                         set_extent_uptodate(tree, start, end, &cached,
2368                                             GFP_ATOMIC);
2369                 }
2370                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2371
2372                 if (whole_page) {
2373                         if (uptodate) {
2374                                 SetPageUptodate(page);
2375                         } else {
2376                                 ClearPageUptodate(page);
2377                                 SetPageError(page);
2378                         }
2379                         unlock_page(page);
2380                 } else {
2381                         if (uptodate) {
2382                                 check_page_uptodate(tree, page);
2383                         } else {
2384                                 ClearPageUptodate(page);
2385                                 SetPageError(page);
2386                         }
2387                         check_page_locked(tree, page);
2388                 }
2389         } while (bvec <= bvec_end);
2390
2391         bio_put(bio);
2392 }
2393
2394 struct bio *
2395 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2396                 gfp_t gfp_flags)
2397 {
2398         struct bio *bio;
2399
2400         bio = bio_alloc(gfp_flags, nr_vecs);
2401
2402         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2403                 while (!bio && (nr_vecs /= 2))
2404                         bio = bio_alloc(gfp_flags, nr_vecs);
2405         }
2406
2407         if (bio) {
2408                 bio->bi_size = 0;
2409                 bio->bi_bdev = bdev;
2410                 bio->bi_sector = first_sector;
2411         }
2412         return bio;
2413 }
2414
2415 /*
2416  * Since writes are async, they will only return -ENOMEM.
2417  * Reads can return the full range of I/O error conditions.
2418  */
2419 static int __must_check submit_one_bio(int rw, struct bio *bio,
2420                                        int mirror_num, unsigned long bio_flags)
2421 {
2422         int ret = 0;
2423         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2424         struct page *page = bvec->bv_page;
2425         struct extent_io_tree *tree = bio->bi_private;
2426         u64 start;
2427
2428         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
2429
2430         bio->bi_private = NULL;
2431
2432         bio_get(bio);
2433
2434         if (tree->ops && tree->ops->submit_bio_hook)
2435                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2436                                            mirror_num, bio_flags, start);
2437         else
2438                 btrfsic_submit_bio(rw, bio);
2439
2440         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2441                 ret = -EOPNOTSUPP;
2442         bio_put(bio);
2443         return ret;
2444 }
2445
2446 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2447                      unsigned long offset, size_t size, struct bio *bio,
2448                      unsigned long bio_flags)
2449 {
2450         int ret = 0;
2451         if (tree->ops && tree->ops->merge_bio_hook)
2452                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2453                                                 bio_flags);
2454         BUG_ON(ret < 0);
2455         return ret;
2456
2457 }
2458
2459 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2460                               struct page *page, sector_t sector,
2461                               size_t size, unsigned long offset,
2462                               struct block_device *bdev,
2463                               struct bio **bio_ret,
2464                               unsigned long max_pages,
2465                               bio_end_io_t end_io_func,
2466                               int mirror_num,
2467                               unsigned long prev_bio_flags,
2468                               unsigned long bio_flags)
2469 {
2470         int ret = 0;
2471         struct bio *bio;
2472         int nr;
2473         int contig = 0;
2474         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2475         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2476         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2477
2478         if (bio_ret && *bio_ret) {
2479                 bio = *bio_ret;
2480                 if (old_compressed)
2481                         contig = bio->bi_sector == sector;
2482                 else
2483                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
2484                                 sector;
2485
2486                 if (prev_bio_flags != bio_flags || !contig ||
2487                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2488                     bio_add_page(bio, page, page_size, offset) < page_size) {
2489                         ret = submit_one_bio(rw, bio, mirror_num,
2490                                              prev_bio_flags);
2491                         if (ret < 0)
2492                                 return ret;
2493                         bio = NULL;
2494                 } else {
2495                         return 0;
2496                 }
2497         }
2498         if (this_compressed)
2499                 nr = BIO_MAX_PAGES;
2500         else
2501                 nr = bio_get_nr_vecs(bdev);
2502
2503         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2504         if (!bio)
2505                 return -ENOMEM;
2506
2507         bio_add_page(bio, page, page_size, offset);
2508         bio->bi_end_io = end_io_func;
2509         bio->bi_private = tree;
2510
2511         if (bio_ret)
2512                 *bio_ret = bio;
2513         else
2514                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2515
2516         return ret;
2517 }
2518
2519 void attach_extent_buffer_page(struct extent_buffer *eb, struct page *page)
2520 {
2521         if (!PagePrivate(page)) {
2522                 SetPagePrivate(page);
2523                 page_cache_get(page);
2524                 set_page_private(page, (unsigned long)eb);
2525         } else {
2526                 WARN_ON(page->private != (unsigned long)eb);
2527         }
2528 }
2529
2530 void set_page_extent_mapped(struct page *page)
2531 {
2532         if (!PagePrivate(page)) {
2533                 SetPagePrivate(page);
2534                 page_cache_get(page);
2535                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2536         }
2537 }
2538
2539 /*
2540  * basic readpage implementation.  Locked extent state structs are inserted
2541  * into the tree that are removed when the IO is done (by the end_io
2542  * handlers)
2543  * XXX JDM: This needs looking at to ensure proper page locking
2544  */
2545 static int __extent_read_full_page(struct extent_io_tree *tree,
2546                                    struct page *page,
2547                                    get_extent_t *get_extent,
2548                                    struct bio **bio, int mirror_num,
2549                                    unsigned long *bio_flags)
2550 {
2551         struct inode *inode = page->mapping->host;
2552         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2553         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2554         u64 end;
2555         u64 cur = start;
2556         u64 extent_offset;
2557         u64 last_byte = i_size_read(inode);
2558         u64 block_start;
2559         u64 cur_end;
2560         sector_t sector;
2561         struct extent_map *em;
2562         struct block_device *bdev;
2563         struct btrfs_ordered_extent *ordered;
2564         int ret;
2565         int nr = 0;
2566         size_t pg_offset = 0;
2567         size_t iosize;
2568         size_t disk_io_size;
2569         size_t blocksize = inode->i_sb->s_blocksize;
2570         unsigned long this_bio_flag = 0;
2571
2572         set_page_extent_mapped(page);
2573
2574         if (!PageUptodate(page)) {
2575                 if (cleancache_get_page(page) == 0) {
2576                         BUG_ON(blocksize != PAGE_SIZE);
2577                         goto out;
2578                 }
2579         }
2580
2581         end = page_end;
2582         while (1) {
2583                 lock_extent(tree, start, end);
2584                 ordered = btrfs_lookup_ordered_extent(inode, start);
2585                 if (!ordered)
2586                         break;
2587                 unlock_extent(tree, start, end);
2588                 btrfs_start_ordered_extent(inode, ordered, 1);
2589                 btrfs_put_ordered_extent(ordered);
2590         }
2591
2592         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2593                 char *userpage;
2594                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2595
2596                 if (zero_offset) {
2597                         iosize = PAGE_CACHE_SIZE - zero_offset;
2598                         userpage = kmap_atomic(page);
2599                         memset(userpage + zero_offset, 0, iosize);
2600                         flush_dcache_page(page);
2601                         kunmap_atomic(userpage);
2602                 }
2603         }
2604         while (cur <= end) {
2605                 if (cur >= last_byte) {
2606                         char *userpage;
2607                         struct extent_state *cached = NULL;
2608
2609                         iosize = PAGE_CACHE_SIZE - pg_offset;
2610                         userpage = kmap_atomic(page);
2611                         memset(userpage + pg_offset, 0, iosize);
2612                         flush_dcache_page(page);
2613                         kunmap_atomic(userpage);
2614                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2615                                             &cached, GFP_NOFS);
2616                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2617                                              &cached, GFP_NOFS);
2618                         break;
2619                 }
2620                 em = get_extent(inode, page, pg_offset, cur,
2621                                 end - cur + 1, 0);
2622                 if (IS_ERR_OR_NULL(em)) {
2623                         SetPageError(page);
2624                         unlock_extent(tree, cur, end);
2625                         break;
2626                 }
2627                 extent_offset = cur - em->start;
2628                 BUG_ON(extent_map_end(em) <= cur);
2629                 BUG_ON(end < cur);
2630
2631                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2632                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2633                         extent_set_compress_type(&this_bio_flag,
2634                                                  em->compress_type);
2635                 }
2636
2637                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2638                 cur_end = min(extent_map_end(em) - 1, end);
2639                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2640                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2641                         disk_io_size = em->block_len;
2642                         sector = em->block_start >> 9;
2643                 } else {
2644                         sector = (em->block_start + extent_offset) >> 9;
2645                         disk_io_size = iosize;
2646                 }
2647                 bdev = em->bdev;
2648                 block_start = em->block_start;
2649                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2650                         block_start = EXTENT_MAP_HOLE;
2651                 free_extent_map(em);
2652                 em = NULL;
2653
2654                 /* we've found a hole, just zero and go on */
2655                 if (block_start == EXTENT_MAP_HOLE) {
2656                         char *userpage;
2657                         struct extent_state *cached = NULL;
2658
2659                         userpage = kmap_atomic(page);
2660                         memset(userpage + pg_offset, 0, iosize);
2661                         flush_dcache_page(page);
2662                         kunmap_atomic(userpage);
2663
2664                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2665                                             &cached, GFP_NOFS);
2666                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2667                                              &cached, GFP_NOFS);
2668                         cur = cur + iosize;
2669                         pg_offset += iosize;
2670                         continue;
2671                 }
2672                 /* the get_extent function already copied into the page */
2673                 if (test_range_bit(tree, cur, cur_end,
2674                                    EXTENT_UPTODATE, 1, NULL)) {
2675                         check_page_uptodate(tree, page);
2676                         unlock_extent(tree, cur, cur + iosize - 1);
2677                         cur = cur + iosize;
2678                         pg_offset += iosize;
2679                         continue;
2680                 }
2681                 /* we have an inline extent but it didn't get marked up
2682                  * to date.  Error out
2683                  */
2684                 if (block_start == EXTENT_MAP_INLINE) {
2685                         SetPageError(page);
2686                         unlock_extent(tree, cur, cur + iosize - 1);
2687                         cur = cur + iosize;
2688                         pg_offset += iosize;
2689                         continue;
2690                 }
2691
2692                 ret = 0;
2693                 if (tree->ops && tree->ops->readpage_io_hook) {
2694                         ret = tree->ops->readpage_io_hook(page, cur,
2695                                                           cur + iosize - 1);
2696                 }
2697                 if (!ret) {
2698                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2699                         pnr -= page->index;
2700                         ret = submit_extent_page(READ, tree, page,
2701                                          sector, disk_io_size, pg_offset,
2702                                          bdev, bio, pnr,
2703                                          end_bio_extent_readpage, mirror_num,
2704                                          *bio_flags,
2705                                          this_bio_flag);
2706                         BUG_ON(ret == -ENOMEM);
2707                         nr++;
2708                         *bio_flags = this_bio_flag;
2709                 }
2710                 if (ret)
2711                         SetPageError(page);
2712                 cur = cur + iosize;
2713                 pg_offset += iosize;
2714         }
2715 out:
2716         if (!nr) {
2717                 if (!PageError(page))
2718                         SetPageUptodate(page);
2719                 unlock_page(page);
2720         }
2721         return 0;
2722 }
2723
2724 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2725                             get_extent_t *get_extent, int mirror_num)
2726 {
2727         struct bio *bio = NULL;
2728         unsigned long bio_flags = 0;
2729         int ret;
2730
2731         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
2732                                       &bio_flags);
2733         if (bio)
2734                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
2735         return ret;
2736 }
2737
2738 static noinline void update_nr_written(struct page *page,
2739                                       struct writeback_control *wbc,
2740                                       unsigned long nr_written)
2741 {
2742         wbc->nr_to_write -= nr_written;
2743         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2744             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2745                 page->mapping->writeback_index = page->index + nr_written;
2746 }
2747
2748 /*
2749  * the writepage semantics are similar to regular writepage.  extent
2750  * records are inserted to lock ranges in the tree, and as dirty areas
2751  * are found, they are marked writeback.  Then the lock bits are removed
2752  * and the end_io handler clears the writeback ranges
2753  */
2754 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2755                               void *data)
2756 {
2757         struct inode *inode = page->mapping->host;
2758         struct extent_page_data *epd = data;
2759         struct extent_io_tree *tree = epd->tree;
2760         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2761         u64 delalloc_start;
2762         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2763         u64 end;
2764         u64 cur = start;
2765         u64 extent_offset;
2766         u64 last_byte = i_size_read(inode);
2767         u64 block_start;
2768         u64 iosize;
2769         sector_t sector;
2770         struct extent_state *cached_state = NULL;
2771         struct extent_map *em;
2772         struct block_device *bdev;
2773         int ret;
2774         int nr = 0;
2775         size_t pg_offset = 0;
2776         size_t blocksize;
2777         loff_t i_size = i_size_read(inode);
2778         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2779         u64 nr_delalloc;
2780         u64 delalloc_end;
2781         int page_started;
2782         int compressed;
2783         int write_flags;
2784         unsigned long nr_written = 0;
2785         bool fill_delalloc = true;
2786
2787         if (wbc->sync_mode == WB_SYNC_ALL)
2788                 write_flags = WRITE_SYNC;
2789         else
2790                 write_flags = WRITE;
2791
2792         trace___extent_writepage(page, inode, wbc);
2793
2794         WARN_ON(!PageLocked(page));
2795
2796         ClearPageError(page);
2797
2798         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2799         if (page->index > end_index ||
2800            (page->index == end_index && !pg_offset)) {
2801                 page->mapping->a_ops->invalidatepage(page, 0);
2802                 unlock_page(page);
2803                 return 0;
2804         }
2805
2806         if (page->index == end_index) {
2807                 char *userpage;
2808
2809                 userpage = kmap_atomic(page);
2810                 memset(userpage + pg_offset, 0,
2811                        PAGE_CACHE_SIZE - pg_offset);
2812                 kunmap_atomic(userpage);
2813                 flush_dcache_page(page);
2814         }
2815         pg_offset = 0;
2816
2817         set_page_extent_mapped(page);
2818
2819         if (!tree->ops || !tree->ops->fill_delalloc)
2820                 fill_delalloc = false;
2821
2822         delalloc_start = start;
2823         delalloc_end = 0;
2824         page_started = 0;
2825         if (!epd->extent_locked && fill_delalloc) {
2826                 u64 delalloc_to_write = 0;
2827                 /*
2828                  * make sure the wbc mapping index is at least updated
2829                  * to this page.
2830                  */
2831                 update_nr_written(page, wbc, 0);
2832
2833                 while (delalloc_end < page_end) {
2834                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2835                                                        page,
2836                                                        &delalloc_start,
2837                                                        &delalloc_end,
2838                                                        128 * 1024 * 1024);
2839                         if (nr_delalloc == 0) {
2840                                 delalloc_start = delalloc_end + 1;
2841                                 continue;
2842                         }
2843                         ret = tree->ops->fill_delalloc(inode, page,
2844                                                        delalloc_start,
2845                                                        delalloc_end,
2846                                                        &page_started,
2847                                                        &nr_written);
2848                         /* File system has been set read-only */
2849                         if (ret) {
2850                                 SetPageError(page);
2851                                 goto done;
2852                         }
2853                         /*
2854                          * delalloc_end is already one less than the total
2855                          * length, so we don't subtract one from
2856                          * PAGE_CACHE_SIZE
2857                          */
2858                         delalloc_to_write += (delalloc_end - delalloc_start +
2859                                               PAGE_CACHE_SIZE) >>
2860                                               PAGE_CACHE_SHIFT;
2861                         delalloc_start = delalloc_end + 1;
2862                 }
2863                 if (wbc->nr_to_write < delalloc_to_write) {
2864                         int thresh = 8192;
2865
2866                         if (delalloc_to_write < thresh * 2)
2867                                 thresh = delalloc_to_write;
2868                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2869                                                  thresh);
2870                 }
2871
2872                 /* did the fill delalloc function already unlock and start
2873                  * the IO?
2874                  */
2875                 if (page_started) {
2876                         ret = 0;
2877                         /*
2878                          * we've unlocked the page, so we can't update
2879                          * the mapping's writeback index, just update
2880                          * nr_to_write.
2881                          */
2882                         wbc->nr_to_write -= nr_written;
2883                         goto done_unlocked;
2884                 }
2885         }
2886         if (tree->ops && tree->ops->writepage_start_hook) {
2887                 ret = tree->ops->writepage_start_hook(page, start,
2888                                                       page_end);
2889                 if (ret) {
2890                         /* Fixup worker will requeue */
2891                         if (ret == -EBUSY)
2892                                 wbc->pages_skipped++;
2893                         else
2894                                 redirty_page_for_writepage(wbc, page);
2895                         update_nr_written(page, wbc, nr_written);
2896                         unlock_page(page);
2897                         ret = 0;
2898                         goto done_unlocked;
2899                 }
2900         }
2901
2902         /*
2903          * we don't want to touch the inode after unlocking the page,
2904          * so we update the mapping writeback index now
2905          */
2906         update_nr_written(page, wbc, nr_written + 1);
2907
2908         end = page_end;
2909         if (last_byte <= start) {
2910                 if (tree->ops && tree->ops->writepage_end_io_hook)
2911                         tree->ops->writepage_end_io_hook(page, start,
2912                                                          page_end, NULL, 1);
2913                 goto done;
2914         }
2915
2916         blocksize = inode->i_sb->s_blocksize;
2917
2918         while (cur <= end) {
2919                 if (cur >= last_byte) {
2920                         if (tree->ops && tree->ops->writepage_end_io_hook)
2921                                 tree->ops->writepage_end_io_hook(page, cur,
2922                                                          page_end, NULL, 1);
2923                         break;
2924                 }
2925                 em = epd->get_extent(inode, page, pg_offset, cur,
2926                                      end - cur + 1, 1);
2927                 if (IS_ERR_OR_NULL(em)) {
2928                         SetPageError(page);
2929                         break;
2930                 }
2931
2932                 extent_offset = cur - em->start;
2933                 BUG_ON(extent_map_end(em) <= cur);
2934                 BUG_ON(end < cur);
2935                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2936                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2937                 sector = (em->block_start + extent_offset) >> 9;
2938                 bdev = em->bdev;
2939                 block_start = em->block_start;
2940                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2941                 free_extent_map(em);
2942                 em = NULL;
2943
2944                 /*
2945                  * compressed and inline extents are written through other
2946                  * paths in the FS
2947                  */
2948                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2949                     block_start == EXTENT_MAP_INLINE) {
2950                         /*
2951                          * end_io notification does not happen here for
2952                          * compressed extents
2953                          */
2954                         if (!compressed && tree->ops &&
2955                             tree->ops->writepage_end_io_hook)
2956                                 tree->ops->writepage_end_io_hook(page, cur,
2957                                                          cur + iosize - 1,
2958                                                          NULL, 1);
2959                         else if (compressed) {
2960                                 /* we don't want to end_page_writeback on
2961                                  * a compressed extent.  this happens
2962                                  * elsewhere
2963                                  */
2964                                 nr++;
2965                         }
2966
2967                         cur += iosize;
2968                         pg_offset += iosize;
2969                         continue;
2970                 }
2971                 /* leave this out until we have a page_mkwrite call */
2972                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2973                                    EXTENT_DIRTY, 0, NULL)) {
2974                         cur = cur + iosize;
2975                         pg_offset += iosize;
2976                         continue;
2977                 }
2978
2979                 if (tree->ops && tree->ops->writepage_io_hook) {
2980                         ret = tree->ops->writepage_io_hook(page, cur,
2981                                                 cur + iosize - 1);
2982                 } else {
2983                         ret = 0;
2984                 }
2985                 if (ret) {
2986                         SetPageError(page);
2987                 } else {
2988                         unsigned long max_nr = end_index + 1;
2989
2990                         set_range_writeback(tree, cur, cur + iosize - 1);
2991                         if (!PageWriteback(page)) {
2992                                 printk(KERN_ERR "btrfs warning page %lu not "
2993                                        "writeback, cur %llu end %llu\n",
2994                                        page->index, (unsigned long long)cur,
2995                                        (unsigned long long)end);
2996                         }
2997
2998                         ret = submit_extent_page(write_flags, tree, page,
2999                                                  sector, iosize, pg_offset,
3000                                                  bdev, &epd->bio, max_nr,
3001                                                  end_bio_extent_writepage,
3002                                                  0, 0, 0);
3003                         if (ret)
3004                                 SetPageError(page);
3005                 }
3006                 cur = cur + iosize;
3007                 pg_offset += iosize;
3008                 nr++;
3009         }
3010 done:
3011         if (nr == 0) {
3012                 /* make sure the mapping tag for page dirty gets cleared */
3013                 set_page_writeback(page);
3014                 end_page_writeback(page);
3015         }
3016         unlock_page(page);
3017
3018 done_unlocked:
3019
3020         /* drop our reference on any cached states */
3021         free_extent_state(cached_state);
3022         return 0;
3023 }
3024
3025 static int eb_wait(void *word)
3026 {
3027         io_schedule();
3028         return 0;
3029 }
3030
3031 static void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3032 {
3033         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3034                     TASK_UNINTERRUPTIBLE);
3035 }
3036
3037 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3038                                      struct btrfs_fs_info *fs_info,
3039                                      struct extent_page_data *epd)
3040 {
3041         unsigned long i, num_pages;
3042         int flush = 0;
3043         int ret = 0;
3044
3045         if (!btrfs_try_tree_write_lock(eb)) {
3046                 flush = 1;
3047                 flush_write_bio(epd);
3048                 btrfs_tree_lock(eb);
3049         }
3050
3051         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3052                 btrfs_tree_unlock(eb);
3053                 if (!epd->sync_io)
3054                         return 0;
3055                 if (!flush) {
3056                         flush_write_bio(epd);
3057                         flush = 1;
3058                 }
3059                 while (1) {
3060                         wait_on_extent_buffer_writeback(eb);
3061                         btrfs_tree_lock(eb);
3062                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3063                                 break;
3064                         btrfs_tree_unlock(eb);
3065                 }
3066         }
3067
3068         /*
3069          * We need to do this to prevent races in people who check if the eb is
3070          * under IO since we can end up having no IO bits set for a short period
3071          * of time.
3072          */
3073         spin_lock(&eb->refs_lock);
3074         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3075                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3076                 spin_unlock(&eb->refs_lock);
3077                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3078                 spin_lock(&fs_info->delalloc_lock);
3079                 if (fs_info->dirty_metadata_bytes >= eb->len)
3080                         fs_info->dirty_metadata_bytes -= eb->len;
3081                 else
3082                         WARN_ON(1);
3083                 spin_unlock(&fs_info->delalloc_lock);
3084                 ret = 1;
3085         } else {
3086                 spin_unlock(&eb->refs_lock);
3087         }
3088
3089         btrfs_tree_unlock(eb);
3090
3091         if (!ret)
3092                 return ret;
3093
3094         num_pages = num_extent_pages(eb->start, eb->len);
3095         for (i = 0; i < num_pages; i++) {
3096                 struct page *p = extent_buffer_page(eb, i);
3097
3098                 if (!trylock_page(p)) {
3099                         if (!flush) {
3100                                 flush_write_bio(epd);
3101                                 flush = 1;
3102                         }
3103                         lock_page(p);
3104                 }
3105         }
3106
3107         return ret;
3108 }
3109
3110 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3111 {
3112         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3113         smp_mb__after_clear_bit();
3114         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3115 }
3116
3117 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3118 {
3119         int uptodate = err == 0;
3120         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
3121         struct extent_buffer *eb;
3122         int done;
3123
3124         do {
3125                 struct page *page = bvec->bv_page;
3126
3127                 bvec--;
3128                 eb = (struct extent_buffer *)page->private;
3129                 BUG_ON(!eb);
3130                 done = atomic_dec_and_test(&eb->io_pages);
3131
3132                 if (!uptodate || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3133                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3134                         ClearPageUptodate(page);
3135                         SetPageError(page);
3136                 }
3137
3138                 end_page_writeback(page);
3139
3140                 if (!done)
3141                         continue;
3142
3143                 end_extent_buffer_writeback(eb);
3144         } while (bvec >= bio->bi_io_vec);
3145
3146         bio_put(bio);
3147
3148 }
3149
3150 static int write_one_eb(struct extent_buffer *eb,
3151                         struct btrfs_fs_info *fs_info,
3152                         struct writeback_control *wbc,
3153                         struct extent_page_data *epd)
3154 {
3155         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3156         u64 offset = eb->start;
3157         unsigned long i, num_pages;
3158         int rw = (epd->sync_io ? WRITE_SYNC : WRITE);
3159         int ret = 0;
3160
3161         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3162         num_pages = num_extent_pages(eb->start, eb->len);
3163         atomic_set(&eb->io_pages, num_pages);
3164         for (i = 0; i < num_pages; i++) {
3165                 struct page *p = extent_buffer_page(eb, i);
3166
3167                 clear_page_dirty_for_io(p);
3168                 set_page_writeback(p);
3169                 ret = submit_extent_page(rw, eb->tree, p, offset >> 9,
3170                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3171                                          -1, end_bio_extent_buffer_writepage,
3172                                          0, 0, 0);
3173                 if (ret) {
3174                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3175                         SetPageError(p);
3176                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3177                                 end_extent_buffer_writeback(eb);
3178                         ret = -EIO;
3179                         break;
3180                 }
3181                 offset += PAGE_CACHE_SIZE;
3182                 update_nr_written(p, wbc, 1);
3183                 unlock_page(p);
3184         }
3185
3186         if (unlikely(ret)) {
3187                 for (; i < num_pages; i++) {
3188                         struct page *p = extent_buffer_page(eb, i);
3189                         unlock_page(p);
3190                 }
3191         }
3192
3193         return ret;
3194 }
3195
3196 int btree_write_cache_pages(struct address_space *mapping,
3197                                    struct writeback_control *wbc)
3198 {
3199         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3200         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3201         struct extent_buffer *eb, *prev_eb = NULL;
3202         struct extent_page_data epd = {
3203                 .bio = NULL,
3204                 .tree = tree,
3205                 .extent_locked = 0,
3206                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3207         };
3208         int ret = 0;
3209         int done = 0;
3210         int nr_to_write_done = 0;
3211         struct pagevec pvec;
3212         int nr_pages;
3213         pgoff_t index;
3214         pgoff_t end;            /* Inclusive */
3215         int scanned = 0;
3216         int tag;
3217
3218         pagevec_init(&pvec, 0);
3219         if (wbc->range_cyclic) {
3220                 index = mapping->writeback_index; /* Start from prev offset */
3221                 end = -1;
3222         } else {
3223                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3224                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3225                 scanned = 1;
3226         }
3227         if (wbc->sync_mode == WB_SYNC_ALL)
3228                 tag = PAGECACHE_TAG_TOWRITE;
3229         else
3230                 tag = PAGECACHE_TAG_DIRTY;
3231 retry:
3232         if (wbc->sync_mode == WB_SYNC_ALL)
3233                 tag_pages_for_writeback(mapping, index, end);
3234         while (!done && !nr_to_write_done && (index <= end) &&
3235                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3236                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3237                 unsigned i;
3238
3239                 scanned = 1;
3240                 for (i = 0; i < nr_pages; i++) {
3241                         struct page *page = pvec.pages[i];
3242
3243                         if (!PagePrivate(page))
3244                                 continue;
3245
3246                         if (!wbc->range_cyclic && page->index > end) {
3247                                 done = 1;
3248                                 break;
3249                         }
3250
3251                         eb = (struct extent_buffer *)page->private;
3252                         if (!eb) {
3253                                 WARN_ON(1);
3254                                 continue;
3255                         }
3256
3257                         if (eb == prev_eb)
3258                                 continue;
3259
3260                         if (!atomic_inc_not_zero(&eb->refs)) {
3261                                 WARN_ON(1);
3262                                 continue;
3263                         }
3264
3265                         prev_eb = eb;
3266                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3267                         if (!ret) {
3268                                 free_extent_buffer(eb);
3269                                 continue;
3270                         }
3271
3272                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3273                         if (ret) {
3274                                 done = 1;
3275                                 free_extent_buffer(eb);
3276                                 break;
3277                         }
3278                         free_extent_buffer(eb);
3279
3280                         /*
3281                          * the filesystem may choose to bump up nr_to_write.
3282                          * We have to make sure to honor the new nr_to_write
3283                          * at any time
3284                          */
3285                         nr_to_write_done = wbc->nr_to_write <= 0;
3286                 }
3287                 pagevec_release(&pvec);
3288                 cond_resched();
3289         }
3290         if (!scanned && !done) {
3291                 /*
3292                  * We hit the last page and there is more work to be done: wrap
3293                  * back to the start of the file
3294                  */
3295                 scanned = 1;
3296                 index = 0;
3297                 goto retry;
3298         }
3299         flush_write_bio(&epd);
3300         return ret;
3301 }
3302
3303 /**
3304  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3305  * @mapping: address space structure to write
3306  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3307  * @writepage: function called for each page
3308  * @data: data passed to writepage function
3309  *
3310  * If a page is already under I/O, write_cache_pages() skips it, even
3311  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3312  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3313  * and msync() need to guarantee that all the data which was dirty at the time
3314  * the call was made get new I/O started against them.  If wbc->sync_mode is
3315  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3316  * existing IO to complete.
3317  */
3318 static int extent_write_cache_pages(struct extent_io_tree *tree,
3319                              struct address_space *mapping,
3320                              struct writeback_control *wbc,
3321                              writepage_t writepage, void *data,
3322                              void (*flush_fn)(void *))
3323 {
3324         struct inode *inode = mapping->host;
3325         int ret = 0;
3326         int done = 0;
3327         int nr_to_write_done = 0;
3328         struct pagevec pvec;
3329         int nr_pages;
3330         pgoff_t index;
3331         pgoff_t end;            /* Inclusive */
3332         int scanned = 0;
3333         int tag;
3334
3335         /*
3336          * We have to hold onto the inode so that ordered extents can do their
3337          * work when the IO finishes.  The alternative to this is failing to add
3338          * an ordered extent if the igrab() fails there and that is a huge pain
3339          * to deal with, so instead just hold onto the inode throughout the
3340          * writepages operation.  If it fails here we are freeing up the inode
3341          * anyway and we'd rather not waste our time writing out stuff that is
3342          * going to be truncated anyway.
3343          */
3344         if (!igrab(inode))
3345                 return 0;
3346
3347         pagevec_init(&pvec, 0);
3348         if (wbc->range_cyclic) {
3349                 index = mapping->writeback_index; /* Start from prev offset */
3350                 end = -1;
3351         } else {
3352                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3353                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3354                 scanned = 1;
3355         }
3356         if (wbc->sync_mode == WB_SYNC_ALL)
3357                 tag = PAGECACHE_TAG_TOWRITE;
3358         else
3359                 tag = PAGECACHE_TAG_DIRTY;
3360 retry:
3361         if (wbc->sync_mode == WB_SYNC_ALL)
3362                 tag_pages_for_writeback(mapping, index, end);
3363         while (!done && !nr_to_write_done && (index <= end) &&
3364                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3365                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3366                 unsigned i;
3367
3368                 scanned = 1;
3369                 for (i = 0; i < nr_pages; i++) {
3370                         struct page *page = pvec.pages[i];
3371
3372                         /*
3373                          * At this point we hold neither mapping->tree_lock nor
3374                          * lock on the page itself: the page may be truncated or
3375                          * invalidated (changing page->mapping to NULL), or even
3376                          * swizzled back from swapper_space to tmpfs file
3377                          * mapping
3378                          */
3379                         if (tree->ops &&
3380                             tree->ops->write_cache_pages_lock_hook) {
3381                                 tree->ops->write_cache_pages_lock_hook(page,
3382                                                                data, flush_fn);
3383                         } else {
3384                                 if (!trylock_page(page)) {
3385                                         flush_fn(data);
3386                                         lock_page(page);
3387                                 }
3388                         }
3389
3390                         if (unlikely(page->mapping != mapping)) {
3391                                 unlock_page(page);
3392                                 continue;
3393                         }
3394
3395                         if (!wbc->range_cyclic && page->index > end) {
3396                                 done = 1;
3397                                 unlock_page(page);
3398                                 continue;
3399                         }
3400
3401                         if (wbc->sync_mode != WB_SYNC_NONE) {
3402                                 if (PageWriteback(page))
3403                                         flush_fn(data);
3404                                 wait_on_page_writeback(page);
3405                         }
3406
3407                         if (PageWriteback(page) ||
3408                             !clear_page_dirty_for_io(page)) {
3409                                 unlock_page(page);
3410                                 continue;
3411                         }
3412
3413                         ret = (*writepage)(page, wbc, data);
3414
3415                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3416                                 unlock_page(page);
3417                                 ret = 0;
3418                         }
3419                         if (ret)
3420                                 done = 1;
3421
3422                         /*
3423                          * the filesystem may choose to bump up nr_to_write.
3424                          * We have to make sure to honor the new nr_to_write
3425                          * at any time
3426                          */
3427                         nr_to_write_done = wbc->nr_to_write <= 0;
3428                 }
3429                 pagevec_release(&pvec);
3430                 cond_resched();
3431         }
3432         if (!scanned && !done) {
3433                 /*
3434                  * We hit the last page and there is more work to be done: wrap
3435                  * back to the start of the file
3436                  */
3437                 scanned = 1;
3438                 index = 0;
3439                 goto retry;
3440         }
3441         btrfs_add_delayed_iput(inode);
3442         return ret;
3443 }
3444
3445 static void flush_epd_write_bio(struct extent_page_data *epd)
3446 {
3447         if (epd->bio) {
3448                 int rw = WRITE;
3449                 int ret;
3450
3451                 if (epd->sync_io)
3452                         rw = WRITE_SYNC;
3453
3454                 ret = submit_one_bio(rw, epd->bio, 0, 0);
3455                 BUG_ON(ret < 0); /* -ENOMEM */
3456                 epd->bio = NULL;
3457         }
3458 }
3459
3460 static noinline void flush_write_bio(void *data)
3461 {
3462         struct extent_page_data *epd = data;
3463         flush_epd_write_bio(epd);
3464 }
3465
3466 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3467                           get_extent_t *get_extent,
3468                           struct writeback_control *wbc)
3469 {
3470         int ret;
3471         struct extent_page_data epd = {
3472                 .bio = NULL,
3473                 .tree = tree,
3474                 .get_extent = get_extent,
3475                 .extent_locked = 0,
3476                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3477         };
3478
3479         ret = __extent_writepage(page, wbc, &epd);
3480
3481         flush_epd_write_bio(&epd);
3482         return ret;
3483 }
3484
3485 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3486                               u64 start, u64 end, get_extent_t *get_extent,
3487                               int mode)
3488 {
3489         int ret = 0;
3490         struct address_space *mapping = inode->i_mapping;
3491         struct page *page;
3492         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3493                 PAGE_CACHE_SHIFT;
3494
3495         struct extent_page_data epd = {
3496                 .bio = NULL,
3497                 .tree = tree,
3498                 .get_extent = get_extent,
3499                 .extent_locked = 1,
3500                 .sync_io = mode == WB_SYNC_ALL,
3501         };
3502         struct writeback_control wbc_writepages = {
3503                 .sync_mode      = mode,
3504                 .nr_to_write    = nr_pages * 2,
3505                 .range_start    = start,
3506                 .range_end      = end + 1,
3507         };
3508
3509         while (start <= end) {
3510                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3511                 if (clear_page_dirty_for_io(page))
3512                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3513                 else {
3514                         if (tree->ops && tree->ops->writepage_end_io_hook)
3515                                 tree->ops->writepage_end_io_hook(page, start,
3516                                                  start + PAGE_CACHE_SIZE - 1,
3517                                                  NULL, 1);
3518                         unlock_page(page);
3519                 }
3520                 page_cache_release(page);
3521                 start += PAGE_CACHE_SIZE;
3522         }
3523
3524         flush_epd_write_bio(&epd);
3525         return ret;
3526 }
3527
3528 int extent_writepages(struct extent_io_tree *tree,
3529                       struct address_space *mapping,
3530                       get_extent_t *get_extent,
3531                       struct writeback_control *wbc)
3532 {
3533         int ret = 0;
3534         struct extent_page_data epd = {
3535                 .bio = NULL,
3536                 .tree = tree,
3537                 .get_extent = get_extent,
3538                 .extent_locked = 0,
3539                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3540         };
3541
3542         ret = extent_write_cache_pages(tree, mapping, wbc,
3543                                        __extent_writepage, &epd,
3544                                        flush_write_bio);
3545         flush_epd_write_bio(&epd);
3546         return ret;
3547 }
3548
3549 int extent_readpages(struct extent_io_tree *tree,
3550                      struct address_space *mapping,
3551                      struct list_head *pages, unsigned nr_pages,
3552                      get_extent_t get_extent)
3553 {
3554         struct bio *bio = NULL;
3555         unsigned page_idx;
3556         unsigned long bio_flags = 0;
3557         struct page *pagepool[16];
3558         struct page *page;
3559         int i = 0;
3560         int nr = 0;
3561
3562         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3563                 page = list_entry(pages->prev, struct page, lru);
3564
3565                 prefetchw(&page->flags);
3566                 list_del(&page->lru);
3567                 if (add_to_page_cache_lru(page, mapping,
3568                                         page->index, GFP_NOFS)) {
3569                         page_cache_release(page);
3570                         continue;
3571                 }
3572
3573                 pagepool[nr++] = page;
3574                 if (nr < ARRAY_SIZE(pagepool))
3575                         continue;
3576                 for (i = 0; i < nr; i++) {
3577                         __extent_read_full_page(tree, pagepool[i], get_extent,
3578                                         &bio, 0, &bio_flags);
3579                         page_cache_release(pagepool[i]);
3580                 }
3581                 nr = 0;
3582         }
3583         for (i = 0; i < nr; i++) {
3584                 __extent_read_full_page(tree, pagepool[i], get_extent,
3585                                         &bio, 0, &bio_flags);
3586                 page_cache_release(pagepool[i]);
3587         }
3588
3589         BUG_ON(!list_empty(pages));
3590         if (bio)
3591                 return submit_one_bio(READ, bio, 0, bio_flags);
3592         return 0;
3593 }
3594
3595 /*
3596  * basic invalidatepage code, this waits on any locked or writeback
3597  * ranges corresponding to the page, and then deletes any extent state
3598  * records from the tree
3599  */
3600 int extent_invalidatepage(struct extent_io_tree *tree,
3601                           struct page *page, unsigned long offset)
3602 {
3603         struct extent_state *cached_state = NULL;
3604         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
3605         u64 end = start + PAGE_CACHE_SIZE - 1;
3606         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3607
3608         start += (offset + blocksize - 1) & ~(blocksize - 1);
3609         if (start > end)
3610                 return 0;
3611
3612         lock_extent_bits(tree, start, end, 0, &cached_state);
3613         wait_on_page_writeback(page);
3614         clear_extent_bit(tree, start, end,
3615                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3616                          EXTENT_DO_ACCOUNTING,
3617                          1, 1, &cached_state, GFP_NOFS);
3618         return 0;
3619 }
3620
3621 /*
3622  * a helper for releasepage, this tests for areas of the page that
3623  * are locked or under IO and drops the related state bits if it is safe
3624  * to drop the page.
3625  */
3626 int try_release_extent_state(struct extent_map_tree *map,
3627                              struct extent_io_tree *tree, struct page *page,
3628                              gfp_t mask)
3629 {
3630         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3631         u64 end = start + PAGE_CACHE_SIZE - 1;
3632         int ret = 1;
3633
3634         if (test_range_bit(tree, start, end,
3635                            EXTENT_IOBITS, 0, NULL))
3636                 ret = 0;
3637         else {
3638                 if ((mask & GFP_NOFS) == GFP_NOFS)
3639                         mask = GFP_NOFS;
3640                 /*
3641                  * at this point we can safely clear everything except the
3642                  * locked bit and the nodatasum bit
3643                  */
3644                 ret = clear_extent_bit(tree, start, end,
3645                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3646                                  0, 0, NULL, mask);
3647
3648                 /* if clear_extent_bit failed for enomem reasons,
3649                  * we can't allow the release to continue.
3650                  */
3651                 if (ret < 0)
3652                         ret = 0;
3653                 else
3654                         ret = 1;
3655         }
3656         return ret;
3657 }
3658
3659 /*
3660  * a helper for releasepage.  As long as there are no locked extents
3661  * in the range corresponding to the page, both state records and extent
3662  * map records are removed
3663  */
3664 int try_release_extent_mapping(struct extent_map_tree *map,
3665                                struct extent_io_tree *tree, struct page *page,
3666                                gfp_t mask)
3667 {
3668         struct extent_map *em;
3669         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3670         u64 end = start + PAGE_CACHE_SIZE - 1;
3671
3672         if ((mask & __GFP_WAIT) &&
3673             page->mapping->host->i_size > 16 * 1024 * 1024) {
3674                 u64 len;
3675                 while (start <= end) {
3676                         len = end - start + 1;
3677                         write_lock(&map->lock);
3678                         em = lookup_extent_mapping(map, start, len);
3679                         if (!em) {
3680                                 write_unlock(&map->lock);
3681                                 break;
3682                         }
3683                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3684                             em->start != start) {
3685                                 write_unlock(&map->lock);
3686                                 free_extent_map(em);
3687                                 break;
3688                         }
3689                         if (!test_range_bit(tree, em->start,
3690                                             extent_map_end(em) - 1,
3691                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
3692                                             0, NULL)) {
3693                                 remove_extent_mapping(map, em);
3694                                 /* once for the rb tree */
3695                                 free_extent_map(em);
3696                         }
3697                         start = extent_map_end(em);
3698                         write_unlock(&map->lock);
3699
3700                         /* once for us */
3701                         free_extent_map(em);
3702                 }
3703         }
3704         return try_release_extent_state(map, tree, page, mask);
3705 }
3706
3707 /*
3708  * helper function for fiemap, which doesn't want to see any holes.
3709  * This maps until we find something past 'last'
3710  */
3711 static struct extent_map *get_extent_skip_holes(struct inode *inode,
3712                                                 u64 offset,
3713                                                 u64 last,
3714                                                 get_extent_t *get_extent)
3715 {
3716         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
3717         struct extent_map *em;
3718         u64 len;
3719
3720         if (offset >= last)
3721                 return NULL;
3722
3723         while(1) {
3724                 len = last - offset;
3725                 if (len == 0)
3726                         break;
3727                 len = (len + sectorsize - 1) & ~(sectorsize - 1);
3728                 em = get_extent(inode, NULL, 0, offset, len, 0);
3729                 if (IS_ERR_OR_NULL(em))
3730                         return em;
3731
3732                 /* if this isn't a hole return it */
3733                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
3734                     em->block_start != EXTENT_MAP_HOLE) {
3735                         return em;
3736                 }
3737
3738                 /* this is a hole, advance to the next extent */
3739                 offset = extent_map_end(em);
3740                 free_extent_map(em);
3741                 if (offset >= last)
3742                         break;
3743         }
3744         return NULL;
3745 }
3746
3747 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
3748                 __u64 start, __u64 len, get_extent_t *get_extent)
3749 {
3750         int ret = 0;
3751         u64 off = start;
3752         u64 max = start + len;
3753         u32 flags = 0;
3754         u32 found_type;
3755         u64 last;
3756         u64 last_for_get_extent = 0;
3757         u64 disko = 0;
3758         u64 isize = i_size_read(inode);
3759         struct btrfs_key found_key;
3760         struct extent_map *em = NULL;
3761         struct extent_state *cached_state = NULL;
3762         struct btrfs_path *path;
3763         struct btrfs_file_extent_item *item;
3764         int end = 0;
3765         u64 em_start = 0;
3766         u64 em_len = 0;
3767         u64 em_end = 0;
3768         unsigned long emflags;
3769
3770         if (len == 0)
3771                 return -EINVAL;
3772
3773         path = btrfs_alloc_path();
3774         if (!path)
3775                 return -ENOMEM;
3776         path->leave_spinning = 1;
3777
3778         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
3779         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
3780
3781         /*
3782          * lookup the last file extent.  We're not using i_size here
3783          * because there might be preallocation past i_size
3784          */
3785         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3786                                        path, btrfs_ino(inode), -1, 0);
3787         if (ret < 0) {
3788                 btrfs_free_path(path);
3789                 return ret;
3790         }
3791         WARN_ON(!ret);
3792         path->slots[0]--;
3793         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3794                               struct btrfs_file_extent_item);
3795         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3796         found_type = btrfs_key_type(&found_key);
3797
3798         /* No extents, but there might be delalloc bits */
3799         if (found_key.objectid != btrfs_ino(inode) ||
3800             found_type != BTRFS_EXTENT_DATA_KEY) {
3801                 /* have to trust i_size as the end */
3802                 last = (u64)-1;
3803                 last_for_get_extent = isize;
3804         } else {
3805                 /*
3806                  * remember the start of the last extent.  There are a
3807                  * bunch of different factors that go into the length of the
3808                  * extent, so its much less complex to remember where it started
3809                  */
3810                 last = found_key.offset;
3811                 last_for_get_extent = last + 1;
3812         }
3813         btrfs_free_path(path);
3814
3815         /*
3816          * we might have some extents allocated but more delalloc past those
3817          * extents.  so, we trust isize unless the start of the last extent is
3818          * beyond isize
3819          */
3820         if (last < isize) {
3821                 last = (u64)-1;
3822                 last_for_get_extent = isize;
3823         }
3824
3825         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3826                          &cached_state);
3827
3828         em = get_extent_skip_holes(inode, start, last_for_get_extent,
3829                                    get_extent);
3830         if (!em)
3831                 goto out;
3832         if (IS_ERR(em)) {
3833                 ret = PTR_ERR(em);
3834                 goto out;
3835         }
3836
3837         while (!end) {
3838                 u64 offset_in_extent;
3839
3840                 /* break if the extent we found is outside the range */
3841                 if (em->start >= max || extent_map_end(em) < off)
3842                         break;
3843
3844                 /*
3845                  * get_extent may return an extent that starts before our
3846                  * requested range.  We have to make sure the ranges
3847                  * we return to fiemap always move forward and don't
3848                  * overlap, so adjust the offsets here
3849                  */
3850                 em_start = max(em->start, off);
3851
3852                 /*
3853                  * record the offset from the start of the extent
3854                  * for adjusting the disk offset below
3855                  */
3856                 offset_in_extent = em_start - em->start;
3857                 em_end = extent_map_end(em);
3858                 em_len = em_end - em_start;
3859                 emflags = em->flags;
3860                 disko = 0;
3861                 flags = 0;
3862
3863                 /*
3864                  * bump off for our next call to get_extent
3865                  */
3866                 off = extent_map_end(em);
3867                 if (off >= max)
3868                         end = 1;
3869
3870                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3871                         end = 1;
3872                         flags |= FIEMAP_EXTENT_LAST;
3873                 } else if (em->block_start == EXTENT_MAP_INLINE) {
3874                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
3875                                   FIEMAP_EXTENT_NOT_ALIGNED);
3876                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3877                         flags |= (FIEMAP_EXTENT_DELALLOC |
3878                                   FIEMAP_EXTENT_UNKNOWN);
3879                 } else {
3880                         disko = em->block_start + offset_in_extent;
3881                 }
3882                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3883                         flags |= FIEMAP_EXTENT_ENCODED;
3884
3885                 free_extent_map(em);
3886                 em = NULL;
3887                 if ((em_start >= last) || em_len == (u64)-1 ||
3888                    (last == (u64)-1 && isize <= em_end)) {
3889                         flags |= FIEMAP_EXTENT_LAST;
3890                         end = 1;
3891                 }
3892
3893                 /* now scan forward to see if this is really the last extent. */
3894                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
3895                                            get_extent);
3896                 if (IS_ERR(em)) {
3897                         ret = PTR_ERR(em);
3898                         goto out;
3899                 }
3900                 if (!em) {
3901                         flags |= FIEMAP_EXTENT_LAST;
3902                         end = 1;
3903                 }
3904                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3905                                               em_len, flags);
3906                 if (ret)
3907                         goto out_free;
3908         }
3909 out_free:
3910         free_extent_map(em);
3911 out:
3912         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3913                              &cached_state, GFP_NOFS);
3914         return ret;
3915 }
3916
3917 inline struct page *extent_buffer_page(struct extent_buffer *eb,
3918                                               unsigned long i)
3919 {
3920         return eb->pages[i];
3921 }
3922
3923 inline unsigned long num_extent_pages(u64 start, u64 len)
3924 {
3925         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3926                 (start >> PAGE_CACHE_SHIFT);
3927 }
3928
3929 static void __free_extent_buffer(struct extent_buffer *eb)
3930 {
3931 #if LEAK_DEBUG
3932         unsigned long flags;
3933         spin_lock_irqsave(&leak_lock, flags);
3934         list_del(&eb->leak_list);
3935         spin_unlock_irqrestore(&leak_lock, flags);
3936 #endif
3937         if (eb->pages && eb->pages != eb->inline_pages)
3938                 kfree(eb->pages);
3939         kmem_cache_free(extent_buffer_cache, eb);
3940 }
3941
3942 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3943                                                    u64 start,
3944                                                    unsigned long len,
3945                                                    gfp_t mask)
3946 {
3947         struct extent_buffer *eb = NULL;
3948 #if LEAK_DEBUG
3949         unsigned long flags;
3950 #endif
3951
3952         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3953         if (eb == NULL)
3954                 return NULL;
3955         eb->start = start;
3956         eb->len = len;
3957         eb->tree = tree;
3958         eb->bflags = 0;
3959         rwlock_init(&eb->lock);
3960         atomic_set(&eb->write_locks, 0);
3961         atomic_set(&eb->read_locks, 0);
3962         atomic_set(&eb->blocking_readers, 0);
3963         atomic_set(&eb->blocking_writers, 0);
3964         atomic_set(&eb->spinning_readers, 0);
3965         atomic_set(&eb->spinning_writers, 0);
3966         eb->lock_nested = 0;
3967         init_waitqueue_head(&eb->write_lock_wq);
3968         init_waitqueue_head(&eb->read_lock_wq);
3969
3970 #if LEAK_DEBUG
3971         spin_lock_irqsave(&leak_lock, flags);
3972         list_add(&eb->leak_list, &buffers);
3973         spin_unlock_irqrestore(&leak_lock, flags);
3974 #endif
3975         spin_lock_init(&eb->refs_lock);
3976         atomic_set(&eb->refs, 1);
3977         atomic_set(&eb->io_pages, 0);
3978
3979         if (len > MAX_INLINE_EXTENT_BUFFER_SIZE) {
3980                 struct page **pages;
3981                 int num_pages = (len + PAGE_CACHE_SIZE - 1) >>
3982                         PAGE_CACHE_SHIFT;
3983                 pages = kzalloc(num_pages, mask);
3984                 if (!pages) {
3985                         __free_extent_buffer(eb);
3986                         return NULL;
3987                 }
3988                 eb->pages = pages;
3989         } else {
3990                 eb->pages = eb->inline_pages;
3991         }
3992
3993         return eb;
3994 }
3995
3996 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
3997 {
3998         unsigned long i;
3999         struct page *p;
4000         struct extent_buffer *new;
4001         unsigned long num_pages = num_extent_pages(src->start, src->len);
4002
4003         new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_ATOMIC);
4004         if (new == NULL)
4005                 return NULL;
4006
4007         for (i = 0; i < num_pages; i++) {
4008                 p = alloc_page(GFP_ATOMIC);
4009                 BUG_ON(!p);
4010                 attach_extent_buffer_page(new, p);
4011                 WARN_ON(PageDirty(p));
4012                 SetPageUptodate(p);
4013                 new->pages[i] = p;
4014         }
4015
4016         copy_extent_buffer(new, src, 0, 0, src->len);
4017         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4018         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4019
4020         return new;
4021 }
4022
4023 struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4024 {
4025         struct extent_buffer *eb;
4026         unsigned long num_pages = num_extent_pages(0, len);
4027         unsigned long i;
4028
4029         eb = __alloc_extent_buffer(NULL, start, len, GFP_ATOMIC);
4030         if (!eb)
4031                 return NULL;
4032
4033         for (i = 0; i < num_pages; i++) {
4034                 eb->pages[i] = alloc_page(GFP_ATOMIC);
4035                 if (!eb->pages[i])
4036                         goto err;
4037         }
4038         set_extent_buffer_uptodate(eb);
4039         btrfs_set_header_nritems(eb, 0);
4040         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4041
4042         return eb;
4043 err:
4044         for (i--; i > 0; i--)
4045                 __free_page(eb->pages[i]);
4046         __free_extent_buffer(eb);
4047         return NULL;
4048 }
4049
4050 static int extent_buffer_under_io(struct extent_buffer *eb)
4051 {
4052         return (atomic_read(&eb->io_pages) ||
4053                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4054                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4055 }
4056
4057 /*
4058  * Helper for releasing extent buffer page.
4059  */
4060 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4061                                                 unsigned long start_idx)
4062 {
4063         unsigned long index;
4064         unsigned long num_pages;
4065         struct page *page;
4066         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4067
4068         BUG_ON(extent_buffer_under_io(eb));
4069
4070         num_pages = num_extent_pages(eb->start, eb->len);
4071         index = start_idx + num_pages;
4072         if (start_idx >= index)
4073                 return;
4074
4075         do {
4076                 index--;
4077                 page = extent_buffer_page(eb, index);
4078                 if (page && mapped) {
4079                         spin_lock(&page->mapping->private_lock);
4080                         /*
4081                          * We do this since we'll remove the pages after we've
4082                          * removed the eb from the radix tree, so we could race
4083                          * and have this page now attached to the new eb.  So
4084                          * only clear page_private if it's still connected to
4085                          * this eb.
4086                          */
4087                         if (PagePrivate(page) &&
4088                             page->private == (unsigned long)eb) {
4089                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4090                                 BUG_ON(PageDirty(page));
4091                                 BUG_ON(PageWriteback(page));
4092                                 /*
4093                                  * We need to make sure we haven't be attached
4094                                  * to a new eb.
4095                                  */
4096                                 ClearPagePrivate(page);
4097                                 set_page_private(page, 0);
4098                                 /* One for the page private */
4099                                 page_cache_release(page);
4100                         }
4101                         spin_unlock(&page->mapping->private_lock);
4102
4103                 }
4104                 if (page) {
4105                         /* One for when we alloced the page */
4106                         page_cache_release(page);
4107                 }
4108         } while (index != start_idx);
4109 }
4110
4111 /*
4112  * Helper for releasing the extent buffer.
4113  */
4114 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4115 {
4116         btrfs_release_extent_buffer_page(eb, 0);
4117         __free_extent_buffer(eb);
4118 }
4119
4120 static void check_buffer_tree_ref(struct extent_buffer *eb)
4121 {
4122         /* the ref bit is tricky.  We have to make sure it is set
4123          * if we have the buffer dirty.   Otherwise the
4124          * code to free a buffer can end up dropping a dirty
4125          * page
4126          *
4127          * Once the ref bit is set, it won't go away while the
4128          * buffer is dirty or in writeback, and it also won't
4129          * go away while we have the reference count on the
4130          * eb bumped.
4131          *
4132          * We can't just set the ref bit without bumping the
4133          * ref on the eb because free_extent_buffer might
4134          * see the ref bit and try to clear it.  If this happens
4135          * free_extent_buffer might end up dropping our original
4136          * ref by mistake and freeing the page before we are able
4137          * to add one more ref.
4138          *
4139          * So bump the ref count first, then set the bit.  If someone
4140          * beat us to it, drop the ref we added.
4141          */
4142         spin_lock(&eb->refs_lock);
4143         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4144                 atomic_inc(&eb->refs);
4145         spin_unlock(&eb->refs_lock);
4146 }
4147
4148 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4149 {
4150         unsigned long num_pages, i;
4151
4152         check_buffer_tree_ref(eb);
4153
4154         num_pages = num_extent_pages(eb->start, eb->len);
4155         for (i = 0; i < num_pages; i++) {
4156                 struct page *p = extent_buffer_page(eb, i);
4157                 mark_page_accessed(p);
4158         }
4159 }
4160
4161 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
4162                                           u64 start, unsigned long len)
4163 {
4164         unsigned long num_pages = num_extent_pages(start, len);
4165         unsigned long i;
4166         unsigned long index = start >> PAGE_CACHE_SHIFT;
4167         struct extent_buffer *eb;
4168         struct extent_buffer *exists = NULL;
4169         struct page *p;
4170         struct address_space *mapping = tree->mapping;
4171         int uptodate = 1;
4172         int ret;
4173
4174         rcu_read_lock();
4175         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4176         if (eb && atomic_inc_not_zero(&eb->refs)) {
4177                 rcu_read_unlock();
4178                 mark_extent_buffer_accessed(eb);
4179                 return eb;
4180         }
4181         rcu_read_unlock();
4182
4183         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
4184         if (!eb)
4185                 return NULL;
4186
4187         for (i = 0; i < num_pages; i++, index++) {
4188                 p = find_or_create_page(mapping, index, GFP_NOFS);
4189                 if (!p) {
4190                         WARN_ON(1);
4191                         goto free_eb;
4192                 }
4193
4194                 spin_lock(&mapping->private_lock);
4195                 if (PagePrivate(p)) {
4196                         /*
4197                          * We could have already allocated an eb for this page
4198                          * and attached one so lets see if we can get a ref on
4199                          * the existing eb, and if we can we know it's good and
4200                          * we can just return that one, else we know we can just
4201                          * overwrite page->private.
4202                          */
4203                         exists = (struct extent_buffer *)p->private;
4204                         if (atomic_inc_not_zero(&exists->refs)) {
4205                                 spin_unlock(&mapping->private_lock);
4206                                 unlock_page(p);
4207                                 page_cache_release(p);
4208                                 mark_extent_buffer_accessed(exists);
4209                                 goto free_eb;
4210                         }
4211
4212                         /*
4213                          * Do this so attach doesn't complain and we need to
4214                          * drop the ref the old guy had.
4215                          */
4216                         ClearPagePrivate(p);
4217                         WARN_ON(PageDirty(p));
4218                         page_cache_release(p);
4219                 }
4220                 attach_extent_buffer_page(eb, p);
4221                 spin_unlock(&mapping->private_lock);
4222                 WARN_ON(PageDirty(p));
4223                 mark_page_accessed(p);
4224                 eb->pages[i] = p;
4225                 if (!PageUptodate(p))
4226                         uptodate = 0;
4227
4228                 /*
4229                  * see below about how we avoid a nasty race with release page
4230                  * and why we unlock later
4231                  */
4232         }
4233         if (uptodate)
4234                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4235 again:
4236         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4237         if (ret)
4238                 goto free_eb;
4239
4240         spin_lock(&tree->buffer_lock);
4241         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
4242         if (ret == -EEXIST) {
4243                 exists = radix_tree_lookup(&tree->buffer,
4244                                                 start >> PAGE_CACHE_SHIFT);
4245                 if (!atomic_inc_not_zero(&exists->refs)) {
4246                         spin_unlock(&tree->buffer_lock);
4247                         radix_tree_preload_end();
4248                         exists = NULL;
4249                         goto again;
4250                 }
4251                 spin_unlock(&tree->buffer_lock);
4252                 radix_tree_preload_end();
4253                 mark_extent_buffer_accessed(exists);
4254                 goto free_eb;
4255         }
4256         /* add one reference for the tree */
4257         check_buffer_tree_ref(eb);
4258         spin_unlock(&tree->buffer_lock);
4259         radix_tree_preload_end();
4260
4261         /*
4262          * there is a race where release page may have
4263          * tried to find this extent buffer in the radix
4264          * but failed.  It will tell the VM it is safe to
4265          * reclaim the, and it will clear the page private bit.
4266          * We must make sure to set the page private bit properly
4267          * after the extent buffer is in the radix tree so
4268          * it doesn't get lost
4269          */
4270         SetPageChecked(eb->pages[0]);
4271         for (i = 1; i < num_pages; i++) {
4272                 p = extent_buffer_page(eb, i);
4273                 ClearPageChecked(p);
4274                 unlock_page(p);
4275         }
4276         unlock_page(eb->pages[0]);
4277         return eb;
4278
4279 free_eb:
4280         for (i = 0; i < num_pages; i++) {
4281                 if (eb->pages[i])
4282                         unlock_page(eb->pages[i]);
4283         }
4284
4285         WARN_ON(!atomic_dec_and_test(&eb->refs));
4286         btrfs_release_extent_buffer(eb);
4287         return exists;
4288 }
4289
4290 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
4291                                          u64 start, unsigned long len)
4292 {
4293         struct extent_buffer *eb;
4294
4295         rcu_read_lock();
4296         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4297         if (eb && atomic_inc_not_zero(&eb->refs)) {
4298                 rcu_read_unlock();
4299                 mark_extent_buffer_accessed(eb);
4300                 return eb;
4301         }
4302         rcu_read_unlock();
4303
4304         return NULL;
4305 }
4306
4307 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4308 {
4309         struct extent_buffer *eb =
4310                         container_of(head, struct extent_buffer, rcu_head);
4311
4312         __free_extent_buffer(eb);
4313 }
4314
4315 /* Expects to have eb->eb_lock already held */
4316 static int release_extent_buffer(struct extent_buffer *eb, gfp_t mask)
4317 {
4318         WARN_ON(atomic_read(&eb->refs) == 0);
4319         if (atomic_dec_and_test(&eb->refs)) {
4320                 if (test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags)) {
4321                         spin_unlock(&eb->refs_lock);
4322                 } else {
4323                         struct extent_io_tree *tree = eb->tree;
4324
4325                         spin_unlock(&eb->refs_lock);
4326
4327                         spin_lock(&tree->buffer_lock);
4328                         radix_tree_delete(&tree->buffer,
4329                                           eb->start >> PAGE_CACHE_SHIFT);
4330                         spin_unlock(&tree->buffer_lock);
4331                 }
4332
4333                 /* Should be safe to release our pages at this point */
4334                 btrfs_release_extent_buffer_page(eb, 0);
4335
4336                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4337                 return 1;
4338         }
4339         spin_unlock(&eb->refs_lock);
4340
4341         return 0;
4342 }
4343
4344 void free_extent_buffer(struct extent_buffer *eb)
4345 {
4346         if (!eb)
4347                 return;
4348
4349         spin_lock(&eb->refs_lock);
4350         if (atomic_read(&eb->refs) == 2 &&
4351             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4352                 atomic_dec(&eb->refs);
4353
4354         if (atomic_read(&eb->refs) == 2 &&
4355             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4356             !extent_buffer_under_io(eb) &&
4357             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4358                 atomic_dec(&eb->refs);
4359
4360         /*
4361          * I know this is terrible, but it's temporary until we stop tracking
4362          * the uptodate bits and such for the extent buffers.
4363          */
4364         release_extent_buffer(eb, GFP_ATOMIC);
4365 }
4366
4367 void free_extent_buffer_stale(struct extent_buffer *eb)
4368 {
4369         if (!eb)
4370                 return;
4371
4372         spin_lock(&eb->refs_lock);
4373         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4374
4375         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4376             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4377                 atomic_dec(&eb->refs);
4378         release_extent_buffer(eb, GFP_NOFS);
4379 }
4380
4381 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4382 {
4383         unsigned long i;
4384         unsigned long num_pages;
4385         struct page *page;
4386
4387         num_pages = num_extent_pages(eb->start, eb->len);
4388
4389         for (i = 0; i < num_pages; i++) {
4390                 page = extent_buffer_page(eb, i);
4391                 if (!PageDirty(page))
4392                         continue;
4393
4394                 lock_page(page);
4395                 WARN_ON(!PagePrivate(page));
4396
4397                 clear_page_dirty_for_io(page);
4398                 spin_lock_irq(&page->mapping->tree_lock);
4399                 if (!PageDirty(page)) {
4400                         radix_tree_tag_clear(&page->mapping->page_tree,
4401                                                 page_index(page),
4402                                                 PAGECACHE_TAG_DIRTY);
4403                 }
4404                 spin_unlock_irq(&page->mapping->tree_lock);
4405                 ClearPageError(page);
4406                 unlock_page(page);
4407         }
4408         WARN_ON(atomic_read(&eb->refs) == 0);
4409 }
4410
4411 int set_extent_buffer_dirty(struct extent_buffer *eb)
4412 {
4413         unsigned long i;
4414         unsigned long num_pages;
4415         int was_dirty = 0;
4416
4417         check_buffer_tree_ref(eb);
4418
4419         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4420
4421         num_pages = num_extent_pages(eb->start, eb->len);
4422         WARN_ON(atomic_read(&eb->refs) == 0);
4423         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4424
4425         for (i = 0; i < num_pages; i++)
4426                 set_page_dirty(extent_buffer_page(eb, i));
4427         return was_dirty;
4428 }
4429
4430 static int range_straddles_pages(u64 start, u64 len)
4431 {
4432         if (len < PAGE_CACHE_SIZE)
4433                 return 1;
4434         if (start & (PAGE_CACHE_SIZE - 1))
4435                 return 1;
4436         if ((start + len) & (PAGE_CACHE_SIZE - 1))
4437                 return 1;
4438         return 0;
4439 }
4440
4441 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4442 {
4443         unsigned long i;
4444         struct page *page;
4445         unsigned long num_pages;
4446
4447         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4448         num_pages = num_extent_pages(eb->start, eb->len);
4449         for (i = 0; i < num_pages; i++) {
4450                 page = extent_buffer_page(eb, i);
4451                 if (page)
4452                         ClearPageUptodate(page);
4453         }
4454         return 0;
4455 }
4456
4457 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4458 {
4459         unsigned long i;
4460         struct page *page;
4461         unsigned long num_pages;
4462
4463         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4464         num_pages = num_extent_pages(eb->start, eb->len);
4465         for (i = 0; i < num_pages; i++) {
4466                 page = extent_buffer_page(eb, i);
4467                 SetPageUptodate(page);
4468         }
4469         return 0;
4470 }
4471
4472 int extent_range_uptodate(struct extent_io_tree *tree,
4473                           u64 start, u64 end)
4474 {
4475         struct page *page;
4476         int ret;
4477         int pg_uptodate = 1;
4478         int uptodate;
4479         unsigned long index;
4480
4481         if (range_straddles_pages(start, end - start + 1)) {
4482                 ret = test_range_bit(tree, start, end,
4483                                      EXTENT_UPTODATE, 1, NULL);
4484                 if (ret)
4485                         return 1;
4486         }
4487         while (start <= end) {
4488                 index = start >> PAGE_CACHE_SHIFT;
4489                 page = find_get_page(tree->mapping, index);
4490                 if (!page)
4491                         return 1;
4492                 uptodate = PageUptodate(page);
4493                 page_cache_release(page);
4494                 if (!uptodate) {
4495                         pg_uptodate = 0;
4496                         break;
4497                 }
4498                 start += PAGE_CACHE_SIZE;
4499         }
4500         return pg_uptodate;
4501 }
4502
4503 int extent_buffer_uptodate(struct extent_buffer *eb)
4504 {
4505         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4506 }
4507
4508 int read_extent_buffer_pages(struct extent_io_tree *tree,
4509                              struct extent_buffer *eb, u64 start, int wait,
4510                              get_extent_t *get_extent, int mirror_num)
4511 {
4512         unsigned long i;
4513         unsigned long start_i;
4514         struct page *page;
4515         int err;
4516         int ret = 0;
4517         int locked_pages = 0;
4518         int all_uptodate = 1;
4519         unsigned long num_pages;
4520         unsigned long num_reads = 0;
4521         struct bio *bio = NULL;
4522         unsigned long bio_flags = 0;
4523
4524         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4525                 return 0;
4526
4527         if (start) {
4528                 WARN_ON(start < eb->start);
4529                 start_i = (start >> PAGE_CACHE_SHIFT) -
4530                         (eb->start >> PAGE_CACHE_SHIFT);
4531         } else {
4532                 start_i = 0;
4533         }
4534
4535         num_pages = num_extent_pages(eb->start, eb->len);
4536         for (i = start_i; i < num_pages; i++) {
4537                 page = extent_buffer_page(eb, i);
4538                 if (wait == WAIT_NONE) {
4539                         if (!trylock_page(page))
4540                                 goto unlock_exit;
4541                 } else {
4542                         lock_page(page);
4543                 }
4544                 locked_pages++;
4545                 if (!PageUptodate(page)) {
4546                         num_reads++;
4547                         all_uptodate = 0;
4548                 }
4549         }
4550         if (all_uptodate) {
4551                 if (start_i == 0)
4552                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4553                 goto unlock_exit;
4554         }
4555
4556         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4557         eb->read_mirror = 0;
4558         atomic_set(&eb->io_pages, num_reads);
4559         for (i = start_i; i < num_pages; i++) {
4560                 page = extent_buffer_page(eb, i);
4561                 if (!PageUptodate(page)) {
4562                         ClearPageError(page);
4563                         err = __extent_read_full_page(tree, page,
4564                                                       get_extent, &bio,
4565                                                       mirror_num, &bio_flags);
4566                         if (err)
4567                                 ret = err;
4568                 } else {
4569                         unlock_page(page);
4570                 }
4571         }
4572
4573         if (bio) {
4574                 err = submit_one_bio(READ, bio, mirror_num, bio_flags);
4575                 if (err)
4576                         return err;
4577         }
4578
4579         if (ret || wait != WAIT_COMPLETE)
4580                 return ret;
4581
4582         for (i = start_i; i < num_pages; i++) {
4583                 page = extent_buffer_page(eb, i);
4584                 wait_on_page_locked(page);
4585                 if (!PageUptodate(page))
4586                         ret = -EIO;
4587         }
4588
4589         return ret;
4590
4591 unlock_exit:
4592         i = start_i;
4593         while (locked_pages > 0) {
4594                 page = extent_buffer_page(eb, i);
4595                 i++;
4596                 unlock_page(page);
4597                 locked_pages--;
4598         }
4599         return ret;
4600 }
4601
4602 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4603                         unsigned long start,
4604                         unsigned long len)
4605 {
4606         size_t cur;
4607         size_t offset;
4608         struct page *page;
4609         char *kaddr;
4610         char *dst = (char *)dstv;
4611         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4612         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4613
4614         WARN_ON(start > eb->len);
4615         WARN_ON(start + len > eb->start + eb->len);
4616
4617         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4618
4619         while (len > 0) {
4620                 page = extent_buffer_page(eb, i);
4621
4622                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4623                 kaddr = page_address(page);
4624                 memcpy(dst, kaddr + offset, cur);
4625
4626                 dst += cur;
4627                 len -= cur;
4628                 offset = 0;
4629                 i++;
4630         }
4631 }
4632
4633 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4634                                unsigned long min_len, char **map,
4635                                unsigned long *map_start,
4636                                unsigned long *map_len)
4637 {
4638         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4639         char *kaddr;
4640         struct page *p;
4641         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4642         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4643         unsigned long end_i = (start_offset + start + min_len - 1) >>
4644                 PAGE_CACHE_SHIFT;
4645
4646         if (i != end_i)
4647                 return -EINVAL;
4648
4649         if (i == 0) {
4650                 offset = start_offset;
4651                 *map_start = 0;
4652         } else {
4653                 offset = 0;
4654                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4655         }
4656
4657         if (start + min_len > eb->len) {
4658                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4659                        "wanted %lu %lu\n", (unsigned long long)eb->start,
4660                        eb->len, start, min_len);
4661                 WARN_ON(1);
4662                 return -EINVAL;
4663         }
4664
4665         p = extent_buffer_page(eb, i);
4666         kaddr = page_address(p);
4667         *map = kaddr + offset;
4668         *map_len = PAGE_CACHE_SIZE - offset;
4669         return 0;
4670 }
4671
4672 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4673                           unsigned long start,
4674                           unsigned long len)
4675 {
4676         size_t cur;
4677         size_t offset;
4678         struct page *page;
4679         char *kaddr;
4680         char *ptr = (char *)ptrv;
4681         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4682         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4683         int ret = 0;
4684
4685         WARN_ON(start > eb->len);
4686         WARN_ON(start + len > eb->start + eb->len);
4687
4688         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4689
4690         while (len > 0) {
4691                 page = extent_buffer_page(eb, i);
4692
4693                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4694
4695                 kaddr = page_address(page);
4696                 ret = memcmp(ptr, kaddr + offset, cur);
4697                 if (ret)
4698                         break;
4699
4700                 ptr += cur;
4701                 len -= cur;
4702                 offset = 0;
4703                 i++;
4704         }
4705         return ret;
4706 }
4707
4708 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4709                          unsigned long start, unsigned long len)
4710 {
4711         size_t cur;
4712         size_t offset;
4713         struct page *page;
4714         char *kaddr;
4715         char *src = (char *)srcv;
4716         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4717         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4718
4719         WARN_ON(start > eb->len);
4720         WARN_ON(start + len > eb->start + eb->len);
4721
4722         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4723
4724         while (len > 0) {
4725                 page = extent_buffer_page(eb, i);
4726                 WARN_ON(!PageUptodate(page));
4727
4728                 cur = min(len, PAGE_CACHE_SIZE - offset);
4729                 kaddr = page_address(page);
4730                 memcpy(kaddr + offset, src, cur);
4731
4732                 src += cur;
4733                 len -= cur;
4734                 offset = 0;
4735                 i++;
4736         }
4737 }
4738
4739 void memset_extent_buffer(struct extent_buffer *eb, char c,
4740                           unsigned long start, unsigned long len)
4741 {
4742         size_t cur;
4743         size_t offset;
4744         struct page *page;
4745         char *kaddr;
4746         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4747         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4748
4749         WARN_ON(start > eb->len);
4750         WARN_ON(start + len > eb->start + eb->len);
4751
4752         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4753
4754         while (len > 0) {
4755                 page = extent_buffer_page(eb, i);
4756                 WARN_ON(!PageUptodate(page));
4757
4758                 cur = min(len, PAGE_CACHE_SIZE - offset);
4759                 kaddr = page_address(page);
4760                 memset(kaddr + offset, c, cur);
4761
4762                 len -= cur;
4763                 offset = 0;
4764                 i++;
4765         }
4766 }
4767
4768 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
4769                         unsigned long dst_offset, unsigned long src_offset,
4770                         unsigned long len)
4771 {
4772         u64 dst_len = dst->len;
4773         size_t cur;
4774         size_t offset;
4775         struct page *page;
4776         char *kaddr;
4777         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4778         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4779
4780         WARN_ON(src->len != dst_len);
4781
4782         offset = (start_offset + dst_offset) &
4783                 ((unsigned long)PAGE_CACHE_SIZE - 1);
4784
4785         while (len > 0) {
4786                 page = extent_buffer_page(dst, i);
4787                 WARN_ON(!PageUptodate(page));
4788
4789                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
4790
4791                 kaddr = page_address(page);
4792                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4793
4794                 src_offset += cur;
4795                 len -= cur;
4796                 offset = 0;
4797                 i++;
4798         }
4799 }
4800
4801 static void move_pages(struct page *dst_page, struct page *src_page,
4802                        unsigned long dst_off, unsigned long src_off,
4803                        unsigned long len)
4804 {
4805         char *dst_kaddr = page_address(dst_page);
4806         if (dst_page == src_page) {
4807                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
4808         } else {
4809                 char *src_kaddr = page_address(src_page);
4810                 char *p = dst_kaddr + dst_off + len;
4811                 char *s = src_kaddr + src_off + len;
4812
4813                 while (len--)
4814                         *--p = *--s;
4815         }
4816 }
4817
4818 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4819 {
4820         unsigned long distance = (src > dst) ? src - dst : dst - src;
4821         return distance < len;
4822 }
4823
4824 static void copy_pages(struct page *dst_page, struct page *src_page,
4825                        unsigned long dst_off, unsigned long src_off,
4826                        unsigned long len)
4827 {
4828         char *dst_kaddr = page_address(dst_page);
4829         char *src_kaddr;
4830         int must_memmove = 0;
4831
4832         if (dst_page != src_page) {
4833                 src_kaddr = page_address(src_page);
4834         } else {
4835                 src_kaddr = dst_kaddr;
4836                 if (areas_overlap(src_off, dst_off, len))
4837                         must_memmove = 1;
4838         }
4839
4840         if (must_memmove)
4841                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
4842         else
4843                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
4844 }
4845
4846 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4847                            unsigned long src_offset, unsigned long len)
4848 {
4849         size_t cur;
4850         size_t dst_off_in_page;
4851         size_t src_off_in_page;
4852         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4853         unsigned long dst_i;
4854         unsigned long src_i;
4855
4856         if (src_offset + len > dst->len) {
4857                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4858                        "len %lu dst len %lu\n", src_offset, len, dst->len);
4859                 BUG_ON(1);
4860         }
4861         if (dst_offset + len > dst->len) {
4862                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4863                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
4864                 BUG_ON(1);
4865         }
4866
4867         while (len > 0) {
4868                 dst_off_in_page = (start_offset + dst_offset) &
4869                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4870                 src_off_in_page = (start_offset + src_offset) &
4871                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4872
4873                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4874                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
4875
4876                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
4877                                                src_off_in_page));
4878                 cur = min_t(unsigned long, cur,
4879                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
4880
4881                 copy_pages(extent_buffer_page(dst, dst_i),
4882                            extent_buffer_page(dst, src_i),
4883                            dst_off_in_page, src_off_in_page, cur);
4884
4885                 src_offset += cur;
4886                 dst_offset += cur;
4887                 len -= cur;
4888         }
4889 }
4890
4891 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4892                            unsigned long src_offset, unsigned long len)
4893 {
4894         size_t cur;
4895         size_t dst_off_in_page;
4896         size_t src_off_in_page;
4897         unsigned long dst_end = dst_offset + len - 1;
4898         unsigned long src_end = src_offset + len - 1;
4899         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4900         unsigned long dst_i;
4901         unsigned long src_i;
4902
4903         if (src_offset + len > dst->len) {
4904                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4905                        "len %lu len %lu\n", src_offset, len, dst->len);
4906                 BUG_ON(1);
4907         }
4908         if (dst_offset + len > dst->len) {
4909                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4910                        "len %lu len %lu\n", dst_offset, len, dst->len);
4911                 BUG_ON(1);
4912         }
4913         if (dst_offset < src_offset) {
4914                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4915                 return;
4916         }
4917         while (len > 0) {
4918                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4919                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4920
4921                 dst_off_in_page = (start_offset + dst_end) &
4922                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4923                 src_off_in_page = (start_offset + src_end) &
4924                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4925
4926                 cur = min_t(unsigned long, len, src_off_in_page + 1);
4927                 cur = min(cur, dst_off_in_page + 1);
4928                 move_pages(extent_buffer_page(dst, dst_i),
4929                            extent_buffer_page(dst, src_i),
4930                            dst_off_in_page - cur + 1,
4931                            src_off_in_page - cur + 1, cur);
4932
4933                 dst_end -= cur;
4934                 src_end -= cur;
4935                 len -= cur;
4936         }
4937 }
4938
4939 int try_release_extent_buffer(struct page *page, gfp_t mask)
4940 {
4941         struct extent_buffer *eb;
4942
4943         /*
4944          * We need to make sure noboody is attaching this page to an eb right
4945          * now.
4946          */
4947         spin_lock(&page->mapping->private_lock);
4948         if (!PagePrivate(page)) {
4949                 spin_unlock(&page->mapping->private_lock);
4950                 return 1;
4951         }
4952
4953         eb = (struct extent_buffer *)page->private;
4954         BUG_ON(!eb);
4955
4956         /*
4957          * This is a little awful but should be ok, we need to make sure that
4958          * the eb doesn't disappear out from under us while we're looking at
4959          * this page.
4960          */
4961         spin_lock(&eb->refs_lock);
4962         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4963                 spin_unlock(&eb->refs_lock);
4964                 spin_unlock(&page->mapping->private_lock);
4965                 return 0;
4966         }
4967         spin_unlock(&page->mapping->private_lock);
4968
4969         if ((mask & GFP_NOFS) == GFP_NOFS)
4970                 mask = GFP_NOFS;
4971
4972         /*
4973          * If tree ref isn't set then we know the ref on this eb is a real ref,
4974          * so just return, this page will likely be freed soon anyway.
4975          */
4976         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4977                 spin_unlock(&eb->refs_lock);
4978                 return 0;
4979         }
4980
4981         return release_extent_buffer(eb, mask);
4982 }