fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "compat.h"
  17 #include "ctree.h"
  18 #include "btrfs_inode.h"
  19 #include "volumes.h"
  20 #include "check-integrity.h"
  21 #include "locking.h"
  22 #include "rcu-string.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26 static struct bio_set *btrfs_bioset;
  27
  28 #ifdef CONFIG_BTRFS_DEBUG
  29 static LIST_HEAD(buffers);
  30 static LIST_HEAD(states);
  31
  32 static DEFINE_SPINLOCK(leak_lock);
  33
  34 static inline
  35 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  36 {
  37         unsigned long flags;
  38
  39         spin_lock_irqsave(&leak_lock, flags);
  40         list_add(new, head);
  41         spin_unlock_irqrestore(&leak_lock, flags);
  42 }
  43
  44 static inline
  45 void btrfs_leak_debug_del(struct list_head *entry)
  46 {
  47         unsigned long flags;
  48
  49         spin_lock_irqsave(&leak_lock, flags);
  50         list_del(entry);
  51         spin_unlock_irqrestore(&leak_lock, flags);
  52 }
  53
  54 static inline
  55 void btrfs_leak_debug_check(void)
  56 {
  57         struct extent_state *state;
  58         struct extent_buffer *eb;
  59
  60         while (!list_empty(&states)) {
  61                 state = list_entry(states.next, struct extent_state, leak_list);
  62                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  63                        "state %lu in tree %p refs %d\n",
  64                        (unsigned long long)state->start,
  65                        (unsigned long long)state->end,
  66                        state->state, state->tree, atomic_read(&state->refs));
  67                 list_del(&state->leak_list);
  68                 kmem_cache_free(extent_state_cache, state);
  69         }
  70
  71         while (!list_empty(&buffers)) {
  72                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  73                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  74                        "refs %d\n", (unsigned long long)eb->start,
  75                        eb->len, atomic_read(&eb->refs));
  76                 list_del(&eb->leak_list);
  77                 kmem_cache_free(extent_buffer_cache, eb);
  78         }
  79 }
  80
  81 #define btrfs_debug_check_extent_io_range(inode, start, end)            \
  82         __btrfs_debug_check_extent_io_range(__func__, (inode), (start), (end))
  83 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  84                 struct inode *inode, u64 start, u64 end)
  85 {
  86         u64 isize = i_size_read(inode);
  87
  88         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  89                 printk_ratelimited(KERN_DEBUG
  90                     "btrfs: %s: ino %llu isize %llu odd range [%llu,%llu]\n",
  91                                 caller,
  92                                 (unsigned long long)btrfs_ino(inode),
  93                                 (unsigned long long)isize,
  94                                 (unsigned long long)start,
  95                                 (unsigned long long)end);
  96         }
  97 }
  98 #else
  99 #define btrfs_leak_debug_add(new, head) do {} while (0)
 100 #define btrfs_leak_debug_del(entry)     do {} while (0)
 101 #define btrfs_leak_debug_check()        do {} while (0)
 102 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 103 #endif
 104
 105 #define BUFFER_LRU_MAX 64
 106
 107 struct tree_entry {
 108         u64 start;
 109         u64 end;
 110         struct rb_node rb_node;
 111 };
 112
 113 struct extent_page_data {
 114         struct bio *bio;
 115         struct extent_io_tree *tree;
 116         get_extent_t *get_extent;
 117         unsigned long bio_flags;
 118
 119         /* tells writepage not to lock the state bits for this range
 120          * it still does the unlocking
 121          */
 122         unsigned int extent_locked:1;
 123
 124         /* tells the submit_bio code to use a WRITE_SYNC */
 125         unsigned int sync_io:1;
 126 };
 127
 128 static noinline void flush_write_bio(void *data);
 129 static inline struct btrfs_fs_info *
 130 tree_fs_info(struct extent_io_tree *tree)
 131 {
 132         return btrfs_sb(tree->mapping->host->i_sb);
 133 }
 134
 135 int __init extent_io_init(void)
 136 {
 137         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 138                         sizeof(struct extent_state), 0,
 139                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 140         if (!extent_state_cache)
 141                 return -ENOMEM;
 142
 143         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 144                         sizeof(struct extent_buffer), 0,
 145                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 146         if (!extent_buffer_cache)
 147                 goto free_state_cache;
 148
 149         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 150                                      offsetof(struct btrfs_io_bio, bio));
 151         if (!btrfs_bioset)
 152                 goto free_buffer_cache;
 153         return 0;
 154
 155 free_buffer_cache:
 156         kmem_cache_destroy(extent_buffer_cache);
 157         extent_buffer_cache = NULL;
 158
 159 free_state_cache:
 160         kmem_cache_destroy(extent_state_cache);
 161         extent_state_cache = NULL;
 162         return -ENOMEM;
 163 }
 164
 165 void extent_io_exit(void)
 166 {
 167         btrfs_leak_debug_check();
 168
 169         /*
 170          * Make sure all delayed rcu free are flushed before we
 171          * destroy caches.
 172          */
 173         rcu_barrier();
 174         if (extent_state_cache)
 175                 kmem_cache_destroy(extent_state_cache);
 176         if (extent_buffer_cache)
 177                 kmem_cache_destroy(extent_buffer_cache);
 178         if (btrfs_bioset)
 179                 bioset_free(btrfs_bioset);
 180 }
 181
 182 void extent_io_tree_init(struct extent_io_tree *tree,
 183                          struct address_space *mapping)
 184 {
 185         tree->state = RB_ROOT;
 186         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 187         tree->ops = NULL;
 188         tree->dirty_bytes = 0;
 189         spin_lock_init(&tree->lock);
 190         spin_lock_init(&tree->buffer_lock);
 191         tree->mapping = mapping;
 192 }
 193
 194 static struct extent_state *alloc_extent_state(gfp_t mask)
 195 {
 196         struct extent_state *state;
 197
 198         state = kmem_cache_alloc(extent_state_cache, mask);
 199         if (!state)
 200                 return state;
 201         state->state = 0;
 202         state->private = 0;
 203         state->tree = NULL;
 204         btrfs_leak_debug_add(&state->leak_list, &states);
 205         atomic_set(&state->refs, 1);
 206         init_waitqueue_head(&state->wq);
 207         trace_alloc_extent_state(state, mask, _RET_IP_);
 208         return state;
 209 }
 210
 211 void free_extent_state(struct extent_state *state)
 212 {
 213         if (!state)
 214                 return;
 215         if (atomic_dec_and_test(&state->refs)) {
 216                 WARN_ON(state->tree);
 217                 btrfs_leak_debug_del(&state->leak_list);
 218                 trace_free_extent_state(state, _RET_IP_);
 219                 kmem_cache_free(extent_state_cache, state);
 220         }
 221 }
 222
 223 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 224                                    struct rb_node *node)
 225 {
 226         struct rb_node **p = &root->rb_node;
 227         struct rb_node *parent = NULL;
 228         struct tree_entry *entry;
 229
 230         while (*p) {
 231                 parent = *p;
 232                 entry = rb_entry(parent, struct tree_entry, rb_node);
 233
 234                 if (offset < entry->start)
 235                         p = &(*p)->rb_left;
 236                 else if (offset > entry->end)
 237                         p = &(*p)->rb_right;
 238                 else
 239                         return parent;
 240         }
 241
 242         rb_link_node(node, parent, p);
 243         rb_insert_color(node, root);
 244         return NULL;
 245 }
 246
 247 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 248                                      struct rb_node **prev_ret,
 249                                      struct rb_node **next_ret)
 250 {
 251         struct rb_root *root = &tree->state;
 252         struct rb_node *n = root->rb_node;
 253         struct rb_node *prev = NULL;
 254         struct rb_node *orig_prev = NULL;
 255         struct tree_entry *entry;
 256         struct tree_entry *prev_entry = NULL;
 257
 258         while (n) {
 259                 entry = rb_entry(n, struct tree_entry, rb_node);
 260                 prev = n;
 261                 prev_entry = entry;
 262
 263                 if (offset < entry->start)
 264                         n = n->rb_left;
 265                 else if (offset > entry->end)
 266                         n = n->rb_right;
 267                 else
 268                         return n;
 269         }
 270
 271         if (prev_ret) {
 272                 orig_prev = prev;
 273                 while (prev && offset > prev_entry->end) {
 274                         prev = rb_next(prev);
 275                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 276                 }
 277                 *prev_ret = prev;
 278                 prev = orig_prev;
 279         }
 280
 281         if (next_ret) {
 282                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 283                 while (prev && offset < prev_entry->start) {
 284                         prev = rb_prev(prev);
 285                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 286                 }
 287                 *next_ret = prev;
 288         }
 289         return NULL;
 290 }
 291
 292 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 293                                           u64 offset)
 294 {
 295         struct rb_node *prev = NULL;
 296         struct rb_node *ret;
 297
 298         ret = __etree_search(tree, offset, &prev, NULL);
 299         if (!ret)
 300                 return prev;
 301         return ret;
 302 }
 303
 304 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 305                      struct extent_state *other)
 306 {
 307         if (tree->ops && tree->ops->merge_extent_hook)
 308                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 309                                              other);
 310 }
 311
 312 /*
 313  * utility function to look for merge candidates inside a given range.
 314  * Any extents with matching state are merged together into a single
 315  * extent in the tree.  Extents with EXTENT_IO in their state field
 316  * are not merged because the end_io handlers need to be able to do
 317  * operations on them without sleeping (or doing allocations/splits).
 318  *
 319  * This should be called with the tree lock held.
 320  */
 321 static void merge_state(struct extent_io_tree *tree,
 322                         struct extent_state *state)
 323 {
 324         struct extent_state *other;
 325         struct rb_node *other_node;
 326
 327         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 328                 return;
 329
 330         other_node = rb_prev(&state->rb_node);
 331         if (other_node) {
 332                 other = rb_entry(other_node, struct extent_state, rb_node);
 333                 if (other->end == state->start - 1 &&
 334                     other->state == state->state) {
 335                         merge_cb(tree, state, other);
 336                         state->start = other->start;
 337                         other->tree = NULL;
 338                         rb_erase(&other->rb_node, &tree->state);
 339                         free_extent_state(other);
 340                 }
 341         }
 342         other_node = rb_next(&state->rb_node);
 343         if (other_node) {
 344                 other = rb_entry(other_node, struct extent_state, rb_node);
 345                 if (other->start == state->end + 1 &&
 346                     other->state == state->state) {
 347                         merge_cb(tree, state, other);
 348                         state->end = other->end;
 349                         other->tree = NULL;
 350                         rb_erase(&other->rb_node, &tree->state);
 351                         free_extent_state(other);
 352                 }
 353         }
 354 }
 355
 356 static void set_state_cb(struct extent_io_tree *tree,
 357                          struct extent_state *state, unsigned long *bits)
 358 {
 359         if (tree->ops && tree->ops->set_bit_hook)
 360                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 361 }
 362
 363 static void clear_state_cb(struct extent_io_tree *tree,
 364                            struct extent_state *state, unsigned long *bits)
 365 {
 366         if (tree->ops && tree->ops->clear_bit_hook)
 367                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 368 }
 369
 370 static void set_state_bits(struct extent_io_tree *tree,
 371                            struct extent_state *state, unsigned long *bits);
 372
 373 /*
 374  * insert an extent_state struct into the tree.  'bits' are set on the
 375  * struct before it is inserted.
 376  *
 377  * This may return -EEXIST if the extent is already there, in which case the
 378  * state struct is freed.
 379  *
 380  * The tree lock is not taken internally.  This is a utility function and
 381  * probably isn't what you want to call (see set/clear_extent_bit).
 382  */
 383 static int insert_state(struct extent_io_tree *tree,
 384                         struct extent_state *state, u64 start, u64 end,
 385                         unsigned long *bits)
 386 {
 387         struct rb_node *node;
 388
 389         if (end < start)
 390                 WARN(1, KERN_ERR "btrfs end < start %llu %llu\n",
 391                        (unsigned long long)end,
 392                        (unsigned long long)start);
 393         state->start = start;
 394         state->end = end;
 395
 396         set_state_bits(tree, state, bits);
 397
 398         node = tree_insert(&tree->state, end, &state->rb_node);
 399         if (node) {
 400                 struct extent_state *found;
 401                 found = rb_entry(node, struct extent_state, rb_node);
 402                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 403                        "%llu %llu\n", (unsigned long long)found->start,
 404                        (unsigned long long)found->end,
 405                        (unsigned long long)start, (unsigned long long)end);
 406                 return -EEXIST;
 407         }
 408         state->tree = tree;
 409         merge_state(tree, state);
 410         return 0;
 411 }
 412
 413 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 414                      u64 split)
 415 {
 416         if (tree->ops && tree->ops->split_extent_hook)
 417                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 418 }
 419
 420 /*
 421  * split a given extent state struct in two, inserting the preallocated
 422  * struct 'prealloc' as the newly created second half.  'split' indicates an
 423  * offset inside 'orig' where it should be split.
 424  *
 425  * Before calling,
 426  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 427  * are two extent state structs in the tree:
 428  * prealloc: [orig->start, split - 1]
 429  * orig: [ split, orig->end ]
 430  *
 431  * The tree locks are not taken by this function. They need to be held
 432  * by the caller.
 433  */
 434 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 435                        struct extent_state *prealloc, u64 split)
 436 {
 437         struct rb_node *node;
 438
 439         split_cb(tree, orig, split);
 440
 441         prealloc->start = orig->start;
 442         prealloc->end = split - 1;
 443         prealloc->state = orig->state;
 444         orig->start = split;
 445
 446         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 447         if (node) {
 448                 free_extent_state(prealloc);
 449                 return -EEXIST;
 450         }
 451         prealloc->tree = tree;
 452         return 0;
 453 }
 454
 455 static struct extent_state *next_state(struct extent_state *state)
 456 {
 457         struct rb_node *next = rb_next(&state->rb_node);
 458         if (next)
 459                 return rb_entry(next, struct extent_state, rb_node);
 460         else
 461                 return NULL;
 462 }
 463
 464 /*
 465  * utility function to clear some bits in an extent state struct.
 466  * it will optionally wake up any one waiting on this state (wake == 1).
 467  *
 468  * If no bits are set on the state struct after clearing things, the
 469  * struct is freed and removed from the tree
 470  */
 471 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 472                                             struct extent_state *state,
 473                                             unsigned long *bits, int wake)
 474 {
 475         struct extent_state *next;
 476         unsigned long bits_to_clear = *bits & ~EXTENT_CTLBITS;
 477
 478         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 479                 u64 range = state->end - state->start + 1;
 480                 WARN_ON(range > tree->dirty_bytes);
 481                 tree->dirty_bytes -= range;
 482         }
 483         clear_state_cb(tree, state, bits);
 484         state->state &= ~bits_to_clear;
 485         if (wake)
 486                 wake_up(&state->wq);
 487         if (state->state == 0) {
 488                 next = next_state(state);
 489                 if (state->tree) {
 490                         rb_erase(&state->rb_node, &tree->state);
 491                         state->tree = NULL;
 492                         free_extent_state(state);
 493                 } else {
 494                         WARN_ON(1);
 495                 }
 496         } else {
 497                 merge_state(tree, state);
 498                 next = next_state(state);
 499         }
 500         return next;
 501 }
 502
 503 static struct extent_state *
 504 alloc_extent_state_atomic(struct extent_state *prealloc)
 505 {
 506         if (!prealloc)
 507                 prealloc = alloc_extent_state(GFP_ATOMIC);
 508
 509         return prealloc;
 510 }
 511
 512 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 513 {
 514         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 515                     "Extent tree was modified by another "
 516                     "thread while locked.");
 517 }
 518
 519 /*
 520  * clear some bits on a range in the tree.  This may require splitting
 521  * or inserting elements in the tree, so the gfp mask is used to
 522  * indicate which allocations or sleeping are allowed.
 523  *
 524  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 525  * the given range from the tree regardless of state (ie for truncate).
 526  *
 527  * the range [start, end] is inclusive.
 528  *
 529  * This takes the tree lock, and returns 0 on success and < 0 on error.
 530  */
 531 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 532                      unsigned long bits, int wake, int delete,
 533                      struct extent_state **cached_state,
 534                      gfp_t mask)
 535 {
 536         struct extent_state *state;
 537         struct extent_state *cached;
 538         struct extent_state *prealloc = NULL;
 539         struct rb_node *node;
 540         u64 last_end;
 541         int err;
 542         int clear = 0;
 543
 544         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
 545
 546         if (bits & EXTENT_DELALLOC)
 547                 bits |= EXTENT_NORESERVE;
 548
 549         if (delete)
 550                 bits |= ~EXTENT_CTLBITS;
 551         bits |= EXTENT_FIRST_DELALLOC;
 552
 553         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 554                 clear = 1;
 555 again:
 556         if (!prealloc && (mask & __GFP_WAIT)) {
 557                 prealloc = alloc_extent_state(mask);
 558                 if (!prealloc)
 559                         return -ENOMEM;
 560         }
 561
 562         spin_lock(&tree->lock);
 563         if (cached_state) {
 564                 cached = *cached_state;
 565
 566                 if (clear) {
 567                         *cached_state = NULL;
 568                         cached_state = NULL;
 569                 }
 570
 571                 if (cached && cached->tree && cached->start <= start &&
 572                     cached->end > start) {
 573                         if (clear)
 574                                 atomic_dec(&cached->refs);
 575                         state = cached;
 576                         goto hit_next;
 577                 }
 578                 if (clear)
 579                         free_extent_state(cached);
 580         }
 581         /*
 582          * this search will find the extents that end after
 583          * our range starts
 584          */
 585         node = tree_search(tree, start);
 586         if (!node)
 587                 goto out;
 588         state = rb_entry(node, struct extent_state, rb_node);
 589 hit_next:
 590         if (state->start > end)
 591                 goto out;
 592         WARN_ON(state->end < start);
 593         last_end = state->end;
 594
 595         /* the state doesn't have the wanted bits, go ahead */
 596         if (!(state->state & bits)) {
 597                 state = next_state(state);
 598                 goto next;
 599         }
 600
 601         /*
 602          *     | ---- desired range ---- |
 603          *  | state | or
 604          *  | ------------- state -------------- |
 605          *
 606          * We need to split the extent we found, and may flip
 607          * bits on second half.
 608          *
 609          * If the extent we found extends past our range, we
 610          * just split and search again.  It'll get split again
 611          * the next time though.
 612          *
 613          * If the extent we found is inside our range, we clear
 614          * the desired bit on it.
 615          */
 616
 617         if (state->start < start) {
 618                 prealloc = alloc_extent_state_atomic(prealloc);
 619                 BUG_ON(!prealloc);
 620                 err = split_state(tree, state, prealloc, start);
 621                 if (err)
 622                         extent_io_tree_panic(tree, err);
 623
 624                 prealloc = NULL;
 625                 if (err)
 626                         goto out;
 627                 if (state->end <= end) {
 628                         state = clear_state_bit(tree, state, &bits, wake);
 629                         goto next;
 630                 }
 631                 goto search_again;
 632         }
 633         /*
 634          * | ---- desired range ---- |
 635          *                        | state |
 636          * We need to split the extent, and clear the bit
 637          * on the first half
 638          */
 639         if (state->start <= end && state->end > end) {
 640                 prealloc = alloc_extent_state_atomic(prealloc);
 641                 BUG_ON(!prealloc);
 642                 err = split_state(tree, state, prealloc, end + 1);
 643                 if (err)
 644                         extent_io_tree_panic(tree, err);
 645
 646                 if (wake)
 647                         wake_up(&state->wq);
 648
 649                 clear_state_bit(tree, prealloc, &bits, wake);
 650
 651                 prealloc = NULL;
 652                 goto out;
 653         }
 654
 655         state = clear_state_bit(tree, state, &bits, wake);
 656 next:
 657         if (last_end == (u64)-1)
 658                 goto out;
 659         start = last_end + 1;
 660         if (start <= end && state && !need_resched())
 661                 goto hit_next;
 662         goto search_again;
 663
 664 out:
 665         spin_unlock(&tree->lock);
 666         if (prealloc)
 667                 free_extent_state(prealloc);
 668
 669         return 0;
 670
 671 search_again:
 672         if (start > end)
 673                 goto out;
 674         spin_unlock(&tree->lock);
 675         if (mask & __GFP_WAIT)
 676                 cond_resched();
 677         goto again;
 678 }
 679
 680 static void wait_on_state(struct extent_io_tree *tree,
 681                           struct extent_state *state)
 682                 __releases(tree->lock)
 683                 __acquires(tree->lock)
 684 {
 685         DEFINE_WAIT(wait);
 686         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 687         spin_unlock(&tree->lock);
 688         schedule();
 689         spin_lock(&tree->lock);
 690         finish_wait(&state->wq, &wait);
 691 }
 692
 693 /*
 694  * waits for one or more bits to clear on a range in the state tree.
 695  * The range [start, end] is inclusive.
 696  * The tree lock is taken by this function
 697  */
 698 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 699                             unsigned long bits)
 700 {
 701         struct extent_state *state;
 702         struct rb_node *node;
 703
 704         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
 705
 706         spin_lock(&tree->lock);
 707 again:
 708         while (1) {
 709                 /*
 710                  * this search will find all the extents that end after
 711                  * our range starts
 712                  */
 713                 node = tree_search(tree, start);
 714                 if (!node)
 715                         break;
 716
 717                 state = rb_entry(node, struct extent_state, rb_node);
 718
 719                 if (state->start > end)
 720                         goto out;
 721
 722                 if (state->state & bits) {
 723                         start = state->start;
 724                         atomic_inc(&state->refs);
 725                         wait_on_state(tree, state);
 726                         free_extent_state(state);
 727                         goto again;
 728                 }
 729                 start = state->end + 1;
 730
 731                 if (start > end)
 732                         break;
 733
 734                 cond_resched_lock(&tree->lock);
 735         }
 736 out:
 737         spin_unlock(&tree->lock);
 738 }
 739
 740 static void set_state_bits(struct extent_io_tree *tree,
 741                            struct extent_state *state,
 742                            unsigned long *bits)
 743 {
 744         unsigned long bits_to_set = *bits & ~EXTENT_CTLBITS;
 745
 746         set_state_cb(tree, state, bits);
 747         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 748                 u64 range = state->end - state->start + 1;
 749                 tree->dirty_bytes += range;
 750         }
 751         state->state |= bits_to_set;
 752 }
 753
 754 static void cache_state(struct extent_state *state,
 755                         struct extent_state **cached_ptr)
 756 {
 757         if (cached_ptr && !(*cached_ptr)) {
 758                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 759                         *cached_ptr = state;
 760                         atomic_inc(&state->refs);
 761                 }
 762         }
 763 }
 764
 765 static void uncache_state(struct extent_state **cached_ptr)
 766 {
 767         if (cached_ptr && (*cached_ptr)) {
 768                 struct extent_state *state = *cached_ptr;
 769                 *cached_ptr = NULL;
 770                 free_extent_state(state);
 771         }
 772 }
 773
 774 /*
 775  * set some bits on a range in the tree.  This may require allocations or
 776  * sleeping, so the gfp mask is used to indicate what is allowed.
 777  *
 778  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 779  * part of the range already has the desired bits set.  The start of the
 780  * existing range is returned in failed_start in this case.
 781  *
 782  * [start, end] is inclusive This takes the tree lock.
 783  */
 784
 785 static int __must_check
 786 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 787                  unsigned long bits, unsigned long exclusive_bits,
 788                  u64 *failed_start, struct extent_state **cached_state,
 789                  gfp_t mask)
 790 {
 791         struct extent_state *state;
 792         struct extent_state *prealloc = NULL;
 793         struct rb_node *node;
 794         int err = 0;
 795         u64 last_start;
 796         u64 last_end;
 797
 798         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
 799
 800         bits |= EXTENT_FIRST_DELALLOC;
 801 again:
 802         if (!prealloc && (mask & __GFP_WAIT)) {
 803                 prealloc = alloc_extent_state(mask);
 804                 BUG_ON(!prealloc);
 805         }
 806
 807         spin_lock(&tree->lock);
 808         if (cached_state && *cached_state) {
 809                 state = *cached_state;
 810                 if (state->start <= start && state->end > start &&
 811                     state->tree) {
 812                         node = &state->rb_node;
 813                         goto hit_next;
 814                 }
 815         }
 816         /*
 817          * this search will find all the extents that end after
 818          * our range starts.
 819          */
 820         node = tree_search(tree, start);
 821         if (!node) {
 822                 prealloc = alloc_extent_state_atomic(prealloc);
 823                 BUG_ON(!prealloc);
 824                 err = insert_state(tree, prealloc, start, end, &bits);
 825                 if (err)
 826                         extent_io_tree_panic(tree, err);
 827
 828                 prealloc = NULL;
 829                 goto out;
 830         }
 831         state = rb_entry(node, struct extent_state, rb_node);
 832 hit_next:
 833         last_start = state->start;
 834         last_end = state->end;
 835
 836         /*
 837          * | ---- desired range ---- |
 838          * | state |
 839          *
 840          * Just lock what we found and keep going
 841          */
 842         if (state->start == start && state->end <= end) {
 843                 if (state->state & exclusive_bits) {
 844                         *failed_start = state->start;
 845                         err = -EEXIST;
 846                         goto out;
 847                 }
 848
 849                 set_state_bits(tree, state, &bits);
 850                 cache_state(state, cached_state);
 851                 merge_state(tree, state);
 852                 if (last_end == (u64)-1)
 853                         goto out;
 854                 start = last_end + 1;
 855                 state = next_state(state);
 856                 if (start < end && state && state->start == start &&
 857                     !need_resched())
 858                         goto hit_next;
 859                 goto search_again;
 860         }
 861
 862         /*
 863          *     | ---- desired range ---- |
 864          * | state |
 865          *   or
 866          * | ------------- state -------------- |
 867          *
 868          * We need to split the extent we found, and may flip bits on
 869          * second half.
 870          *
 871          * If the extent we found extends past our
 872          * range, we just split and search again.  It'll get split
 873          * again the next time though.
 874          *
 875          * If the extent we found is inside our range, we set the
 876          * desired bit on it.
 877          */
 878         if (state->start < start) {
 879                 if (state->state & exclusive_bits) {
 880                         *failed_start = start;
 881                         err = -EEXIST;
 882                         goto out;
 883                 }
 884
 885                 prealloc = alloc_extent_state_atomic(prealloc);
 886                 BUG_ON(!prealloc);
 887                 err = split_state(tree, state, prealloc, start);
 888                 if (err)
 889                         extent_io_tree_panic(tree, err);
 890
 891                 prealloc = NULL;
 892                 if (err)
 893                         goto out;
 894                 if (state->end <= end) {
 895                         set_state_bits(tree, state, &bits);
 896                         cache_state(state, cached_state);
 897                         merge_state(tree, state);
 898                         if (last_end == (u64)-1)
 899                                 goto out;
 900                         start = last_end + 1;
 901                         state = next_state(state);
 902                         if (start < end && state && state->start == start &&
 903                             !need_resched())
 904                                 goto hit_next;
 905                 }
 906                 goto search_again;
 907         }
 908         /*
 909          * | ---- desired range ---- |
 910          *     | state | or               | state |
 911          *
 912          * There's a hole, we need to insert something in it and
 913          * ignore the extent we found.
 914          */
 915         if (state->start > start) {
 916                 u64 this_end;
 917                 if (end < last_start)
 918                         this_end = end;
 919                 else
 920                         this_end = last_start - 1;
 921
 922                 prealloc = alloc_extent_state_atomic(prealloc);
 923                 BUG_ON(!prealloc);
 924
 925                 /*
 926                  * Avoid to free 'prealloc' if it can be merged with
 927                  * the later extent.
 928                  */
 929                 err = insert_state(tree, prealloc, start, this_end,
 930                                    &bits);
 931                 if (err)
 932                         extent_io_tree_panic(tree, err);
 933
 934                 cache_state(prealloc, cached_state);
 935                 prealloc = NULL;
 936                 start = this_end + 1;
 937                 goto search_again;
 938         }
 939         /*
 940          * | ---- desired range ---- |
 941          *                        | state |
 942          * We need to split the extent, and set the bit
 943          * on the first half
 944          */
 945         if (state->start <= end && state->end > end) {
 946                 if (state->state & exclusive_bits) {
 947                         *failed_start = start;
 948                         err = -EEXIST;
 949                         goto out;
 950                 }
 951
 952                 prealloc = alloc_extent_state_atomic(prealloc);
 953                 BUG_ON(!prealloc);
 954                 err = split_state(tree, state, prealloc, end + 1);
 955                 if (err)
 956                         extent_io_tree_panic(tree, err);
 957
 958                 set_state_bits(tree, prealloc, &bits);
 959                 cache_state(prealloc, cached_state);
 960                 merge_state(tree, prealloc);
 961                 prealloc = NULL;
 962                 goto out;
 963         }
 964
 965         goto search_again;
 966
 967 out:
 968         spin_unlock(&tree->lock);
 969         if (prealloc)
 970                 free_extent_state(prealloc);
 971
 972         return err;
 973
 974 search_again:
 975         if (start > end)
 976                 goto out;
 977         spin_unlock(&tree->lock);
 978         if (mask & __GFP_WAIT)
 979                 cond_resched();
 980         goto again;
 981 }
 982
 983 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 984                    unsigned long bits, u64 * failed_start,
 985                    struct extent_state **cached_state, gfp_t mask)
 986 {
 987         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
 988                                 cached_state, mask);
 989 }
 990
 991
 992 /**
 993  * convert_extent_bit - convert all bits in a given range from one bit to
 994  *                      another
 995  * @tree:       the io tree to search
 996  * @start:      the start offset in bytes
 997  * @end:        the end offset in bytes (inclusive)
 998  * @bits:       the bits to set in this range
 999  * @clear_bits: the bits to clear in this range
1000  * @cached_state:       state that we're going to cache
1001  * @mask:       the allocation mask
1002  *
1003  * This will go through and set bits for the given range.  If any states exist
1004  * already in this range they are set with the given bit and cleared of the
1005  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1006  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1007  * boundary bits like LOCK.
1008  */
1009 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1010                        unsigned long bits, unsigned long clear_bits,
1011                        struct extent_state **cached_state, gfp_t mask)
1012 {
1013         struct extent_state *state;
1014         struct extent_state *prealloc = NULL;
1015         struct rb_node *node;
1016         int err = 0;
1017         u64 last_start;
1018         u64 last_end;
1019
1020         btrfs_debug_check_extent_io_range(tree->mapping->host, start, end);
1021
1022 again:
1023         if (!prealloc && (mask & __GFP_WAIT)) {
1024                 prealloc = alloc_extent_state(mask);
1025                 if (!prealloc)
1026                         return -ENOMEM;
1027         }
1028
1029         spin_lock(&tree->lock);
1030         if (cached_state && *cached_state) {
1031                 state = *cached_state;
1032                 if (state->start <= start && state->end > start &&
1033                     state->tree) {
1034                         node = &state->rb_node;
1035                         goto hit_next;
1036                 }
1037         }
1038
1039         /*
1040          * this search will find all the extents that end after
1041          * our range starts.
1042          */
1043         node = tree_search(tree, start);
1044         if (!node) {
1045                 prealloc = alloc_extent_state_atomic(prealloc);
1046                 if (!prealloc) {
1047                         err = -ENOMEM;
1048                         goto out;
1049                 }
1050                 err = insert_state(tree, prealloc, start, end, &bits);
1051                 prealloc = NULL;
1052                 if (err)
1053                         extent_io_tree_panic(tree, err);
1054                 goto out;
1055         }
1056         state = rb_entry(node, struct extent_state, rb_node);
1057 hit_next:
1058         last_start = state->start;
1059         last_end = state->end;
1060
1061         /*
1062          * | ---- desired range ---- |
1063          * | state |
1064          *
1065          * Just lock what we found and keep going
1066          */
1067         if (state->start == start && state->end <= end) {
1068                 set_state_bits(tree, state, &bits);
1069                 cache_state(state, cached_state);
1070                 state = clear_state_bit(tree, state, &clear_bits, 0);
1071                 if (last_end == (u64)-1)
1072                         goto out;
1073                 start = last_end + 1;
1074                 if (start < end && state && state->start == start &&
1075                     !need_resched())
1076                         goto hit_next;
1077                 goto search_again;
1078         }
1079
1080         /*
1081          *     | ---- desired range ---- |
1082          * | state |
1083          *   or
1084          * | ------------- state -------------- |
1085          *
1086          * We need to split the extent we found, and may flip bits on
1087          * second half.
1088          *
1089          * If the extent we found extends past our
1090          * range, we just split and search again.  It'll get split
1091          * again the next time though.
1092          *
1093          * If the extent we found is inside our range, we set the
1094          * desired bit on it.
1095          */
1096         if (state->start < start) {
1097                 prealloc = alloc_extent_state_atomic(prealloc);
1098                 if (!prealloc) {
1099                         err = -ENOMEM;
1100                         goto out;
1101                 }
1102                 err = split_state(tree, state, prealloc, start);
1103                 if (err)
1104                         extent_io_tree_panic(tree, err);
1105                 prealloc = NULL;
1106                 if (err)
1107                         goto out;
1108                 if (state->end <= end) {
1109                         set_state_bits(tree, state, &bits);
1110                         cache_state(state, cached_state);
1111                         state = clear_state_bit(tree, state, &clear_bits, 0);
1112                         if (last_end == (u64)-1)
1113                                 goto out;
1114                         start = last_end + 1;
1115                         if (start < end && state && state->start == start &&
1116                             !need_resched())
1117                                 goto hit_next;
1118                 }
1119                 goto search_again;
1120         }
1121         /*
1122          * | ---- desired range ---- |
1123          *     | state | or               | state |
1124          *
1125          * There's a hole, we need to insert something in it and
1126          * ignore the extent we found.
1127          */
1128         if (state->start > start) {
1129                 u64 this_end;
1130                 if (end < last_start)
1131                         this_end = end;
1132                 else
1133                         this_end = last_start - 1;
1134
1135                 prealloc = alloc_extent_state_atomic(prealloc);
1136                 if (!prealloc) {
1137                         err = -ENOMEM;
1138                         goto out;
1139                 }
1140
1141                 /*
1142                  * Avoid to free 'prealloc' if it can be merged with
1143                  * the later extent.
1144                  */
1145                 err = insert_state(tree, prealloc, start, this_end,
1146                                    &bits);
1147                 if (err)
1148                         extent_io_tree_panic(tree, err);
1149                 cache_state(prealloc, cached_state);
1150                 prealloc = NULL;
1151                 start = this_end + 1;
1152                 goto search_again;
1153         }
1154         /*
1155          * | ---- desired range ---- |
1156          *                        | state |
1157          * We need to split the extent, and set the bit
1158          * on the first half
1159          */
1160         if (state->start <= end && state->end > end) {
1161                 prealloc = alloc_extent_state_atomic(prealloc);
1162                 if (!prealloc) {
1163                         err = -ENOMEM;
1164                         goto out;
1165                 }
1166
1167                 err = split_state(tree, state, prealloc, end + 1);
1168                 if (err)
1169                         extent_io_tree_panic(tree, err);
1170
1171                 set_state_bits(tree, prealloc, &bits);
1172                 cache_state(prealloc, cached_state);
1173                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1174                 prealloc = NULL;
1175                 goto out;
1176         }
1177
1178         goto search_again;
1179
1180 out:
1181         spin_unlock(&tree->lock);
1182         if (prealloc)
1183                 free_extent_state(prealloc);
1184
1185         return err;
1186
1187 search_again:
1188         if (start > end)
1189                 goto out;
1190         spin_unlock(&tree->lock);
1191         if (mask & __GFP_WAIT)
1192                 cond_resched();
1193         goto again;
1194 }
1195
1196 /* wrappers around set/clear extent bit */
1197 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1198                      gfp_t mask)
1199 {
1200         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1201                               NULL, mask);
1202 }
1203
1204 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1205                     unsigned long bits, gfp_t mask)
1206 {
1207         return set_extent_bit(tree, start, end, bits, NULL,
1208                               NULL, mask);
1209 }
1210
1211 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1212                       unsigned long bits, gfp_t mask)
1213 {
1214         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1215 }
1216
1217 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1218                         struct extent_state **cached_state, gfp_t mask)
1219 {
1220         return set_extent_bit(tree, start, end,
1221                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1222                               NULL, cached_state, mask);
1223 }
1224
1225 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1226                       struct extent_state **cached_state, gfp_t mask)
1227 {
1228         return set_extent_bit(tree, start, end,
1229                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1230                               NULL, cached_state, mask);
1231 }
1232
1233 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1234                        gfp_t mask)
1235 {
1236         return clear_extent_bit(tree, start, end,
1237                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1238                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1239 }
1240
1241 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1242                      gfp_t mask)
1243 {
1244         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1245                               NULL, mask);
1246 }
1247
1248 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1249                         struct extent_state **cached_state, gfp_t mask)
1250 {
1251         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1252                               cached_state, mask);
1253 }
1254
1255 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1256                           struct extent_state **cached_state, gfp_t mask)
1257 {
1258         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1259                                 cached_state, mask);
1260 }
1261
1262 /*
1263  * either insert or lock state struct between start and end use mask to tell
1264  * us if waiting is desired.
1265  */
1266 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1267                      unsigned long bits, struct extent_state **cached_state)
1268 {
1269         int err;
1270         u64 failed_start;
1271         while (1) {
1272                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1273                                        EXTENT_LOCKED, &failed_start,
1274                                        cached_state, GFP_NOFS);
1275                 if (err == -EEXIST) {
1276                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1277                         start = failed_start;
1278                 } else
1279                         break;
1280                 WARN_ON(start > end);
1281         }
1282         return err;
1283 }
1284
1285 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1286 {
1287         return lock_extent_bits(tree, start, end, 0, NULL);
1288 }
1289
1290 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1291 {
1292         int err;
1293         u64 failed_start;
1294
1295         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1296                                &failed_start, NULL, GFP_NOFS);
1297         if (err == -EEXIST) {
1298                 if (failed_start > start)
1299                         clear_extent_bit(tree, start, failed_start - 1,
1300                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1301                 return 0;
1302         }
1303         return 1;
1304 }
1305
1306 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1307                          struct extent_state **cached, gfp_t mask)
1308 {
1309         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1310                                 mask);
1311 }
1312
1313 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1314 {
1315         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1316                                 GFP_NOFS);
1317 }
1318
1319 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1320 {
1321         unsigned long index = start >> PAGE_CACHE_SHIFT;
1322         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1323         struct page *page;
1324
1325         while (index <= end_index) {
1326                 page = find_get_page(inode->i_mapping, index);
1327                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1328                 clear_page_dirty_for_io(page);
1329                 page_cache_release(page);
1330                 index++;
1331         }
1332         return 0;
1333 }
1334
1335 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1336 {
1337         unsigned long index = start >> PAGE_CACHE_SHIFT;
1338         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1339         struct page *page;
1340
1341         while (index <= end_index) {
1342                 page = find_get_page(inode->i_mapping, index);
1343                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1344                 account_page_redirty(page);
1345                 __set_page_dirty_nobuffers(page);
1346                 page_cache_release(page);
1347                 index++;
1348         }
1349         return 0;
1350 }
1351
1352 /*
1353  * helper function to set both pages and extents in the tree writeback
1354  */
1355 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1356 {
1357         unsigned long index = start >> PAGE_CACHE_SHIFT;
1358         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1359         struct page *page;
1360
1361         while (index <= end_index) {
1362                 page = find_get_page(tree->mapping, index);
1363                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1364                 set_page_writeback(page);
1365                 page_cache_release(page);
1366                 index++;
1367         }
1368         return 0;
1369 }
1370
1371 /* find the first state struct with 'bits' set after 'start', and
1372  * return it.  tree->lock must be held.  NULL will returned if
1373  * nothing was found after 'start'
1374  */
1375 static struct extent_state *
1376 find_first_extent_bit_state(struct extent_io_tree *tree,
1377                             u64 start, unsigned long bits)
1378 {
1379         struct rb_node *node;
1380         struct extent_state *state;
1381
1382         /*
1383          * this search will find all the extents that end after
1384          * our range starts.
1385          */
1386         node = tree_search(tree, start);
1387         if (!node)
1388                 goto out;
1389
1390         while (1) {
1391                 state = rb_entry(node, struct extent_state, rb_node);
1392                 if (state->end >= start && (state->state & bits))
1393                         return state;
1394
1395                 node = rb_next(node);
1396                 if (!node)
1397                         break;
1398         }
1399 out:
1400         return NULL;
1401 }
1402
1403 /*
1404  * find the first offset in the io tree with 'bits' set. zero is
1405  * returned if we find something, and *start_ret and *end_ret are
1406  * set to reflect the state struct that was found.
1407  *
1408  * If nothing was found, 1 is returned. If found something, return 0.
1409  */
1410 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1411                           u64 *start_ret, u64 *end_ret, unsigned long bits,
1412                           struct extent_state **cached_state)
1413 {
1414         struct extent_state *state;
1415         struct rb_node *n;
1416         int ret = 1;
1417
1418         spin_lock(&tree->lock);
1419         if (cached_state && *cached_state) {
1420                 state = *cached_state;
1421                 if (state->end == start - 1 && state->tree) {
1422                         n = rb_next(&state->rb_node);
1423                         while (n) {
1424                                 state = rb_entry(n, struct extent_state,
1425                                                  rb_node);
1426                                 if (state->state & bits)
1427                                         goto got_it;
1428                                 n = rb_next(n);
1429                         }
1430                         free_extent_state(*cached_state);
1431                         *cached_state = NULL;
1432                         goto out;
1433                 }
1434                 free_extent_state(*cached_state);
1435                 *cached_state = NULL;
1436         }
1437
1438         state = find_first_extent_bit_state(tree, start, bits);
1439 got_it:
1440         if (state) {
1441                 cache_state(state, cached_state);
1442                 *start_ret = state->start;
1443                 *end_ret = state->end;
1444                 ret = 0;
1445         }
1446 out:
1447         spin_unlock(&tree->lock);
1448         return ret;
1449 }
1450
1451 /*
1452  * find a contiguous range of bytes in the file marked as delalloc, not
1453  * more than 'max_bytes'.  start and end are used to return the range,
1454  *
1455  * 1 is returned if we find something, 0 if nothing was in the tree
1456  */
1457 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1458                                         u64 *start, u64 *end, u64 max_bytes,
1459                                         struct extent_state **cached_state)
1460 {
1461         struct rb_node *node;
1462         struct extent_state *state;
1463         u64 cur_start = *start;
1464         u64 found = 0;
1465         u64 total_bytes = 0;
1466
1467         spin_lock(&tree->lock);
1468
1469         /*
1470          * this search will find all the extents that end after
1471          * our range starts.
1472          */
1473         node = tree_search(tree, cur_start);
1474         if (!node) {
1475                 if (!found)
1476                         *end = (u64)-1;
1477                 goto out;
1478         }
1479
1480         while (1) {
1481                 state = rb_entry(node, struct extent_state, rb_node);
1482                 if (found && (state->start != cur_start ||
1483                               (state->state & EXTENT_BOUNDARY))) {
1484                         goto out;
1485                 }
1486                 if (!(state->state & EXTENT_DELALLOC)) {
1487                         if (!found)
1488                                 *end = state->end;
1489                         goto out;
1490                 }
1491                 if (!found) {
1492                         *start = state->start;
1493                         *cached_state = state;
1494                         atomic_inc(&state->refs);
1495                 }
1496                 found++;
1497                 *end = state->end;
1498                 cur_start = state->end + 1;
1499                 node = rb_next(node);
1500                 if (!node)
1501                         break;
1502                 total_bytes += state->end - state->start + 1;
1503                 if (total_bytes >= max_bytes)
1504                         break;
1505         }
1506 out:
1507         spin_unlock(&tree->lock);
1508         return found;
1509 }
1510
1511 static noinline void __unlock_for_delalloc(struct inode *inode,
1512                                            struct page *locked_page,
1513                                            u64 start, u64 end)
1514 {
1515         int ret;
1516         struct page *pages[16];
1517         unsigned long index = start >> PAGE_CACHE_SHIFT;
1518         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1519         unsigned long nr_pages = end_index - index + 1;
1520         int i;
1521
1522         if (index == locked_page->index && end_index == index)
1523                 return;
1524
1525         while (nr_pages > 0) {
1526                 ret = find_get_pages_contig(inode->i_mapping, index,
1527                                      min_t(unsigned long, nr_pages,
1528                                      ARRAY_SIZE(pages)), pages);
1529                 for (i = 0; i < ret; i++) {
1530                         if (pages[i] != locked_page)
1531                                 unlock_page(pages[i]);
1532                         page_cache_release(pages[i]);
1533                 }
1534                 nr_pages -= ret;
1535                 index += ret;
1536                 cond_resched();
1537         }
1538 }
1539
1540 static noinline int lock_delalloc_pages(struct inode *inode,
1541                                         struct page *locked_page,
1542                                         u64 delalloc_start,
1543                                         u64 delalloc_end)
1544 {
1545         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1546         unsigned long start_index = index;
1547         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1548         unsigned long pages_locked = 0;
1549         struct page *pages[16];
1550         unsigned long nrpages;
1551         int ret;
1552         int i;
1553
1554         /* the caller is responsible for locking the start index */
1555         if (index == locked_page->index && index == end_index)
1556                 return 0;
1557
1558         /* skip the page at the start index */
1559         nrpages = end_index - index + 1;
1560         while (nrpages > 0) {
1561                 ret = find_get_pages_contig(inode->i_mapping, index,
1562                                      min_t(unsigned long,
1563                                      nrpages, ARRAY_SIZE(pages)), pages);
1564                 if (ret == 0) {
1565                         ret = -EAGAIN;
1566                         goto done;
1567                 }
1568                 /* now we have an array of pages, lock them all */
1569                 for (i = 0; i < ret; i++) {
1570                         /*
1571                          * the caller is taking responsibility for
1572                          * locked_page
1573                          */
1574                         if (pages[i] != locked_page) {
1575                                 lock_page(pages[i]);
1576                                 if (!PageDirty(pages[i]) ||
1577                                     pages[i]->mapping != inode->i_mapping) {
1578                                         ret = -EAGAIN;
1579                                         unlock_page(pages[i]);
1580                                         page_cache_release(pages[i]);
1581                                         goto done;
1582                                 }
1583                         }
1584                         page_cache_release(pages[i]);
1585                         pages_locked++;
1586                 }
1587                 nrpages -= ret;
1588                 index += ret;
1589                 cond_resched();
1590         }
1591         ret = 0;
1592 done:
1593         if (ret && pages_locked) {
1594                 __unlock_for_delalloc(inode, locked_page,
1595                               delalloc_start,
1596                               ((u64)(start_index + pages_locked - 1)) <<
1597                               PAGE_CACHE_SHIFT);
1598         }
1599         return ret;
1600 }
1601
1602 /*
1603  * find a contiguous range of bytes in the file marked as delalloc, not
1604  * more than 'max_bytes'.  start and end are used to return the range,
1605  *
1606  * 1 is returned if we find something, 0 if nothing was in the tree
1607  */
1608 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1609                                              struct extent_io_tree *tree,
1610                                              struct page *locked_page,
1611                                              u64 *start, u64 *end,
1612                                              u64 max_bytes)
1613 {
1614         u64 delalloc_start;
1615         u64 delalloc_end;
1616         u64 found;
1617         struct extent_state *cached_state = NULL;
1618         int ret;
1619         int loops = 0;
1620
1621 again:
1622         /* step one, find a bunch of delalloc bytes starting at start */
1623         delalloc_start = *start;
1624         delalloc_end = 0;
1625         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1626                                     max_bytes, &cached_state);
1627         if (!found || delalloc_end <= *start) {
1628                 *start = delalloc_start;
1629                 *end = delalloc_end;
1630                 free_extent_state(cached_state);
1631                 return found;
1632         }
1633
1634         /*
1635          * start comes from the offset of locked_page.  We have to lock
1636          * pages in order, so we can't process delalloc bytes before
1637          * locked_page
1638          */
1639         if (delalloc_start < *start)
1640                 delalloc_start = *start;
1641
1642         /*
1643          * make sure to limit the number of pages we try to lock down
1644          * if we're looping.
1645          */
1646         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1647                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1648
1649         /* step two, lock all the pages after the page that has start */
1650         ret = lock_delalloc_pages(inode, locked_page,
1651                                   delalloc_start, delalloc_end);
1652         if (ret == -EAGAIN) {
1653                 /* some of the pages are gone, lets avoid looping by
1654                  * shortening the size of the delalloc range we're searching
1655                  */
1656                 free_extent_state(cached_state);
1657                 if (!loops) {
1658                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1659                         max_bytes = PAGE_CACHE_SIZE - offset;
1660                         loops = 1;
1661                         goto again;
1662                 } else {
1663                         found = 0;
1664                         goto out_failed;
1665                 }
1666         }
1667         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1668
1669         /* step three, lock the state bits for the whole range */
1670         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1671
1672         /* then test to make sure it is all still delalloc */
1673         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1674                              EXTENT_DELALLOC, 1, cached_state);
1675         if (!ret) {
1676                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1677                                      &cached_state, GFP_NOFS);
1678                 __unlock_for_delalloc(inode, locked_page,
1679                               delalloc_start, delalloc_end);
1680                 cond_resched();
1681                 goto again;
1682         }
1683         free_extent_state(cached_state);
1684         *start = delalloc_start;
1685         *end = delalloc_end;
1686 out_failed:
1687         return found;
1688 }
1689
1690 int extent_clear_unlock_delalloc(struct inode *inode,
1691                                 struct extent_io_tree *tree,
1692                                 u64 start, u64 end, struct page *locked_page,
1693                                 unsigned long op)
1694 {
1695         int ret;
1696         struct page *pages[16];
1697         unsigned long index = start >> PAGE_CACHE_SHIFT;
1698         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1699         unsigned long nr_pages = end_index - index + 1;
1700         int i;
1701         unsigned long clear_bits = 0;
1702
1703         if (op & EXTENT_CLEAR_UNLOCK)
1704                 clear_bits |= EXTENT_LOCKED;
1705         if (op & EXTENT_CLEAR_DIRTY)
1706                 clear_bits |= EXTENT_DIRTY;
1707
1708         if (op & EXTENT_CLEAR_DELALLOC)
1709                 clear_bits |= EXTENT_DELALLOC;
1710
1711         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1712         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1713                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1714                     EXTENT_SET_PRIVATE2)))
1715                 return 0;
1716
1717         while (nr_pages > 0) {
1718                 ret = find_get_pages_contig(inode->i_mapping, index,
1719                                      min_t(unsigned long,
1720                                      nr_pages, ARRAY_SIZE(pages)), pages);
1721                 for (i = 0; i < ret; i++) {
1722
1723                         if (op & EXTENT_SET_PRIVATE2)
1724                                 SetPagePrivate2(pages[i]);
1725
1726                         if (pages[i] == locked_page) {
1727                                 page_cache_release(pages[i]);
1728                                 continue;
1729                         }
1730                         if (op & EXTENT_CLEAR_DIRTY)
1731                                 clear_page_dirty_for_io(pages[i]);
1732                         if (op & EXTENT_SET_WRITEBACK)
1733                                 set_page_writeback(pages[i]);
1734                         if (op & EXTENT_END_WRITEBACK)
1735                                 end_page_writeback(pages[i]);
1736                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1737                                 unlock_page(pages[i]);
1738                         page_cache_release(pages[i]);
1739                 }
1740                 nr_pages -= ret;
1741                 index += ret;
1742                 cond_resched();
1743         }
1744         return 0;
1745 }
1746
1747 /*
1748  * count the number of bytes in the tree that have a given bit(s)
1749  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1750  * cached.  The total number found is returned.
1751  */
1752 u64 count_range_bits(struct extent_io_tree *tree,
1753                      u64 *start, u64 search_end, u64 max_bytes,
1754                      unsigned long bits, int contig)
1755 {
1756         struct rb_node *node;
1757         struct extent_state *state;
1758         u64 cur_start = *start;
1759         u64 total_bytes = 0;
1760         u64 last = 0;
1761         int found = 0;
1762
1763         if (search_end <= cur_start) {
1764                 WARN_ON(1);
1765                 return 0;
1766         }
1767
1768         spin_lock(&tree->lock);
1769         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1770                 total_bytes = tree->dirty_bytes;
1771                 goto out;
1772         }
1773         /*
1774          * this search will find all the extents that end after
1775          * our range starts.
1776          */
1777         node = tree_search(tree, cur_start);
1778         if (!node)
1779                 goto out;
1780
1781         while (1) {
1782                 state = rb_entry(node, struct extent_state, rb_node);
1783                 if (state->start > search_end)
1784                         break;
1785                 if (contig && found && state->start > last + 1)
1786                         break;
1787                 if (state->end >= cur_start && (state->state & bits) == bits) {
1788                         total_bytes += min(search_end, state->end) + 1 -
1789                                        max(cur_start, state->start);
1790                         if (total_bytes >= max_bytes)
1791                                 break;
1792                         if (!found) {
1793                                 *start = max(cur_start, state->start);
1794                                 found = 1;
1795                         }
1796                         last = state->end;
1797                 } else if (contig && found) {
1798                         break;
1799                 }
1800                 node = rb_next(node);
1801                 if (!node)
1802                         break;
1803         }
1804 out:
1805         spin_unlock(&tree->lock);
1806         return total_bytes;
1807 }
1808
1809 /*
1810  * set the private field for a given byte offset in the tree.  If there isn't
1811  * an extent_state there already, this does nothing.
1812  */
1813 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1814 {
1815         struct rb_node *node;
1816         struct extent_state *state;
1817         int ret = 0;
1818
1819         spin_lock(&tree->lock);
1820         /*
1821          * this search will find all the extents that end after
1822          * our range starts.
1823          */
1824         node = tree_search(tree, start);
1825         if (!node) {
1826                 ret = -ENOENT;
1827                 goto out;
1828         }
1829         state = rb_entry(node, struct extent_state, rb_node);
1830         if (state->start != start) {
1831                 ret = -ENOENT;
1832                 goto out;
1833         }
1834         state->private = private;
1835 out:
1836         spin_unlock(&tree->lock);
1837         return ret;
1838 }
1839
1840 void extent_cache_csums_dio(struct extent_io_tree *tree, u64 start, u32 csums[],
1841                             int count)
1842 {
1843         struct rb_node *node;
1844         struct extent_state *state;
1845
1846         spin_lock(&tree->lock);
1847         /*
1848          * this search will find all the extents that end after
1849          * our range starts.
1850          */
1851         node = tree_search(tree, start);
1852         BUG_ON(!node);
1853
1854         state = rb_entry(node, struct extent_state, rb_node);
1855         BUG_ON(state->start != start);
1856
1857         while (count) {
1858                 state->private = *csums++;
1859                 count--;
1860                 state = next_state(state);
1861         }
1862         spin_unlock(&tree->lock);
1863 }
1864
1865 static inline u64 __btrfs_get_bio_offset(struct bio *bio, int bio_index)
1866 {
1867         struct bio_vec *bvec = bio->bi_io_vec + bio_index;
1868
1869         return page_offset(bvec->bv_page) + bvec->bv_offset;
1870 }
1871
1872 void extent_cache_csums(struct extent_io_tree *tree, struct bio *bio, int bio_index,
1873                         u32 csums[], int count)
1874 {
1875         struct rb_node *node;
1876         struct extent_state *state = NULL;
1877         u64 start;
1878
1879         spin_lock(&tree->lock);
1880         do {
1881                 start = __btrfs_get_bio_offset(bio, bio_index);
1882                 if (state == NULL || state->start != start) {
1883                         node = tree_search(tree, start);
1884                         BUG_ON(!node);
1885
1886                         state = rb_entry(node, struct extent_state, rb_node);
1887                         BUG_ON(state->start != start);
1888                 }
1889                 state->private = *csums++;
1890                 count--;
1891                 bio_index++;
1892
1893                 state = next_state(state);
1894         } while (count);
1895         spin_unlock(&tree->lock);
1896 }
1897
1898 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1899 {
1900         struct rb_node *node;
1901         struct extent_state *state;
1902         int ret = 0;
1903
1904         spin_lock(&tree->lock);
1905         /*
1906          * this search will find all the extents that end after
1907          * our range starts.
1908          */
1909         node = tree_search(tree, start);
1910         if (!node) {
1911                 ret = -ENOENT;
1912                 goto out;
1913         }
1914         state = rb_entry(node, struct extent_state, rb_node);
1915         if (state->start != start) {
1916                 ret = -ENOENT;
1917                 goto out;
1918         }
1919         *private = state->private;
1920 out:
1921         spin_unlock(&tree->lock);
1922         return ret;
1923 }
1924
1925 /*
1926  * searches a range in the state tree for a given mask.
1927  * If 'filled' == 1, this returns 1 only if every extent in the tree
1928  * has the bits set.  Otherwise, 1 is returned if any bit in the
1929  * range is found set.
1930  */
1931 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1932                    unsigned long bits, int filled, struct extent_state *cached)
1933 {
1934         struct extent_state *state = NULL;
1935         struct rb_node *node;
1936         int bitset = 0;
1937
1938         spin_lock(&tree->lock);
1939         if (cached && cached->tree && cached->start <= start &&
1940             cached->end > start)
1941                 node = &cached->rb_node;
1942         else
1943                 node = tree_search(tree, start);
1944         while (node && start <= end) {
1945                 state = rb_entry(node, struct extent_state, rb_node);
1946
1947                 if (filled && state->start > start) {
1948                         bitset = 0;
1949                         break;
1950                 }
1951
1952                 if (state->start > end)
1953                         break;
1954
1955                 if (state->state & bits) {
1956                         bitset = 1;
1957                         if (!filled)
1958                                 break;
1959                 } else if (filled) {
1960                         bitset = 0;
1961                         break;
1962                 }
1963
1964                 if (state->end == (u64)-1)
1965                         break;
1966
1967                 start = state->end + 1;
1968                 if (start > end)
1969                         break;
1970                 node = rb_next(node);
1971                 if (!node) {
1972                         if (filled)
1973                                 bitset = 0;
1974                         break;
1975                 }
1976         }
1977         spin_unlock(&tree->lock);
1978         return bitset;
1979 }
1980
1981 /*
1982  * helper function to set a given page up to date if all the
1983  * extents in the tree for that page are up to date
1984  */
1985 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1986 {
1987         u64 start = page_offset(page);
1988         u64 end = start + PAGE_CACHE_SIZE - 1;
1989         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1990                 SetPageUptodate(page);
1991 }
1992
1993 /*
1994  * When IO fails, either with EIO or csum verification fails, we
1995  * try other mirrors that might have a good copy of the data.  This
1996  * io_failure_record is used to record state as we go through all the
1997  * mirrors.  If another mirror has good data, the page is set up to date
1998  * and things continue.  If a good mirror can't be found, the original
1999  * bio end_io callback is called to indicate things have failed.
2000  */
2001 struct io_failure_record {
2002         struct page *page;
2003         u64 start;
2004         u64 len;
2005         u64 logical;
2006         unsigned long bio_flags;
2007         int this_mirror;
2008         int failed_mirror;
2009         int in_validation;
2010 };
2011
2012 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
2013                                 int did_repair)
2014 {
2015         int ret;
2016         int err = 0;
2017         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2018
2019         set_state_private(failure_tree, rec->start, 0);
2020         ret = clear_extent_bits(failure_tree, rec->start,
2021                                 rec->start + rec->len - 1,
2022                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2023         if (ret)
2024                 err = ret;
2025
2026         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
2027                                 rec->start + rec->len - 1,
2028                                 EXTENT_DAMAGED, GFP_NOFS);
2029         if (ret && !err)
2030                 err = ret;
2031
2032         kfree(rec);
2033         return err;
2034 }
2035
2036 static void repair_io_failure_callback(struct bio *bio, int err)
2037 {
2038         complete(bio->bi_private);
2039 }
2040
2041 /*
2042  * this bypasses the standard btrfs submit functions deliberately, as
2043  * the standard behavior is to write all copies in a raid setup. here we only
2044  * want to write the one bad copy. so we do the mapping for ourselves and issue
2045  * submit_bio directly.
2046  * to avoid any synchronization issues, wait for the data after writing, which
2047  * actually prevents the read that triggered the error from finishing.
2048  * currently, there can be no more than two copies of every data bit. thus,
2049  * exactly one rewrite is required.
2050  */
2051 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 start,
2052                         u64 length, u64 logical, struct page *page,
2053                         int mirror_num)
2054 {
2055         struct bio *bio;
2056         struct btrfs_device *dev;
2057         DECLARE_COMPLETION_ONSTACK(compl);
2058         u64 map_length = 0;
2059         u64 sector;
2060         struct btrfs_bio *bbio = NULL;
2061         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2062         int ret;
2063
2064         BUG_ON(!mirror_num);
2065
2066         /* we can't repair anything in raid56 yet */
2067         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2068                 return 0;
2069
2070         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2071         if (!bio)
2072                 return -EIO;
2073         bio->bi_private = &compl;
2074         bio->bi_end_io = repair_io_failure_callback;
2075         bio->bi_size = 0;
2076         map_length = length;
2077
2078         ret = btrfs_map_block(fs_info, WRITE, logical,
2079                               &map_length, &bbio, mirror_num);
2080         if (ret) {
2081                 bio_put(bio);
2082                 return -EIO;
2083         }
2084         BUG_ON(mirror_num != bbio->mirror_num);
2085         sector = bbio->stripes[mirror_num-1].physical >> 9;
2086         bio->bi_sector = sector;
2087         dev = bbio->stripes[mirror_num-1].dev;
2088         kfree(bbio);
2089         if (!dev || !dev->bdev || !dev->writeable) {
2090                 bio_put(bio);
2091                 return -EIO;
2092         }
2093         bio->bi_bdev = dev->bdev;
2094         bio_add_page(bio, page, length, start - page_offset(page));
2095         btrfsic_submit_bio(WRITE_SYNC, bio);
2096         wait_for_completion(&compl);
2097
2098         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2099                 /* try to remap that extent elsewhere? */
2100                 bio_put(bio);
2101                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2102                 return -EIO;
2103         }
2104
2105         printk_ratelimited_in_rcu(KERN_INFO "btrfs read error corrected: ino %lu off %llu "
2106                       "(dev %s sector %llu)\n", page->mapping->host->i_ino,
2107                       start, rcu_str_deref(dev->name), sector);
2108
2109         bio_put(bio);
2110         return 0;
2111 }
2112
2113 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2114                          int mirror_num)
2115 {
2116         u64 start = eb->start;
2117         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2118         int ret = 0;
2119
2120         for (i = 0; i < num_pages; i++) {
2121                 struct page *p = extent_buffer_page(eb, i);
2122                 ret = repair_io_failure(root->fs_info, start, PAGE_CACHE_SIZE,
2123                                         start, p, mirror_num);
2124                 if (ret)
2125                         break;
2126                 start += PAGE_CACHE_SIZE;
2127         }
2128
2129         return ret;
2130 }
2131
2132 /*
2133  * each time an IO finishes, we do a fast check in the IO failure tree
2134  * to see if we need to process or clean up an io_failure_record
2135  */
2136 static int clean_io_failure(u64 start, struct page *page)
2137 {
2138         u64 private;
2139         u64 private_failure;
2140         struct io_failure_record *failrec;
2141         struct btrfs_fs_info *fs_info;
2142         struct extent_state *state;
2143         int num_copies;
2144         int did_repair = 0;
2145         int ret;
2146         struct inode *inode = page->mapping->host;
2147
2148         private = 0;
2149         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2150                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2151         if (!ret)
2152                 return 0;
2153
2154         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2155                                 &private_failure);
2156         if (ret)
2157                 return 0;
2158
2159         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2160         BUG_ON(!failrec->this_mirror);
2161
2162         if (failrec->in_validation) {
2163                 /* there was no real error, just free the record */
2164                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2165                          failrec->start);
2166                 did_repair = 1;
2167                 goto out;
2168         }
2169
2170         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2171         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2172                                             failrec->start,
2173                                             EXTENT_LOCKED);
2174         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2175
2176         if (state && state->start == failrec->start) {
2177                 fs_info = BTRFS_I(inode)->root->fs_info;
2178                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2179                                               failrec->len);
2180                 if (num_copies > 1)  {
2181                         ret = repair_io_failure(fs_info, start, failrec->len,
2182                                                 failrec->logical, page,
2183                                                 failrec->failed_mirror);
2184                         did_repair = !ret;
2185                 }
2186                 ret = 0;
2187         }
2188
2189 out:
2190         if (!ret)
2191                 ret = free_io_failure(inode, failrec, did_repair);
2192
2193         return ret;
2194 }
2195
2196 /*
2197  * this is a generic handler for readpage errors (default
2198  * readpage_io_failed_hook). if other copies exist, read those and write back
2199  * good data to the failed position. does not investigate in remapping the
2200  * failed extent elsewhere, hoping the device will be smart enough to do this as
2201  * needed
2202  */
2203
2204 static int bio_readpage_error(struct bio *failed_bio, struct page *page,
2205                                 u64 start, u64 end, int failed_mirror,
2206                                 struct extent_state *state)
2207 {
2208         struct io_failure_record *failrec = NULL;
2209         u64 private;
2210         struct extent_map *em;
2211         struct inode *inode = page->mapping->host;
2212         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2213         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2214         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2215         struct bio *bio;
2216         int num_copies;
2217         int ret;
2218         int read_mode;
2219         u64 logical;
2220
2221         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2222
2223         ret = get_state_private(failure_tree, start, &private);
2224         if (ret) {
2225                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2226                 if (!failrec)
2227                         return -ENOMEM;
2228                 failrec->start = start;
2229                 failrec->len = end - start + 1;
2230                 failrec->this_mirror = 0;
2231                 failrec->bio_flags = 0;
2232                 failrec->in_validation = 0;
2233
2234                 read_lock(&em_tree->lock);
2235                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2236                 if (!em) {
2237                         read_unlock(&em_tree->lock);
2238                         kfree(failrec);
2239                         return -EIO;
2240                 }
2241
2242                 if (em->start > start || em->start + em->len < start) {
2243                         free_extent_map(em);
2244                         em = NULL;
2245                 }
2246                 read_unlock(&em_tree->lock);
2247
2248                 if (!em) {
2249                         kfree(failrec);
2250                         return -EIO;
2251                 }
2252                 logical = start - em->start;
2253                 logical = em->block_start + logical;
2254                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2255                         logical = em->block_start;
2256                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2257                         extent_set_compress_type(&failrec->bio_flags,
2258                                                  em->compress_type);
2259                 }
2260                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2261                          "len=%llu\n", logical, start, failrec->len);
2262                 failrec->logical = logical;
2263                 free_extent_map(em);
2264
2265                 /* set the bits in the private failure tree */
2266                 ret = set_extent_bits(failure_tree, start, end,
2267                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2268                 if (ret >= 0)
2269                         ret = set_state_private(failure_tree, start,
2270                                                 (u64)(unsigned long)failrec);
2271                 /* set the bits in the inode's tree */
2272                 if (ret >= 0)
2273                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2274                                                 GFP_NOFS);
2275                 if (ret < 0) {
2276                         kfree(failrec);
2277                         return ret;
2278                 }
2279         } else {
2280                 failrec = (struct io_failure_record *)(unsigned long)private;
2281                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2282                          "start=%llu, len=%llu, validation=%d\n",
2283                          failrec->logical, failrec->start, failrec->len,
2284                          failrec->in_validation);
2285                 /*
2286                  * when data can be on disk more than twice, add to failrec here
2287                  * (e.g. with a list for failed_mirror) to make
2288                  * clean_io_failure() clean all those errors at once.
2289                  */
2290         }
2291         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2292                                       failrec->logical, failrec->len);
2293         if (num_copies == 1) {
2294                 /*
2295                  * we only have a single copy of the data, so don't bother with
2296                  * all the retry and error correction code that follows. no
2297                  * matter what the error is, it is very likely to persist.
2298                  */
2299                 pr_debug("bio_readpage_error: cannot repair, num_copies == 1. "
2300                          "state=%p, num_copies=%d, next_mirror %d, "
2301                          "failed_mirror %d\n", state, num_copies,
2302                          failrec->this_mirror, failed_mirror);
2303                 free_io_failure(inode, failrec, 0);
2304                 return -EIO;
2305         }
2306
2307         if (!state) {
2308                 spin_lock(&tree->lock);
2309                 state = find_first_extent_bit_state(tree, failrec->start,
2310                                                     EXTENT_LOCKED);
2311                 if (state && state->start != failrec->start)
2312                         state = NULL;
2313                 spin_unlock(&tree->lock);
2314         }
2315
2316         /*
2317          * there are two premises:
2318          *      a) deliver good data to the caller
2319          *      b) correct the bad sectors on disk
2320          */
2321         if (failed_bio->bi_vcnt > 1) {
2322                 /*
2323                  * to fulfill b), we need to know the exact failing sectors, as
2324                  * we don't want to rewrite any more than the failed ones. thus,
2325                  * we need separate read requests for the failed bio
2326                  *
2327                  * if the following BUG_ON triggers, our validation request got
2328                  * merged. we need separate requests for our algorithm to work.
2329                  */
2330                 BUG_ON(failrec->in_validation);
2331                 failrec->in_validation = 1;
2332                 failrec->this_mirror = failed_mirror;
2333                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2334         } else {
2335                 /*
2336                  * we're ready to fulfill a) and b) alongside. get a good copy
2337                  * of the failed sector and if we succeed, we have setup
2338                  * everything for repair_io_failure to do the rest for us.
2339                  */
2340                 if (failrec->in_validation) {
2341                         BUG_ON(failrec->this_mirror != failed_mirror);
2342                         failrec->in_validation = 0;
2343                         failrec->this_mirror = 0;
2344                 }
2345                 failrec->failed_mirror = failed_mirror;
2346                 failrec->this_mirror++;
2347                 if (failrec->this_mirror == failed_mirror)
2348                         failrec->this_mirror++;
2349                 read_mode = READ_SYNC;
2350         }
2351
2352         if (!state || failrec->this_mirror > num_copies) {
2353                 pr_debug("bio_readpage_error: (fail) state=%p, num_copies=%d, "
2354                          "next_mirror %d, failed_mirror %d\n", state,
2355                          num_copies, failrec->this_mirror, failed_mirror);
2356                 free_io_failure(inode, failrec, 0);
2357                 return -EIO;
2358         }
2359
2360         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2361         if (!bio) {
2362                 free_io_failure(inode, failrec, 0);
2363                 return -EIO;
2364         }
2365         bio->bi_private = state;
2366         bio->bi_end_io = failed_bio->bi_end_io;
2367         bio->bi_sector = failrec->logical >> 9;
2368         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2369         bio->bi_size = 0;
2370
2371         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2372
2373         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2374                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2375                  failrec->this_mirror, num_copies, failrec->in_validation);
2376
2377         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2378                                          failrec->this_mirror,
2379                                          failrec->bio_flags, 0);
2380         return ret;
2381 }
2382
2383 /* lots and lots of room for performance fixes in the end_bio funcs */
2384
2385 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2386 {
2387         int uptodate = (err == 0);
2388         struct extent_io_tree *tree;
2389         int ret;
2390
2391         tree = &BTRFS_I(page->mapping->host)->io_tree;
2392
2393         if (tree->ops && tree->ops->writepage_end_io_hook) {
2394                 ret = tree->ops->writepage_end_io_hook(page, start,
2395                                                end, NULL, uptodate);
2396                 if (ret)
2397                         uptodate = 0;
2398         }
2399
2400         if (!uptodate) {
2401                 ClearPageUptodate(page);
2402                 SetPageError(page);
2403         }
2404         return 0;
2405 }
2406
2407 /*
2408  * after a writepage IO is done, we need to:
2409  * clear the uptodate bits on error
2410  * clear the writeback bits in the extent tree for this IO
2411  * end_page_writeback if the page has no more pending IO
2412  *
2413  * Scheduling is not allowed, so the extent state tree is expected
2414  * to have one and only one object corresponding to this IO.
2415  */
2416 static void end_bio_extent_writepage(struct bio *bio, int err)
2417 {
2418         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2419         struct extent_io_tree *tree;
2420         u64 start;
2421         u64 end;
2422
2423         do {
2424                 struct page *page = bvec->bv_page;
2425                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2426
2427                 /* We always issue full-page reads, but if some block
2428                  * in a page fails to read, blk_update_request() will
2429                  * advance bv_offset and adjust bv_len to compensate.
2430                  * Print a warning for nonzero offsets, and an error
2431                  * if they don't add up to a full page.  */
2432                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE)
2433                         printk("%s page write in btrfs with offset %u and length %u\n",
2434                                bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE
2435                                ? KERN_ERR "partial" : KERN_INFO "incomplete",
2436                                bvec->bv_offset, bvec->bv_len);
2437
2438                 start = page_offset(page);
2439                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2440
2441                 if (--bvec >= bio->bi_io_vec)
2442                         prefetchw(&bvec->bv_page->flags);
2443
2444                 if (end_extent_writepage(page, err, start, end))
2445                         continue;
2446
2447                 end_page_writeback(page);
2448         } while (bvec >= bio->bi_io_vec);
2449
2450         bio_put(bio);
2451 }
2452
2453 /*
2454  * after a readpage IO is done, we need to:
2455  * clear the uptodate bits on error
2456  * set the uptodate bits if things worked
2457  * set the page up to date if all extents in the tree are uptodate
2458  * clear the lock bit in the extent tree
2459  * unlock the page if there are no other extents locked for it
2460  *
2461  * Scheduling is not allowed, so the extent state tree is expected
2462  * to have one and only one object corresponding to this IO.
2463  */
2464 static void end_bio_extent_readpage(struct bio *bio, int err)
2465 {
2466         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2467         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2468         struct bio_vec *bvec = bio->bi_io_vec;
2469         struct extent_io_tree *tree;
2470         u64 start;
2471         u64 end;
2472         int mirror;
2473         int ret;
2474
2475         if (err)
2476                 uptodate = 0;
2477
2478         do {
2479                 struct page *page = bvec->bv_page;
2480                 struct extent_state *cached = NULL;
2481                 struct extent_state *state;
2482                 struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2483                 struct inode *inode = page->mapping->host;
2484
2485                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2486                          "mirror=%lu\n", (u64)bio->bi_sector, err,
2487                          io_bio->mirror_num);
2488                 tree = &BTRFS_I(inode)->io_tree;
2489
2490                 /* We always issue full-page reads, but if some block
2491                  * in a page fails to read, blk_update_request() will
2492                  * advance bv_offset and adjust bv_len to compensate.
2493                  * Print a warning for nonzero offsets, and an error
2494                  * if they don't add up to a full page.  */
2495                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE)
2496                         printk("%s page read in btrfs with offset %u and length %u\n",
2497                                bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE
2498                                ? KERN_ERR "partial" : KERN_INFO "incomplete",
2499                                bvec->bv_offset, bvec->bv_len);
2500
2501                 start = page_offset(page);
2502                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2503
2504                 if (++bvec <= bvec_end)
2505                         prefetchw(&bvec->bv_page->flags);
2506
2507                 spin_lock(&tree->lock);
2508                 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
2509                 if (state && state->start == start) {
2510                         /*
2511                          * take a reference on the state, unlock will drop
2512                          * the ref
2513                          */
2514                         cache_state(state, &cached);
2515                 }
2516                 spin_unlock(&tree->lock);
2517
2518                 mirror = io_bio->mirror_num;
2519                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
2520                         ret = tree->ops->readpage_end_io_hook(page, start, end,
2521                                                               state, mirror);
2522                         if (ret)
2523                                 uptodate = 0;
2524                         else
2525                                 clean_io_failure(start, page);
2526                 }
2527
2528                 if (!uptodate && tree->ops && tree->ops->readpage_io_failed_hook) {
2529                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2530                         if (!ret && !err &&
2531                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2532                                 uptodate = 1;
2533                 } else if (!uptodate) {
2534                         /*
2535                          * The generic bio_readpage_error handles errors the
2536                          * following way: If possible, new read requests are
2537                          * created and submitted and will end up in
2538                          * end_bio_extent_readpage as well (if we're lucky, not
2539                          * in the !uptodate case). In that case it returns 0 and
2540                          * we just go on with the next page in our bio. If it
2541                          * can't handle the error it will return -EIO and we
2542                          * remain responsible for that page.
2543                          */
2544                         ret = bio_readpage_error(bio, page, start, end, mirror, NULL);
2545                         if (ret == 0) {
2546                                 uptodate =
2547                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2548                                 if (err)
2549                                         uptodate = 0;
2550                                 uncache_state(&cached);
2551                                 continue;
2552                         }
2553                 }
2554
2555                 if (uptodate && tree->track_uptodate) {
2556                         set_extent_uptodate(tree, start, end, &cached,
2557                                             GFP_ATOMIC);
2558                 }
2559                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2560
2561                 if (uptodate) {
2562                         loff_t i_size = i_size_read(inode);
2563                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2564                         unsigned offset;
2565
2566                         /* Zero out the end if this page straddles i_size */
2567                         offset = i_size & (PAGE_CACHE_SIZE-1);
2568                         if (page->index == end_index && offset)
2569                                 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2570                         SetPageUptodate(page);
2571                 } else {
2572                         ClearPageUptodate(page);
2573                         SetPageError(page);
2574                 }
2575                 unlock_page(page);
2576         } while (bvec <= bvec_end);
2577
2578         bio_put(bio);
2579 }
2580
2581 /*
2582  * this allocates from the btrfs_bioset.  We're returning a bio right now
2583  * but you can call btrfs_io_bio for the appropriate container_of magic
2584  */
2585 struct bio *
2586 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2587                 gfp_t gfp_flags)
2588 {
2589         struct bio *bio;
2590
2591         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2592
2593         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2594                 while (!bio && (nr_vecs /= 2)) {
2595                         bio = bio_alloc_bioset(gfp_flags,
2596                                                nr_vecs, btrfs_bioset);
2597                 }
2598         }
2599
2600         if (bio) {
2601                 bio->bi_size = 0;
2602                 bio->bi_bdev = bdev;
2603                 bio->bi_sector = first_sector;
2604         }
2605         return bio;
2606 }
2607
2608 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2609 {
2610         return bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2611 }
2612
2613
2614 /* this also allocates from the btrfs_bioset */
2615 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2616 {
2617         return bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2618 }
2619
2620
2621 static int __must_check submit_one_bio(int rw, struct bio *bio,
2622                                        int mirror_num, unsigned long bio_flags)
2623 {
2624         int ret = 0;
2625         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2626         struct page *page = bvec->bv_page;
2627         struct extent_io_tree *tree = bio->bi_private;
2628         u64 start;
2629
2630         start = page_offset(page) + bvec->bv_offset;
2631
2632         bio->bi_private = NULL;
2633
2634         bio_get(bio);
2635
2636         if (tree->ops && tree->ops->submit_bio_hook)
2637                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2638                                            mirror_num, bio_flags, start);
2639         else
2640                 btrfsic_submit_bio(rw, bio);
2641
2642         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2643                 ret = -EOPNOTSUPP;
2644         bio_put(bio);
2645         return ret;
2646 }
2647
2648 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2649                      unsigned long offset, size_t size, struct bio *bio,
2650                      unsigned long bio_flags)
2651 {
2652         int ret = 0;
2653         if (tree->ops && tree->ops->merge_bio_hook)
2654                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2655                                                 bio_flags);
2656         BUG_ON(ret < 0);
2657         return ret;
2658
2659 }
2660
2661 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2662                               struct page *page, sector_t sector,
2663                               size_t size, unsigned long offset,
2664                               struct block_device *bdev,
2665                               struct bio **bio_ret,
2666                               unsigned long max_pages,
2667                               bio_end_io_t end_io_func,
2668                               int mirror_num,
2669                               unsigned long prev_bio_flags,
2670                               unsigned long bio_flags)
2671 {
2672         int ret = 0;
2673         struct bio *bio;
2674         int nr;
2675         int contig = 0;
2676         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2677         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2678         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2679
2680         if (bio_ret && *bio_ret) {
2681                 bio = *bio_ret;
2682                 if (old_compressed)
2683                         contig = bio->bi_sector == sector;
2684                 else
2685                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
2686                                 sector;
2687
2688                 if (prev_bio_flags != bio_flags || !contig ||
2689                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2690                     bio_add_page(bio, page, page_size, offset) < page_size) {
2691                         ret = submit_one_bio(rw, bio, mirror_num,
2692                                              prev_bio_flags);
2693                         if (ret < 0)
2694                                 return ret;
2695                         bio = NULL;
2696                 } else {
2697                         return 0;
2698                 }
2699         }
2700         if (this_compressed)
2701                 nr = BIO_MAX_PAGES;
2702         else
2703                 nr = bio_get_nr_vecs(bdev);
2704
2705         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2706         if (!bio)
2707                 return -ENOMEM;
2708
2709         bio_add_page(bio, page, page_size, offset);
2710         bio->bi_end_io = end_io_func;
2711         bio->bi_private = tree;
2712
2713         if (bio_ret)
2714                 *bio_ret = bio;
2715         else
2716                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2717
2718         return ret;
2719 }
2720
2721 static void attach_extent_buffer_page(struct extent_buffer *eb,
2722                                       struct page *page)
2723 {
2724         if (!PagePrivate(page)) {
2725                 SetPagePrivate(page);
2726                 page_cache_get(page);
2727                 set_page_private(page, (unsigned long)eb);
2728         } else {
2729                 WARN_ON(page->private != (unsigned long)eb);
2730         }
2731 }
2732
2733 void set_page_extent_mapped(struct page *page)
2734 {
2735         if (!PagePrivate(page)) {
2736                 SetPagePrivate(page);
2737                 page_cache_get(page);
2738                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2739         }
2740 }
2741
2742 /*
2743  * basic readpage implementation.  Locked extent state structs are inserted
2744  * into the tree that are removed when the IO is done (by the end_io
2745  * handlers)
2746  * XXX JDM: This needs looking at to ensure proper page locking
2747  */
2748 static int __extent_read_full_page(struct extent_io_tree *tree,
2749                                    struct page *page,
2750                                    get_extent_t *get_extent,
2751                                    struct bio **bio, int mirror_num,
2752                                    unsigned long *bio_flags, int rw)
2753 {
2754         struct inode *inode = page->mapping->host;
2755         u64 start = page_offset(page);
2756         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2757         u64 end;
2758         u64 cur = start;
2759         u64 extent_offset;
2760         u64 last_byte = i_size_read(inode);
2761         u64 block_start;
2762         u64 cur_end;
2763         sector_t sector;
2764         struct extent_map *em;
2765         struct block_device *bdev;
2766         struct btrfs_ordered_extent *ordered;
2767         int ret;
2768         int nr = 0;
2769         size_t pg_offset = 0;
2770         size_t iosize;
2771         size_t disk_io_size;
2772         size_t blocksize = inode->i_sb->s_blocksize;
2773         unsigned long this_bio_flag = 0;
2774
2775         set_page_extent_mapped(page);
2776
2777         if (!PageUptodate(page)) {
2778                 if (cleancache_get_page(page) == 0) {
2779                         BUG_ON(blocksize != PAGE_SIZE);
2780                         goto out;
2781                 }
2782         }
2783
2784         end = page_end;
2785         while (1) {
2786                 lock_extent(tree, start, end);
2787                 ordered = btrfs_lookup_ordered_extent(inode, start);
2788                 if (!ordered)
2789                         break;
2790                 unlock_extent(tree, start, end);
2791                 btrfs_start_ordered_extent(inode, ordered, 1);
2792                 btrfs_put_ordered_extent(ordered);
2793         }
2794
2795         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2796                 char *userpage;
2797                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2798
2799                 if (zero_offset) {
2800                         iosize = PAGE_CACHE_SIZE - zero_offset;
2801                         userpage = kmap_atomic(page);
2802                         memset(userpage + zero_offset, 0, iosize);
2803                         flush_dcache_page(page);
2804                         kunmap_atomic(userpage);
2805                 }
2806         }
2807         while (cur <= end) {
2808                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2809
2810                 if (cur >= last_byte) {
2811                         char *userpage;
2812                         struct extent_state *cached = NULL;
2813
2814                         iosize = PAGE_CACHE_SIZE - pg_offset;
2815                         userpage = kmap_atomic(page);
2816                         memset(userpage + pg_offset, 0, iosize);
2817                         flush_dcache_page(page);
2818                         kunmap_atomic(userpage);
2819                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2820                                             &cached, GFP_NOFS);
2821                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2822                                              &cached, GFP_NOFS);
2823                         break;
2824                 }
2825                 em = get_extent(inode, page, pg_offset, cur,
2826                                 end - cur + 1, 0);
2827                 if (IS_ERR_OR_NULL(em)) {
2828                         SetPageError(page);
2829                         unlock_extent(tree, cur, end);
2830                         break;
2831                 }
2832                 extent_offset = cur - em->start;
2833                 BUG_ON(extent_map_end(em) <= cur);
2834                 BUG_ON(end < cur);
2835
2836                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2837                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2838                         extent_set_compress_type(&this_bio_flag,
2839                                                  em->compress_type);
2840                 }
2841
2842                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2843                 cur_end = min(extent_map_end(em) - 1, end);
2844                 iosize = ALIGN(iosize, blocksize);
2845                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2846                         disk_io_size = em->block_len;
2847                         sector = em->block_start >> 9;
2848                 } else {
2849                         sector = (em->block_start + extent_offset) >> 9;
2850                         disk_io_size = iosize;
2851                 }
2852                 bdev = em->bdev;
2853                 block_start = em->block_start;
2854                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2855                         block_start = EXTENT_MAP_HOLE;
2856                 free_extent_map(em);
2857                 em = NULL;
2858
2859                 /* we've found a hole, just zero and go on */
2860                 if (block_start == EXTENT_MAP_HOLE) {
2861                         char *userpage;
2862                         struct extent_state *cached = NULL;
2863
2864                         userpage = kmap_atomic(page);
2865                         memset(userpage + pg_offset, 0, iosize);
2866                         flush_dcache_page(page);
2867                         kunmap_atomic(userpage);
2868
2869                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2870                                             &cached, GFP_NOFS);
2871                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2872                                              &cached, GFP_NOFS);
2873                         cur = cur + iosize;
2874                         pg_offset += iosize;
2875                         continue;
2876                 }
2877                 /* the get_extent function already copied into the page */
2878                 if (test_range_bit(tree, cur, cur_end,
2879                                    EXTENT_UPTODATE, 1, NULL)) {
2880                         check_page_uptodate(tree, page);
2881                         unlock_extent(tree, cur, cur + iosize - 1);
2882                         cur = cur + iosize;
2883                         pg_offset += iosize;
2884                         continue;
2885                 }
2886                 /* we have an inline extent but it didn't get marked up
2887                  * to date.  Error out
2888                  */
2889                 if (block_start == EXTENT_MAP_INLINE) {
2890                         SetPageError(page);
2891                         unlock_extent(tree, cur, cur + iosize - 1);
2892                         cur = cur + iosize;
2893                         pg_offset += iosize;
2894                         continue;
2895                 }
2896
2897                 pnr -= page->index;
2898                 ret = submit_extent_page(rw, tree, page,
2899                                          sector, disk_io_size, pg_offset,
2900                                          bdev, bio, pnr,
2901                                          end_bio_extent_readpage, mirror_num,
2902                                          *bio_flags,
2903                                          this_bio_flag);
2904                 if (!ret) {
2905                         nr++;
2906                         *bio_flags = this_bio_flag;
2907                 } else {
2908                         SetPageError(page);
2909                         unlock_extent(tree, cur, cur + iosize - 1);
2910                 }
2911                 cur = cur + iosize;
2912                 pg_offset += iosize;
2913         }
2914 out:
2915         if (!nr) {
2916                 if (!PageError(page))
2917                         SetPageUptodate(page);
2918                 unlock_page(page);
2919         }
2920         return 0;
2921 }
2922
2923 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2924                             get_extent_t *get_extent, int mirror_num)
2925 {
2926         struct bio *bio = NULL;
2927         unsigned long bio_flags = 0;
2928         int ret;
2929
2930         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
2931                                       &bio_flags, READ);
2932         if (bio)
2933                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
2934         return ret;
2935 }
2936
2937 static noinline void update_nr_written(struct page *page,
2938                                       struct writeback_control *wbc,
2939                                       unsigned long nr_written)
2940 {
2941         wbc->nr_to_write -= nr_written;
2942         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2943             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2944                 page->mapping->writeback_index = page->index + nr_written;
2945 }
2946
2947 /*
2948  * the writepage semantics are similar to regular writepage.  extent
2949  * records are inserted to lock ranges in the tree, and as dirty areas
2950  * are found, they are marked writeback.  Then the lock bits are removed
2951  * and the end_io handler clears the writeback ranges
2952  */
2953 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2954                               void *data)
2955 {
2956         struct inode *inode = page->mapping->host;
2957         struct extent_page_data *epd = data;
2958         struct extent_io_tree *tree = epd->tree;
2959         u64 start = page_offset(page);
2960         u64 delalloc_start;
2961         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2962         u64 end;
2963         u64 cur = start;
2964         u64 extent_offset;
2965         u64 last_byte = i_size_read(inode);
2966         u64 block_start;
2967         u64 iosize;
2968         sector_t sector;
2969         struct extent_state *cached_state = NULL;
2970         struct extent_map *em;
2971         struct block_device *bdev;
2972         int ret;
2973         int nr = 0;
2974         size_t pg_offset = 0;
2975         size_t blocksize;
2976         loff_t i_size = i_size_read(inode);
2977         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2978         u64 nr_delalloc;
2979         u64 delalloc_end;
2980         int page_started;
2981         int compressed;
2982         int write_flags;
2983         unsigned long nr_written = 0;
2984         bool fill_delalloc = true;
2985
2986         if (wbc->sync_mode == WB_SYNC_ALL)
2987                 write_flags = WRITE_SYNC;
2988         else
2989                 write_flags = WRITE;
2990
2991         trace___extent_writepage(page, inode, wbc);
2992
2993         WARN_ON(!PageLocked(page));
2994
2995         ClearPageError(page);
2996
2997         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2998         if (page->index > end_index ||
2999            (page->index == end_index && !pg_offset)) {
3000                 page->mapping->a_ops->invalidatepage(page, 0);
3001                 unlock_page(page);
3002                 return 0;
3003         }
3004
3005         if (page->index == end_index) {
3006                 char *userpage;
3007
3008                 userpage = kmap_atomic(page);
3009                 memset(userpage + pg_offset, 0,
3010                        PAGE_CACHE_SIZE - pg_offset);
3011                 kunmap_atomic(userpage);
3012                 flush_dcache_page(page);
3013         }
3014         pg_offset = 0;
3015
3016         set_page_extent_mapped(page);
3017
3018         if (!tree->ops || !tree->ops->fill_delalloc)
3019                 fill_delalloc = false;
3020
3021         delalloc_start = start;
3022         delalloc_end = 0;
3023         page_started = 0;
3024         if (!epd->extent_locked && fill_delalloc) {
3025                 u64 delalloc_to_write = 0;
3026                 /*
3027                  * make sure the wbc mapping index is at least updated
3028                  * to this page.
3029                  */
3030                 update_nr_written(page, wbc, 0);
3031
3032                 while (delalloc_end < page_end) {
3033                         nr_delalloc = find_lock_delalloc_range(inode, tree,
3034                                                        page,
3035                                                        &delalloc_start,
3036                                                        &delalloc_end,
3037                                                        128 * 1024 * 1024);
3038                         if (nr_delalloc == 0) {
3039                                 delalloc_start = delalloc_end + 1;
3040                                 continue;
3041                         }
3042                         ret = tree->ops->fill_delalloc(inode, page,
3043                                                        delalloc_start,
3044                                                        delalloc_end,
3045                                                        &page_started,
3046                                                        &nr_written);
3047                         /* File system has been set read-only */
3048                         if (ret) {
3049                                 SetPageError(page);
3050                                 goto done;
3051                         }
3052                         /*
3053                          * delalloc_end is already one less than the total
3054                          * length, so we don't subtract one from
3055                          * PAGE_CACHE_SIZE
3056                          */
3057                         delalloc_to_write += (delalloc_end - delalloc_start +
3058                                               PAGE_CACHE_SIZE) >>
3059                                               PAGE_CACHE_SHIFT;
3060                         delalloc_start = delalloc_end + 1;
3061                 }
3062                 if (wbc->nr_to_write < delalloc_to_write) {
3063                         int thresh = 8192;
3064
3065                         if (delalloc_to_write < thresh * 2)
3066                                 thresh = delalloc_to_write;
3067                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
3068                                                  thresh);
3069                 }
3070
3071                 /* did the fill delalloc function already unlock and start
3072                  * the IO?
3073                  */
3074                 if (page_started) {
3075                         ret = 0;
3076                         /*
3077                          * we've unlocked the page, so we can't update
3078                          * the mapping's writeback index, just update
3079                          * nr_to_write.
3080                          */
3081                         wbc->nr_to_write -= nr_written;
3082                         goto done_unlocked;
3083                 }
3084         }
3085         if (tree->ops && tree->ops->writepage_start_hook) {
3086                 ret = tree->ops->writepage_start_hook(page, start,
3087                                                       page_end);
3088                 if (ret) {
3089                         /* Fixup worker will requeue */
3090                         if (ret == -EBUSY)
3091                                 wbc->pages_skipped++;
3092                         else
3093                                 redirty_page_for_writepage(wbc, page);
3094                         update_nr_written(page, wbc, nr_written);
3095                         unlock_page(page);
3096                         ret = 0;
3097                         goto done_unlocked;
3098                 }
3099         }
3100
3101         /*
3102          * we don't want to touch the inode after unlocking the page,
3103          * so we update the mapping writeback index now
3104          */
3105         update_nr_written(page, wbc, nr_written + 1);
3106
3107         end = page_end;
3108         if (last_byte <= start) {
3109                 if (tree->ops && tree->ops->writepage_end_io_hook)
3110                         tree->ops->writepage_end_io_hook(page, start,
3111                                                          page_end, NULL, 1);
3112                 goto done;
3113         }
3114
3115         blocksize = inode->i_sb->s_blocksize;
3116
3117         while (cur <= end) {
3118                 if (cur >= last_byte) {
3119                         if (tree->ops && tree->ops->writepage_end_io_hook)
3120                                 tree->ops->writepage_end_io_hook(page, cur,
3121                                                          page_end, NULL, 1);
3122                         break;
3123                 }
3124                 em = epd->get_extent(inode, page, pg_offset, cur,
3125                                      end - cur + 1, 1);
3126                 if (IS_ERR_OR_NULL(em)) {
3127                         SetPageError(page);
3128                         break;
3129                 }
3130
3131                 extent_offset = cur - em->start;
3132                 BUG_ON(extent_map_end(em) <= cur);
3133                 BUG_ON(end < cur);
3134                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3135                 iosize = ALIGN(iosize, blocksize);
3136                 sector = (em->block_start + extent_offset) >> 9;
3137                 bdev = em->bdev;
3138                 block_start = em->block_start;
3139                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3140                 free_extent_map(em);
3141                 em = NULL;
3142
3143                 /*
3144                  * compressed and inline extents are written through other
3145                  * paths in the FS
3146                  */
3147                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3148                     block_start == EXTENT_MAP_INLINE) {
3149                         /*
3150                          * end_io notification does not happen here for
3151                          * compressed extents
3152                          */
3153                         if (!compressed && tree->ops &&
3154                             tree->ops->writepage_end_io_hook)
3155                                 tree->ops->writepage_end_io_hook(page, cur,
3156                                                          cur + iosize - 1,
3157                                                          NULL, 1);
3158                         else if (compressed) {
3159                                 /* we don't want to end_page_writeback on
3160                                  * a compressed extent.  this happens
3161                                  * elsewhere
3162                                  */
3163                                 nr++;
3164                         }
3165
3166                         cur += iosize;
3167                         pg_offset += iosize;
3168                         continue;
3169                 }
3170                 /* leave this out until we have a page_mkwrite call */
3171                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
3172                                    EXTENT_DIRTY, 0, NULL)) {
3173                         cur = cur + iosize;
3174                         pg_offset += iosize;
3175                         continue;
3176                 }
3177
3178                 if (tree->ops && tree->ops->writepage_io_hook) {
3179                         ret = tree->ops->writepage_io_hook(page, cur,
3180                                                 cur + iosize - 1);
3181                 } else {
3182                         ret = 0;
3183                 }
3184                 if (ret) {
3185                         SetPageError(page);
3186                 } else {
3187                         unsigned long max_nr = end_index + 1;
3188
3189                         set_range_writeback(tree, cur, cur + iosize - 1);
3190                         if (!PageWriteback(page)) {
3191                                 printk(KERN_ERR "btrfs warning page %lu not "
3192                                        "writeback, cur %llu end %llu\n",
3193                                        page->index, (unsigned long long)cur,
3194                                        (unsigned long long)end);
3195                         }
3196
3197                         ret = submit_extent_page(write_flags, tree, page,
3198                                                  sector, iosize, pg_offset,
3199                                                  bdev, &epd->bio, max_nr,
3200                                                  end_bio_extent_writepage,
3201                                                  0, 0, 0);
3202                         if (ret)
3203                                 SetPageError(page);
3204                 }
3205                 cur = cur + iosize;
3206                 pg_offset += iosize;
3207                 nr++;
3208         }
3209 done:
3210         if (nr == 0) {
3211                 /* make sure the mapping tag for page dirty gets cleared */
3212                 set_page_writeback(page);
3213                 end_page_writeback(page);
3214         }
3215         unlock_page(page);
3216
3217 done_unlocked:
3218
3219         /* drop our reference on any cached states */
3220         free_extent_state(cached_state);
3221         return 0;
3222 }
3223
3224 static int eb_wait(void *word)
3225 {
3226         io_schedule();
3227         return 0;
3228 }
3229
3230 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3231 {
3232         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3233                     TASK_UNINTERRUPTIBLE);
3234 }
3235
3236 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3237                                      struct btrfs_fs_info *fs_info,
3238                                      struct extent_page_data *epd)
3239 {
3240         unsigned long i, num_pages;
3241         int flush = 0;
3242         int ret = 0;
3243
3244         if (!btrfs_try_tree_write_lock(eb)) {
3245                 flush = 1;
3246                 flush_write_bio(epd);
3247                 btrfs_tree_lock(eb);
3248         }
3249
3250         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3251                 btrfs_tree_unlock(eb);
3252                 if (!epd->sync_io)
3253                         return 0;
3254                 if (!flush) {
3255                         flush_write_bio(epd);
3256                         flush = 1;
3257                 }
3258                 while (1) {
3259                         wait_on_extent_buffer_writeback(eb);
3260                         btrfs_tree_lock(eb);
3261                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3262                                 break;
3263                         btrfs_tree_unlock(eb);
3264                 }
3265         }
3266
3267         /*
3268          * We need to do this to prevent races in people who check if the eb is
3269          * under IO since we can end up having no IO bits set for a short period
3270          * of time.
3271          */
3272         spin_lock(&eb->refs_lock);
3273         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3274                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3275                 spin_unlock(&eb->refs_lock);
3276                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3277                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3278                                      -eb->len,
3279                                      fs_info->dirty_metadata_batch);
3280                 ret = 1;
3281         } else {
3282                 spin_unlock(&eb->refs_lock);
3283         }
3284
3285         btrfs_tree_unlock(eb);
3286
3287         if (!ret)
3288                 return ret;
3289
3290         num_pages = num_extent_pages(eb->start, eb->len);
3291         for (i = 0; i < num_pages; i++) {
3292                 struct page *p = extent_buffer_page(eb, i);
3293
3294                 if (!trylock_page(p)) {
3295                         if (!flush) {
3296                                 flush_write_bio(epd);
3297                                 flush = 1;
3298                         }
3299                         lock_page(p);
3300                 }
3301         }
3302
3303         return ret;
3304 }
3305
3306 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3307 {
3308         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3309         smp_mb__after_clear_bit();
3310         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3311 }
3312
3313 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3314 {
3315         int uptodate = err == 0;
3316         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
3317         struct extent_buffer *eb;
3318         int done;
3319
3320         do {
3321                 struct page *page = bvec->bv_page;
3322
3323                 bvec--;
3324                 eb = (struct extent_buffer *)page->private;
3325                 BUG_ON(!eb);
3326                 done = atomic_dec_and_test(&eb->io_pages);
3327
3328                 if (!uptodate || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3329                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3330                         ClearPageUptodate(page);
3331                         SetPageError(page);
3332                 }
3333
3334                 end_page_writeback(page);
3335
3336                 if (!done)
3337                         continue;
3338
3339                 end_extent_buffer_writeback(eb);
3340         } while (bvec >= bio->bi_io_vec);
3341
3342         bio_put(bio);
3343
3344 }
3345
3346 static int write_one_eb(struct extent_buffer *eb,
3347                         struct btrfs_fs_info *fs_info,
3348                         struct writeback_control *wbc,
3349                         struct extent_page_data *epd)
3350 {
3351         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3352         u64 offset = eb->start;
3353         unsigned long i, num_pages;
3354         unsigned long bio_flags = 0;
3355         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3356         int ret = 0;
3357
3358         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3359         num_pages = num_extent_pages(eb->start, eb->len);
3360         atomic_set(&eb->io_pages, num_pages);
3361         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3362                 bio_flags = EXTENT_BIO_TREE_LOG;
3363
3364         for (i = 0; i < num_pages; i++) {
3365                 struct page *p = extent_buffer_page(eb, i);
3366
3367                 clear_page_dirty_for_io(p);
3368                 set_page_writeback(p);
3369                 ret = submit_extent_page(rw, eb->tree, p, offset >> 9,
3370                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3371                                          -1, end_bio_extent_buffer_writepage,
3372                                          0, epd->bio_flags, bio_flags);
3373                 epd->bio_flags = bio_flags;
3374                 if (ret) {
3375                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3376                         SetPageError(p);
3377                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3378                                 end_extent_buffer_writeback(eb);
3379                         ret = -EIO;
3380                         break;
3381                 }
3382                 offset += PAGE_CACHE_SIZE;
3383                 update_nr_written(p, wbc, 1);
3384                 unlock_page(p);
3385         }
3386
3387         if (unlikely(ret)) {
3388                 for (; i < num_pages; i++) {
3389                         struct page *p = extent_buffer_page(eb, i);
3390                         unlock_page(p);
3391                 }
3392         }
3393
3394         return ret;
3395 }
3396
3397 int btree_write_cache_pages(struct address_space *mapping,
3398                                    struct writeback_control *wbc)
3399 {
3400         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3401         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3402         struct extent_buffer *eb, *prev_eb = NULL;
3403         struct extent_page_data epd = {
3404                 .bio = NULL,
3405                 .tree = tree,
3406                 .extent_locked = 0,
3407                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3408                 .bio_flags = 0,
3409         };
3410         int ret = 0;
3411         int done = 0;
3412         int nr_to_write_done = 0;
3413         struct pagevec pvec;
3414         int nr_pages;
3415         pgoff_t index;
3416         pgoff_t end;            /* Inclusive */
3417         int scanned = 0;
3418         int tag;
3419
3420         pagevec_init(&pvec, 0);
3421         if (wbc->range_cyclic) {
3422                 index = mapping->writeback_index; /* Start from prev offset */
3423                 end = -1;
3424         } else {
3425                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3426                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3427                 scanned = 1;
3428         }
3429         if (wbc->sync_mode == WB_SYNC_ALL)
3430                 tag = PAGECACHE_TAG_TOWRITE;
3431         else
3432                 tag = PAGECACHE_TAG_DIRTY;
3433 retry:
3434         if (wbc->sync_mode == WB_SYNC_ALL)
3435                 tag_pages_for_writeback(mapping, index, end);
3436         while (!done && !nr_to_write_done && (index <= end) &&
3437                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3438                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3439                 unsigned i;
3440
3441                 scanned = 1;
3442                 for (i = 0; i < nr_pages; i++) {
3443                         struct page *page = pvec.pages[i];
3444
3445                         if (!PagePrivate(page))
3446                                 continue;
3447
3448                         if (!wbc->range_cyclic && page->index > end) {
3449                                 done = 1;
3450                                 break;
3451                         }
3452
3453                         spin_lock(&mapping->private_lock);
3454                         if (!PagePrivate(page)) {
3455                                 spin_unlock(&mapping->private_lock);
3456                                 continue;
3457                         }
3458
3459                         eb = (struct extent_buffer *)page->private;
3460
3461                         /*
3462                          * Shouldn't happen and normally this would be a BUG_ON
3463                          * but no sense in crashing the users box for something
3464                          * we can survive anyway.
3465                          */
3466                         if (!eb) {
3467                                 spin_unlock(&mapping->private_lock);
3468                                 WARN_ON(1);
3469                                 continue;
3470                         }
3471
3472                         if (eb == prev_eb) {
3473                                 spin_unlock(&mapping->private_lock);
3474                                 continue;
3475                         }
3476
3477                         ret = atomic_inc_not_zero(&eb->refs);
3478                         spin_unlock(&mapping->private_lock);
3479                         if (!ret)
3480                                 continue;
3481
3482                         prev_eb = eb;
3483                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3484                         if (!ret) {
3485                                 free_extent_buffer(eb);
3486                                 continue;
3487                         }
3488
3489                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3490                         if (ret) {
3491                                 done = 1;
3492                                 free_extent_buffer(eb);
3493                                 break;
3494                         }
3495                         free_extent_buffer(eb);
3496
3497                         /*
3498                          * the filesystem may choose to bump up nr_to_write.
3499                          * We have to make sure to honor the new nr_to_write
3500                          * at any time
3501                          */
3502                         nr_to_write_done = wbc->nr_to_write <= 0;
3503                 }
3504                 pagevec_release(&pvec);
3505                 cond_resched();
3506         }
3507         if (!scanned && !done) {
3508                 /*
3509                  * We hit the last page and there is more work to be done: wrap
3510                  * back to the start of the file
3511                  */
3512                 scanned = 1;
3513                 index = 0;
3514                 goto retry;
3515         }
3516         flush_write_bio(&epd);
3517         return ret;
3518 }
3519
3520 /**
3521  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3522  * @mapping: address space structure to write
3523  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3524  * @writepage: function called for each page
3525  * @data: data passed to writepage function
3526  *
3527  * If a page is already under I/O, write_cache_pages() skips it, even
3528  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3529  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3530  * and msync() need to guarantee that all the data which was dirty at the time
3531  * the call was made get new I/O started against them.  If wbc->sync_mode is
3532  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3533  * existing IO to complete.
3534  */
3535 static int extent_write_cache_pages(struct extent_io_tree *tree,
3536                              struct address_space *mapping,
3537                              struct writeback_control *wbc,
3538                              writepage_t writepage, void *data,
3539                              void (*flush_fn)(void *))
3540 {
3541         struct inode *inode = mapping->host;
3542         int ret = 0;
3543         int done = 0;
3544         int nr_to_write_done = 0;
3545         struct pagevec pvec;
3546         int nr_pages;
3547         pgoff_t index;
3548         pgoff_t end;            /* Inclusive */
3549         int scanned = 0;
3550         int tag;
3551
3552         /*
3553          * We have to hold onto the inode so that ordered extents can do their
3554          * work when the IO finishes.  The alternative to this is failing to add
3555          * an ordered extent if the igrab() fails there and that is a huge pain
3556          * to deal with, so instead just hold onto the inode throughout the
3557          * writepages operation.  If it fails here we are freeing up the inode
3558          * anyway and we'd rather not waste our time writing out stuff that is
3559          * going to be truncated anyway.
3560          */
3561         if (!igrab(inode))
3562                 return 0;
3563
3564         pagevec_init(&pvec, 0);
3565         if (wbc->range_cyclic) {
3566                 index = mapping->writeback_index; /* Start from prev offset */
3567                 end = -1;
3568         } else {
3569                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3570                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3571                 scanned = 1;
3572         }
3573         if (wbc->sync_mode == WB_SYNC_ALL)
3574                 tag = PAGECACHE_TAG_TOWRITE;
3575         else
3576                 tag = PAGECACHE_TAG_DIRTY;
3577 retry:
3578         if (wbc->sync_mode == WB_SYNC_ALL)
3579                 tag_pages_for_writeback(mapping, index, end);
3580         while (!done && !nr_to_write_done && (index <= end) &&
3581                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3582                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3583                 unsigned i;
3584
3585                 scanned = 1;
3586                 for (i = 0; i < nr_pages; i++) {
3587                         struct page *page = pvec.pages[i];
3588
3589                         /*
3590                          * At this point we hold neither mapping->tree_lock nor
3591                          * lock on the page itself: the page may be truncated or
3592                          * invalidated (changing page->mapping to NULL), or even
3593                          * swizzled back from swapper_space to tmpfs file
3594                          * mapping
3595                          */
3596                         if (!trylock_page(page)) {
3597                                 flush_fn(data);
3598                                 lock_page(page);
3599                         }
3600
3601                         if (unlikely(page->mapping != mapping)) {
3602                                 unlock_page(page);
3603                                 continue;
3604                         }
3605
3606                         if (!wbc->range_cyclic && page->index > end) {
3607                                 done = 1;
3608                                 unlock_page(page);
3609                                 continue;
3610                         }
3611
3612                         if (wbc->sync_mode != WB_SYNC_NONE) {
3613                                 if (PageWriteback(page))
3614                                         flush_fn(data);
3615                                 wait_on_page_writeback(page);
3616                         }
3617
3618                         if (PageWriteback(page) ||
3619                             !clear_page_dirty_for_io(page)) {
3620                                 unlock_page(page);
3621                                 continue;
3622                         }
3623
3624                         ret = (*writepage)(page, wbc, data);
3625
3626                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3627                                 unlock_page(page);
3628                                 ret = 0;
3629                         }
3630                         if (ret)
3631                                 done = 1;
3632
3633                         /*
3634                          * the filesystem may choose to bump up nr_to_write.
3635                          * We have to make sure to honor the new nr_to_write
3636                          * at any time
3637                          */
3638                         nr_to_write_done = wbc->nr_to_write <= 0;
3639                 }
3640                 pagevec_release(&pvec);
3641                 cond_resched();
3642         }
3643         if (!scanned && !done) {
3644                 /*
3645                  * We hit the last page and there is more work to be done: wrap
3646                  * back to the start of the file
3647                  */
3648                 scanned = 1;
3649                 index = 0;
3650                 goto retry;
3651         }
3652         btrfs_add_delayed_iput(inode);
3653         return ret;
3654 }
3655
3656 static void flush_epd_write_bio(struct extent_page_data *epd)
3657 {
3658         if (epd->bio) {
3659                 int rw = WRITE;
3660                 int ret;
3661
3662                 if (epd->sync_io)
3663                         rw = WRITE_SYNC;
3664
3665                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3666                 BUG_ON(ret < 0); /* -ENOMEM */
3667                 epd->bio = NULL;
3668         }
3669 }
3670
3671 static noinline void flush_write_bio(void *data)
3672 {
3673         struct extent_page_data *epd = data;
3674         flush_epd_write_bio(epd);
3675 }
3676
3677 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3678                           get_extent_t *get_extent,
3679                           struct writeback_control *wbc)
3680 {
3681         int ret;
3682         struct extent_page_data epd = {
3683                 .bio = NULL,
3684                 .tree = tree,
3685                 .get_extent = get_extent,
3686                 .extent_locked = 0,
3687                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3688                 .bio_flags = 0,
3689         };
3690
3691         ret = __extent_writepage(page, wbc, &epd);
3692
3693         flush_epd_write_bio(&epd);
3694         return ret;
3695 }
3696
3697 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3698                               u64 start, u64 end, get_extent_t *get_extent,
3699                               int mode)
3700 {
3701         int ret = 0;
3702         struct address_space *mapping = inode->i_mapping;
3703         struct page *page;
3704         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3705                 PAGE_CACHE_SHIFT;
3706
3707         struct extent_page_data epd = {
3708                 .bio = NULL,
3709                 .tree = tree,
3710                 .get_extent = get_extent,
3711                 .extent_locked = 1,
3712                 .sync_io = mode == WB_SYNC_ALL,
3713                 .bio_flags = 0,
3714         };
3715         struct writeback_control wbc_writepages = {
3716                 .sync_mode      = mode,
3717                 .nr_to_write    = nr_pages * 2,
3718                 .range_start    = start,
3719                 .range_end      = end + 1,
3720         };
3721
3722         while (start <= end) {
3723                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3724                 if (clear_page_dirty_for_io(page))
3725                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3726                 else {
3727                         if (tree->ops && tree->ops->writepage_end_io_hook)
3728                                 tree->ops->writepage_end_io_hook(page, start,
3729                                                  start + PAGE_CACHE_SIZE - 1,
3730                                                  NULL, 1);
3731                         unlock_page(page);
3732                 }
3733                 page_cache_release(page);
3734                 start += PAGE_CACHE_SIZE;
3735         }
3736
3737         flush_epd_write_bio(&epd);
3738         return ret;
3739 }
3740
3741 int extent_writepages(struct extent_io_tree *tree,
3742                       struct address_space *mapping,
3743                       get_extent_t *get_extent,
3744                       struct writeback_control *wbc)
3745 {
3746         int ret = 0;
3747         struct extent_page_data epd = {
3748                 .bio = NULL,
3749                 .tree = tree,
3750                 .get_extent = get_extent,
3751                 .extent_locked = 0,
3752                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3753                 .bio_flags = 0,
3754         };
3755
3756         ret = extent_write_cache_pages(tree, mapping, wbc,
3757                                        __extent_writepage, &epd,
3758                                        flush_write_bio);
3759         flush_epd_write_bio(&epd);
3760         return ret;
3761 }
3762
3763 int extent_readpages(struct extent_io_tree *tree,
3764                      struct address_space *mapping,
3765                      struct list_head *pages, unsigned nr_pages,
3766                      get_extent_t get_extent)
3767 {
3768         struct bio *bio = NULL;
3769         unsigned page_idx;
3770         unsigned long bio_flags = 0;
3771         struct page *pagepool[16];
3772         struct page *page;
3773         int i = 0;
3774         int nr = 0;
3775
3776         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3777                 page = list_entry(pages->prev, struct page, lru);
3778
3779                 prefetchw(&page->flags);
3780                 list_del(&page->lru);
3781                 if (add_to_page_cache_lru(page, mapping,
3782                                         page->index, GFP_NOFS)) {
3783                         page_cache_release(page);
3784                         continue;
3785                 }
3786
3787                 pagepool[nr++] = page;
3788                 if (nr < ARRAY_SIZE(pagepool))
3789                         continue;
3790                 for (i = 0; i < nr; i++) {
3791                         __extent_read_full_page(tree, pagepool[i], get_extent,
3792                                         &bio, 0, &bio_flags, READ);
3793                         page_cache_release(pagepool[i]);
3794                 }
3795                 nr = 0;
3796         }
3797         for (i = 0; i < nr; i++) {
3798                 __extent_read_full_page(tree, pagepool[i], get_extent,
3799                                         &bio, 0, &bio_flags, READ);
3800                 page_cache_release(pagepool[i]);
3801         }
3802
3803         BUG_ON(!list_empty(pages));
3804         if (bio)
3805                 return submit_one_bio(READ, bio, 0, bio_flags);
3806         return 0;
3807 }
3808
3809 /*
3810  * basic invalidatepage code, this waits on any locked or writeback
3811  * ranges corresponding to the page, and then deletes any extent state
3812  * records from the tree
3813  */
3814 int extent_invalidatepage(struct extent_io_tree *tree,
3815                           struct page *page, unsigned long offset)
3816 {
3817         struct extent_state *cached_state = NULL;
3818         u64 start = page_offset(page);
3819         u64 end = start + PAGE_CACHE_SIZE - 1;
3820         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3821
3822         start += ALIGN(offset, blocksize);
3823         if (start > end)
3824                 return 0;
3825
3826         lock_extent_bits(tree, start, end, 0, &cached_state);
3827         wait_on_page_writeback(page);
3828         clear_extent_bit(tree, start, end,
3829                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3830                          EXTENT_DO_ACCOUNTING,
3831                          1, 1, &cached_state, GFP_NOFS);
3832         return 0;
3833 }
3834
3835 /*
3836  * a helper for releasepage, this tests for areas of the page that
3837  * are locked or under IO and drops the related state bits if it is safe
3838  * to drop the page.
3839  */
3840 static int try_release_extent_state(struct extent_map_tree *map,
3841                                     struct extent_io_tree *tree,
3842                                     struct page *page, gfp_t mask)
3843 {
3844         u64 start = page_offset(page);
3845         u64 end = start + PAGE_CACHE_SIZE - 1;
3846         int ret = 1;
3847
3848         if (test_range_bit(tree, start, end,
3849                            EXTENT_IOBITS, 0, NULL))
3850                 ret = 0;
3851         else {
3852                 if ((mask & GFP_NOFS) == GFP_NOFS)
3853                         mask = GFP_NOFS;
3854                 /*
3855                  * at this point we can safely clear everything except the
3856                  * locked bit and the nodatasum bit
3857                  */
3858                 ret = clear_extent_bit(tree, start, end,
3859                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3860                                  0, 0, NULL, mask);
3861
3862                 /* if clear_extent_bit failed for enomem reasons,
3863                  * we can't allow the release to continue.
3864                  */
3865                 if (ret < 0)
3866                         ret = 0;
3867                 else
3868                         ret = 1;
3869         }
3870         return ret;
3871 }
3872
3873 /*
3874  * a helper for releasepage.  As long as there are no locked extents
3875  * in the range corresponding to the page, both state records and extent
3876  * map records are removed
3877  */
3878 int try_release_extent_mapping(struct extent_map_tree *map,
3879                                struct extent_io_tree *tree, struct page *page,
3880                                gfp_t mask)
3881 {
3882         struct extent_map *em;
3883         u64 start = page_offset(page);
3884         u64 end = start + PAGE_CACHE_SIZE - 1;
3885
3886         if ((mask & __GFP_WAIT) &&
3887             page->mapping->host->i_size > 16 * 1024 * 1024) {
3888                 u64 len;
3889                 while (start <= end) {
3890                         len = end - start + 1;
3891                         write_lock(&map->lock);
3892                         em = lookup_extent_mapping(map, start, len);
3893                         if (!em) {
3894                                 write_unlock(&map->lock);
3895                                 break;
3896                         }
3897                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3898                             em->start != start) {
3899                                 write_unlock(&map->lock);
3900                                 free_extent_map(em);
3901                                 break;
3902                         }
3903                         if (!test_range_bit(tree, em->start,
3904                                             extent_map_end(em) - 1,
3905                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
3906                                             0, NULL)) {
3907                                 remove_extent_mapping(map, em);
3908                                 /* once for the rb tree */
3909                                 free_extent_map(em);
3910                         }
3911                         start = extent_map_end(em);
3912                         write_unlock(&map->lock);
3913
3914                         /* once for us */
3915                         free_extent_map(em);
3916                 }
3917         }
3918         return try_release_extent_state(map, tree, page, mask);
3919 }
3920
3921 /*
3922  * helper function for fiemap, which doesn't want to see any holes.
3923  * This maps until we find something past 'last'
3924  */
3925 static struct extent_map *get_extent_skip_holes(struct inode *inode,
3926                                                 u64 offset,
3927                                                 u64 last,
3928                                                 get_extent_t *get_extent)
3929 {
3930         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
3931         struct extent_map *em;
3932         u64 len;
3933
3934         if (offset >= last)
3935                 return NULL;
3936
3937         while(1) {
3938                 len = last - offset;
3939                 if (len == 0)
3940                         break;
3941                 len = ALIGN(len, sectorsize);
3942                 em = get_extent(inode, NULL, 0, offset, len, 0);
3943                 if (IS_ERR_OR_NULL(em))
3944                         return em;
3945
3946                 /* if this isn't a hole return it */
3947                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
3948                     em->block_start != EXTENT_MAP_HOLE) {
3949                         return em;
3950                 }
3951
3952                 /* this is a hole, advance to the next extent */
3953                 offset = extent_map_end(em);
3954                 free_extent_map(em);
3955                 if (offset >= last)
3956                         break;
3957         }
3958         return NULL;
3959 }
3960
3961 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
3962                 __u64 start, __u64 len, get_extent_t *get_extent)
3963 {
3964         int ret = 0;
3965         u64 off = start;
3966         u64 max = start + len;
3967         u32 flags = 0;
3968         u32 found_type;
3969         u64 last;
3970         u64 last_for_get_extent = 0;
3971         u64 disko = 0;
3972         u64 isize = i_size_read(inode);
3973         struct btrfs_key found_key;
3974         struct extent_map *em = NULL;
3975         struct extent_state *cached_state = NULL;
3976         struct btrfs_path *path;
3977         struct btrfs_file_extent_item *item;
3978         int end = 0;
3979         u64 em_start = 0;
3980         u64 em_len = 0;
3981         u64 em_end = 0;
3982         unsigned long emflags;
3983
3984         if (len == 0)
3985                 return -EINVAL;
3986
3987         path = btrfs_alloc_path();
3988         if (!path)
3989                 return -ENOMEM;
3990         path->leave_spinning = 1;
3991
3992         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
3993         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
3994
3995         /*
3996          * lookup the last file extent.  We're not using i_size here
3997          * because there might be preallocation past i_size
3998          */
3999         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
4000                                        path, btrfs_ino(inode), -1, 0);
4001         if (ret < 0) {
4002                 btrfs_free_path(path);
4003                 return ret;
4004         }
4005         WARN_ON(!ret);
4006         path->slots[0]--;
4007         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
4008                               struct btrfs_file_extent_item);
4009         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4010         found_type = btrfs_key_type(&found_key);
4011
4012         /* No extents, but there might be delalloc bits */
4013         if (found_key.objectid != btrfs_ino(inode) ||
4014             found_type != BTRFS_EXTENT_DATA_KEY) {
4015                 /* have to trust i_size as the end */
4016                 last = (u64)-1;
4017                 last_for_get_extent = isize;
4018         } else {
4019                 /*
4020                  * remember the start of the last extent.  There are a
4021                  * bunch of different factors that go into the length of the
4022                  * extent, so its much less complex to remember where it started
4023                  */
4024                 last = found_key.offset;
4025                 last_for_get_extent = last + 1;
4026         }
4027         btrfs_free_path(path);
4028
4029         /*
4030          * we might have some extents allocated but more delalloc past those
4031          * extents.  so, we trust isize unless the start of the last extent is
4032          * beyond isize
4033          */
4034         if (last < isize) {
4035                 last = (u64)-1;
4036                 last_for_get_extent = isize;
4037         }
4038
4039         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4040                          &cached_state);
4041
4042         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4043                                    get_extent);
4044         if (!em)
4045                 goto out;
4046         if (IS_ERR(em)) {
4047                 ret = PTR_ERR(em);
4048                 goto out;
4049         }
4050
4051         while (!end) {
4052                 u64 offset_in_extent;
4053
4054                 /* break if the extent we found is outside the range */
4055                 if (em->start >= max || extent_map_end(em) < off)
4056                         break;
4057
4058                 /*
4059                  * get_extent may return an extent that starts before our
4060                  * requested range.  We have to make sure the ranges
4061                  * we return to fiemap always move forward and don't
4062                  * overlap, so adjust the offsets here
4063                  */
4064                 em_start = max(em->start, off);
4065
4066                 /*
4067                  * record the offset from the start of the extent
4068                  * for adjusting the disk offset below
4069                  */
4070                 offset_in_extent = em_start - em->start;
4071                 em_end = extent_map_end(em);
4072                 em_len = em_end - em_start;
4073                 emflags = em->flags;
4074                 disko = 0;
4075                 flags = 0;
4076
4077                 /*
4078                  * bump off for our next call to get_extent
4079                  */
4080                 off = extent_map_end(em);
4081                 if (off >= max)
4082                         end = 1;
4083
4084                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4085                         end = 1;
4086                         flags |= FIEMAP_EXTENT_LAST;
4087                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4088                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4089                                   FIEMAP_EXTENT_NOT_ALIGNED);
4090                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4091                         flags |= (FIEMAP_EXTENT_DELALLOC |
4092                                   FIEMAP_EXTENT_UNKNOWN);
4093                 } else {
4094                         disko = em->block_start + offset_in_extent;
4095                 }
4096                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4097                         flags |= FIEMAP_EXTENT_ENCODED;
4098
4099                 free_extent_map(em);
4100                 em = NULL;
4101                 if ((em_start >= last) || em_len == (u64)-1 ||
4102                    (last == (u64)-1 && isize <= em_end)) {
4103                         flags |= FIEMAP_EXTENT_LAST;
4104                         end = 1;
4105                 }
4106
4107                 /* now scan forward to see if this is really the last extent. */
4108                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4109                                            get_extent);
4110                 if (IS_ERR(em)) {
4111                         ret = PTR_ERR(em);
4112                         goto out;
4113                 }
4114                 if (!em) {
4115                         flags |= FIEMAP_EXTENT_LAST;
4116                         end = 1;
4117                 }
4118                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4119                                               em_len, flags);
4120                 if (ret)
4121                         goto out_free;
4122         }
4123 out_free:
4124         free_extent_map(em);
4125 out:
4126         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4127                              &cached_state, GFP_NOFS);
4128         return ret;
4129 }
4130
4131 static void __free_extent_buffer(struct extent_buffer *eb)
4132 {
4133         btrfs_leak_debug_del(&eb->leak_list);
4134         kmem_cache_free(extent_buffer_cache, eb);
4135 }
4136
4137 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
4138                                                    u64 start,
4139                                                    unsigned long len,
4140                                                    gfp_t mask)
4141 {
4142         struct extent_buffer *eb = NULL;
4143
4144         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4145         if (eb == NULL)
4146                 return NULL;
4147         eb->start = start;
4148         eb->len = len;
4149         eb->tree = tree;
4150         eb->bflags = 0;
4151         rwlock_init(&eb->lock);
4152         atomic_set(&eb->write_locks, 0);
4153         atomic_set(&eb->read_locks, 0);
4154         atomic_set(&eb->blocking_readers, 0);
4155         atomic_set(&eb->blocking_writers, 0);
4156         atomic_set(&eb->spinning_readers, 0);
4157         atomic_set(&eb->spinning_writers, 0);
4158         eb->lock_nested = 0;
4159         init_waitqueue_head(&eb->write_lock_wq);
4160         init_waitqueue_head(&eb->read_lock_wq);
4161
4162         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4163
4164         spin_lock_init(&eb->refs_lock);
4165         atomic_set(&eb->refs, 1);
4166         atomic_set(&eb->io_pages, 0);
4167
4168         /*
4169          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4170          */
4171         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4172                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4173         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4174
4175         return eb;
4176 }
4177
4178 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4179 {
4180         unsigned long i;
4181         struct page *p;
4182         struct extent_buffer *new;
4183         unsigned long num_pages = num_extent_pages(src->start, src->len);
4184
4185         new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_ATOMIC);
4186         if (new == NULL)
4187                 return NULL;
4188
4189         for (i = 0; i < num_pages; i++) {
4190                 p = alloc_page(GFP_ATOMIC);
4191                 BUG_ON(!p);
4192                 attach_extent_buffer_page(new, p);
4193                 WARN_ON(PageDirty(p));
4194                 SetPageUptodate(p);
4195                 new->pages[i] = p;
4196         }
4197
4198         copy_extent_buffer(new, src, 0, 0, src->len);
4199         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4200         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4201
4202         return new;
4203 }
4204
4205 struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4206 {
4207         struct extent_buffer *eb;
4208         unsigned long num_pages = num_extent_pages(0, len);
4209         unsigned long i;
4210
4211         eb = __alloc_extent_buffer(NULL, start, len, GFP_ATOMIC);
4212         if (!eb)
4213                 return NULL;
4214
4215         for (i = 0; i < num_pages; i++) {
4216                 eb->pages[i] = alloc_page(GFP_ATOMIC);
4217                 if (!eb->pages[i])
4218                         goto err;
4219         }
4220         set_extent_buffer_uptodate(eb);
4221         btrfs_set_header_nritems(eb, 0);
4222         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4223
4224         return eb;
4225 err:
4226         for (; i > 0; i--)
4227                 __free_page(eb->pages[i - 1]);
4228         __free_extent_buffer(eb);
4229         return NULL;
4230 }
4231
4232 static int extent_buffer_under_io(struct extent_buffer *eb)
4233 {
4234         return (atomic_read(&eb->io_pages) ||
4235                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4236                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4237 }
4238
4239 /*
4240  * Helper for releasing extent buffer page.
4241  */
4242 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4243                                                 unsigned long start_idx)
4244 {
4245         unsigned long index;
4246         unsigned long num_pages;
4247         struct page *page;
4248         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4249
4250         BUG_ON(extent_buffer_under_io(eb));
4251
4252         num_pages = num_extent_pages(eb->start, eb->len);
4253         index = start_idx + num_pages;
4254         if (start_idx >= index)
4255                 return;
4256
4257         do {
4258                 index--;
4259                 page = extent_buffer_page(eb, index);
4260                 if (page && mapped) {
4261                         spin_lock(&page->mapping->private_lock);
4262                         /*
4263                          * We do this since we'll remove the pages after we've
4264                          * removed the eb from the radix tree, so we could race
4265                          * and have this page now attached to the new eb.  So
4266                          * only clear page_private if it's still connected to
4267                          * this eb.
4268                          */
4269                         if (PagePrivate(page) &&
4270                             page->private == (unsigned long)eb) {
4271                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4272                                 BUG_ON(PageDirty(page));
4273                                 BUG_ON(PageWriteback(page));
4274                                 /*
4275                                  * We need to make sure we haven't be attached
4276                                  * to a new eb.
4277                                  */
4278                                 ClearPagePrivate(page);
4279                                 set_page_private(page, 0);
4280                                 /* One for the page private */
4281                                 page_cache_release(page);
4282                         }
4283                         spin_unlock(&page->mapping->private_lock);
4284
4285                 }
4286                 if (page) {
4287                         /* One for when we alloced the page */
4288                         page_cache_release(page);
4289                 }
4290         } while (index != start_idx);
4291 }
4292
4293 /*
4294  * Helper for releasing the extent buffer.
4295  */
4296 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4297 {
4298         btrfs_release_extent_buffer_page(eb, 0);
4299         __free_extent_buffer(eb);
4300 }
4301
4302 static void check_buffer_tree_ref(struct extent_buffer *eb)
4303 {
4304         int refs;
4305         /* the ref bit is tricky.  We have to make sure it is set
4306          * if we have the buffer dirty.   Otherwise the
4307          * code to free a buffer can end up dropping a dirty
4308          * page
4309          *
4310          * Once the ref bit is set, it won't go away while the
4311          * buffer is dirty or in writeback, and it also won't
4312          * go away while we have the reference count on the
4313          * eb bumped.
4314          *
4315          * We can't just set the ref bit without bumping the
4316          * ref on the eb because free_extent_buffer might
4317          * see the ref bit and try to clear it.  If this happens
4318          * free_extent_buffer might end up dropping our original
4319          * ref by mistake and freeing the page before we are able
4320          * to add one more ref.
4321          *
4322          * So bump the ref count first, then set the bit.  If someone
4323          * beat us to it, drop the ref we added.
4324          */
4325         refs = atomic_read(&eb->refs);
4326         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4327                 return;
4328
4329         spin_lock(&eb->refs_lock);
4330         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4331                 atomic_inc(&eb->refs);
4332         spin_unlock(&eb->refs_lock);
4333 }
4334
4335 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4336 {
4337         unsigned long num_pages, i;
4338
4339         check_buffer_tree_ref(eb);
4340
4341         num_pages = num_extent_pages(eb->start, eb->len);
4342         for (i = 0; i < num_pages; i++) {
4343                 struct page *p = extent_buffer_page(eb, i);
4344                 mark_page_accessed(p);
4345         }
4346 }
4347
4348 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
4349                                           u64 start, unsigned long len)
4350 {
4351         unsigned long num_pages = num_extent_pages(start, len);
4352         unsigned long i;
4353         unsigned long index = start >> PAGE_CACHE_SHIFT;
4354         struct extent_buffer *eb;
4355         struct extent_buffer *exists = NULL;
4356         struct page *p;
4357         struct address_space *mapping = tree->mapping;
4358         int uptodate = 1;
4359         int ret;
4360
4361         rcu_read_lock();
4362         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4363         if (eb && atomic_inc_not_zero(&eb->refs)) {
4364                 rcu_read_unlock();
4365                 mark_extent_buffer_accessed(eb);
4366                 return eb;
4367         }
4368         rcu_read_unlock();
4369
4370         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
4371         if (!eb)
4372                 return NULL;
4373
4374         for (i = 0; i < num_pages; i++, index++) {
4375                 p = find_or_create_page(mapping, index, GFP_NOFS);
4376                 if (!p)
4377                         goto free_eb;
4378
4379                 spin_lock(&mapping->private_lock);
4380                 if (PagePrivate(p)) {
4381                         /*
4382                          * We could have already allocated an eb for this page
4383                          * and attached one so lets see if we can get a ref on
4384                          * the existing eb, and if we can we know it's good and
4385                          * we can just return that one, else we know we can just
4386                          * overwrite page->private.
4387                          */
4388                         exists = (struct extent_buffer *)p->private;
4389                         if (atomic_inc_not_zero(&exists->refs)) {
4390                                 spin_unlock(&mapping->private_lock);
4391                                 unlock_page(p);
4392                                 page_cache_release(p);
4393                                 mark_extent_buffer_accessed(exists);
4394                                 goto free_eb;
4395                         }
4396
4397                         /*
4398                          * Do this so attach doesn't complain and we need to
4399                          * drop the ref the old guy had.
4400                          */
4401                         ClearPagePrivate(p);
4402                         WARN_ON(PageDirty(p));
4403                         page_cache_release(p);
4404                 }
4405                 attach_extent_buffer_page(eb, p);
4406                 spin_unlock(&mapping->private_lock);
4407                 WARN_ON(PageDirty(p));
4408                 mark_page_accessed(p);
4409                 eb->pages[i] = p;
4410                 if (!PageUptodate(p))
4411                         uptodate = 0;
4412
4413                 /*
4414                  * see below about how we avoid a nasty race with release page
4415                  * and why we unlock later
4416                  */
4417         }
4418         if (uptodate)
4419                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4420 again:
4421         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4422         if (ret)
4423                 goto free_eb;
4424
4425         spin_lock(&tree->buffer_lock);
4426         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
4427         if (ret == -EEXIST) {
4428                 exists = radix_tree_lookup(&tree->buffer,
4429                                                 start >> PAGE_CACHE_SHIFT);
4430                 if (!atomic_inc_not_zero(&exists->refs)) {
4431                         spin_unlock(&tree->buffer_lock);
4432                         radix_tree_preload_end();
4433                         exists = NULL;
4434                         goto again;
4435                 }
4436                 spin_unlock(&tree->buffer_lock);
4437                 radix_tree_preload_end();
4438                 mark_extent_buffer_accessed(exists);
4439                 goto free_eb;
4440         }
4441         /* add one reference for the tree */
4442         check_buffer_tree_ref(eb);
4443         spin_unlock(&tree->buffer_lock);
4444         radix_tree_preload_end();
4445
4446         /*
4447          * there is a race where release page may have
4448          * tried to find this extent buffer in the radix
4449          * but failed.  It will tell the VM it is safe to
4450          * reclaim the, and it will clear the page private bit.
4451          * We must make sure to set the page private bit properly
4452          * after the extent buffer is in the radix tree so
4453          * it doesn't get lost
4454          */
4455         SetPageChecked(eb->pages[0]);
4456         for (i = 1; i < num_pages; i++) {
4457                 p = extent_buffer_page(eb, i);
4458                 ClearPageChecked(p);
4459                 unlock_page(p);
4460         }
4461         unlock_page(eb->pages[0]);
4462         return eb;
4463
4464 free_eb:
4465         for (i = 0; i < num_pages; i++) {
4466                 if (eb->pages[i])
4467                         unlock_page(eb->pages[i]);
4468         }
4469
4470         WARN_ON(!atomic_dec_and_test(&eb->refs));
4471         btrfs_release_extent_buffer(eb);
4472         return exists;
4473 }
4474
4475 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
4476                                          u64 start, unsigned long len)
4477 {
4478         struct extent_buffer *eb;
4479
4480         rcu_read_lock();
4481         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4482         if (eb && atomic_inc_not_zero(&eb->refs)) {
4483                 rcu_read_unlock();
4484                 mark_extent_buffer_accessed(eb);
4485                 return eb;
4486         }
4487         rcu_read_unlock();
4488
4489         return NULL;
4490 }
4491
4492 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4493 {
4494         struct extent_buffer *eb =
4495                         container_of(head, struct extent_buffer, rcu_head);
4496
4497         __free_extent_buffer(eb);
4498 }
4499
4500 /* Expects to have eb->eb_lock already held */
4501 static int release_extent_buffer(struct extent_buffer *eb)
4502 {
4503         WARN_ON(atomic_read(&eb->refs) == 0);
4504         if (atomic_dec_and_test(&eb->refs)) {
4505                 if (test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags)) {
4506                         spin_unlock(&eb->refs_lock);
4507                 } else {
4508                         struct extent_io_tree *tree = eb->tree;
4509
4510                         spin_unlock(&eb->refs_lock);
4511
4512                         spin_lock(&tree->buffer_lock);
4513                         radix_tree_delete(&tree->buffer,
4514                                           eb->start >> PAGE_CACHE_SHIFT);
4515                         spin_unlock(&tree->buffer_lock);
4516                 }
4517
4518                 /* Should be safe to release our pages at this point */
4519                 btrfs_release_extent_buffer_page(eb, 0);
4520                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4521                 return 1;
4522         }
4523         spin_unlock(&eb->refs_lock);
4524
4525         return 0;
4526 }
4527
4528 void free_extent_buffer(struct extent_buffer *eb)
4529 {
4530         int refs;
4531         int old;
4532         if (!eb)
4533                 return;
4534
4535         while (1) {
4536                 refs = atomic_read(&eb->refs);
4537                 if (refs <= 3)
4538                         break;
4539                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
4540                 if (old == refs)
4541                         return;
4542         }
4543
4544         spin_lock(&eb->refs_lock);
4545         if (atomic_read(&eb->refs) == 2 &&
4546             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4547                 atomic_dec(&eb->refs);
4548
4549         if (atomic_read(&eb->refs) == 2 &&
4550             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4551             !extent_buffer_under_io(eb) &&
4552             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4553                 atomic_dec(&eb->refs);
4554
4555         /*
4556          * I know this is terrible, but it's temporary until we stop tracking
4557          * the uptodate bits and such for the extent buffers.
4558          */
4559         release_extent_buffer(eb);
4560 }
4561
4562 void free_extent_buffer_stale(struct extent_buffer *eb)
4563 {
4564         if (!eb)
4565                 return;
4566
4567         spin_lock(&eb->refs_lock);
4568         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4569
4570         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4571             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4572                 atomic_dec(&eb->refs);
4573         release_extent_buffer(eb);
4574 }
4575
4576 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4577 {
4578         unsigned long i;
4579         unsigned long num_pages;
4580         struct page *page;
4581
4582         num_pages = num_extent_pages(eb->start, eb->len);
4583
4584         for (i = 0; i < num_pages; i++) {
4585                 page = extent_buffer_page(eb, i);
4586                 if (!PageDirty(page))
4587                         continue;
4588
4589                 lock_page(page);
4590                 WARN_ON(!PagePrivate(page));
4591
4592                 clear_page_dirty_for_io(page);
4593                 spin_lock_irq(&page->mapping->tree_lock);
4594                 if (!PageDirty(page)) {
4595                         radix_tree_tag_clear(&page->mapping->page_tree,
4596                                                 page_index(page),
4597                                                 PAGECACHE_TAG_DIRTY);
4598                 }
4599                 spin_unlock_irq(&page->mapping->tree_lock);
4600                 ClearPageError(page);
4601                 unlock_page(page);
4602         }
4603         WARN_ON(atomic_read(&eb->refs) == 0);
4604 }
4605
4606 int set_extent_buffer_dirty(struct extent_buffer *eb)
4607 {
4608         unsigned long i;
4609         unsigned long num_pages;
4610         int was_dirty = 0;
4611
4612         check_buffer_tree_ref(eb);
4613
4614         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4615
4616         num_pages = num_extent_pages(eb->start, eb->len);
4617         WARN_ON(atomic_read(&eb->refs) == 0);
4618         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4619
4620         for (i = 0; i < num_pages; i++)
4621                 set_page_dirty(extent_buffer_page(eb, i));
4622         return was_dirty;
4623 }
4624
4625 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4626 {
4627         unsigned long i;
4628         struct page *page;
4629         unsigned long num_pages;
4630
4631         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4632         num_pages = num_extent_pages(eb->start, eb->len);
4633         for (i = 0; i < num_pages; i++) {
4634                 page = extent_buffer_page(eb, i);
4635                 if (page)
4636                         ClearPageUptodate(page);
4637         }
4638         return 0;
4639 }
4640
4641 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4642 {
4643         unsigned long i;
4644         struct page *page;
4645         unsigned long num_pages;
4646
4647         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4648         num_pages = num_extent_pages(eb->start, eb->len);
4649         for (i = 0; i < num_pages; i++) {
4650                 page = extent_buffer_page(eb, i);
4651                 SetPageUptodate(page);
4652         }
4653         return 0;
4654 }
4655
4656 int extent_buffer_uptodate(struct extent_buffer *eb)
4657 {
4658         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4659 }
4660
4661 int read_extent_buffer_pages(struct extent_io_tree *tree,
4662                              struct extent_buffer *eb, u64 start, int wait,
4663                              get_extent_t *get_extent, int mirror_num)
4664 {
4665         unsigned long i;
4666         unsigned long start_i;
4667         struct page *page;
4668         int err;
4669         int ret = 0;
4670         int locked_pages = 0;
4671         int all_uptodate = 1;
4672         unsigned long num_pages;
4673         unsigned long num_reads = 0;
4674         struct bio *bio = NULL;
4675         unsigned long bio_flags = 0;
4676
4677         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4678                 return 0;
4679
4680         if (start) {
4681                 WARN_ON(start < eb->start);
4682                 start_i = (start >> PAGE_CACHE_SHIFT) -
4683                         (eb->start >> PAGE_CACHE_SHIFT);
4684         } else {
4685                 start_i = 0;
4686         }
4687
4688         num_pages = num_extent_pages(eb->start, eb->len);
4689         for (i = start_i; i < num_pages; i++) {
4690                 page = extent_buffer_page(eb, i);
4691                 if (wait == WAIT_NONE) {
4692                         if (!trylock_page(page))
4693                                 goto unlock_exit;
4694                 } else {
4695                         lock_page(page);
4696                 }
4697                 locked_pages++;
4698                 if (!PageUptodate(page)) {
4699                         num_reads++;
4700                         all_uptodate = 0;
4701                 }
4702         }
4703         if (all_uptodate) {
4704                 if (start_i == 0)
4705                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4706                 goto unlock_exit;
4707         }
4708
4709         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4710         eb->read_mirror = 0;
4711         atomic_set(&eb->io_pages, num_reads);
4712         for (i = start_i; i < num_pages; i++) {
4713                 page = extent_buffer_page(eb, i);
4714                 if (!PageUptodate(page)) {
4715                         ClearPageError(page);
4716                         err = __extent_read_full_page(tree, page,
4717                                                       get_extent, &bio,
4718                                                       mirror_num, &bio_flags,
4719                                                       READ | REQ_META);
4720                         if (err)
4721                                 ret = err;
4722                 } else {
4723                         unlock_page(page);
4724                 }
4725         }
4726
4727         if (bio) {
4728                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
4729                                      bio_flags);
4730                 if (err)
4731                         return err;
4732         }
4733
4734         if (ret || wait != WAIT_COMPLETE)
4735                 return ret;
4736
4737         for (i = start_i; i < num_pages; i++) {
4738                 page = extent_buffer_page(eb, i);
4739                 wait_on_page_locked(page);
4740                 if (!PageUptodate(page))
4741                         ret = -EIO;
4742         }
4743
4744         return ret;
4745
4746 unlock_exit:
4747         i = start_i;
4748         while (locked_pages > 0) {
4749                 page = extent_buffer_page(eb, i);
4750                 i++;
4751                 unlock_page(page);
4752                 locked_pages--;
4753         }
4754         return ret;
4755 }
4756
4757 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4758                         unsigned long start,
4759                         unsigned long len)
4760 {
4761         size_t cur;
4762         size_t offset;
4763         struct page *page;
4764         char *kaddr;
4765         char *dst = (char *)dstv;
4766         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4767         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4768
4769         WARN_ON(start > eb->len);
4770         WARN_ON(start + len > eb->start + eb->len);
4771
4772         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4773
4774         while (len > 0) {
4775                 page = extent_buffer_page(eb, i);
4776
4777                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4778                 kaddr = page_address(page);
4779                 memcpy(dst, kaddr + offset, cur);
4780
4781                 dst += cur;
4782                 len -= cur;
4783                 offset = 0;
4784                 i++;
4785         }
4786 }
4787
4788 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4789                                unsigned long min_len, char **map,
4790                                unsigned long *map_start,
4791                                unsigned long *map_len)
4792 {
4793         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4794         char *kaddr;
4795         struct page *p;
4796         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4797         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4798         unsigned long end_i = (start_offset + start + min_len - 1) >>
4799                 PAGE_CACHE_SHIFT;
4800
4801         if (i != end_i)
4802                 return -EINVAL;
4803
4804         if (i == 0) {
4805                 offset = start_offset;
4806                 *map_start = 0;
4807         } else {
4808                 offset = 0;
4809                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4810         }
4811
4812         if (start + min_len > eb->len) {
4813                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4814                        "wanted %lu %lu\n", (unsigned long long)eb->start,
4815                        eb->len, start, min_len);
4816                 return -EINVAL;
4817         }
4818
4819         p = extent_buffer_page(eb, i);
4820         kaddr = page_address(p);
4821         *map = kaddr + offset;
4822         *map_len = PAGE_CACHE_SIZE - offset;
4823         return 0;
4824 }
4825
4826 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4827                           unsigned long start,
4828                           unsigned long len)
4829 {
4830         size_t cur;
4831         size_t offset;
4832         struct page *page;
4833         char *kaddr;
4834         char *ptr = (char *)ptrv;
4835         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4836         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4837         int ret = 0;
4838
4839         WARN_ON(start > eb->len);
4840         WARN_ON(start + len > eb->start + eb->len);
4841
4842         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4843
4844         while (len > 0) {
4845                 page = extent_buffer_page(eb, i);
4846
4847                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4848
4849                 kaddr = page_address(page);
4850                 ret = memcmp(ptr, kaddr + offset, cur);
4851                 if (ret)
4852                         break;
4853
4854                 ptr += cur;
4855                 len -= cur;
4856                 offset = 0;
4857                 i++;
4858         }
4859         return ret;
4860 }
4861
4862 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4863                          unsigned long start, unsigned long len)
4864 {
4865         size_t cur;
4866         size_t offset;
4867         struct page *page;
4868         char *kaddr;
4869         char *src = (char *)srcv;
4870         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4871         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4872
4873         WARN_ON(start > eb->len);
4874         WARN_ON(start + len > eb->start + eb->len);
4875
4876         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4877
4878         while (len > 0) {
4879                 page = extent_buffer_page(eb, i);
4880                 WARN_ON(!PageUptodate(page));
4881
4882                 cur = min(len, PAGE_CACHE_SIZE - offset);
4883                 kaddr = page_address(page);
4884                 memcpy(kaddr + offset, src, cur);
4885
4886                 src += cur;
4887                 len -= cur;
4888                 offset = 0;
4889                 i++;
4890         }
4891 }
4892
4893 void memset_extent_buffer(struct extent_buffer *eb, char c,
4894                           unsigned long start, unsigned long len)
4895 {
4896         size_t cur;
4897         size_t offset;
4898         struct page *page;
4899         char *kaddr;
4900         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4901         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4902
4903         WARN_ON(start > eb->len);
4904         WARN_ON(start + len > eb->start + eb->len);
4905
4906         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4907
4908         while (len > 0) {
4909                 page = extent_buffer_page(eb, i);
4910                 WARN_ON(!PageUptodate(page));
4911
4912                 cur = min(len, PAGE_CACHE_SIZE - offset);
4913                 kaddr = page_address(page);
4914                 memset(kaddr + offset, c, cur);
4915
4916                 len -= cur;
4917                 offset = 0;
4918                 i++;
4919         }
4920 }
4921
4922 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
4923                         unsigned long dst_offset, unsigned long src_offset,
4924                         unsigned long len)
4925 {
4926         u64 dst_len = dst->len;
4927         size_t cur;
4928         size_t offset;
4929         struct page *page;
4930         char *kaddr;
4931         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4932         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4933
4934         WARN_ON(src->len != dst_len);
4935
4936         offset = (start_offset + dst_offset) &
4937                 ((unsigned long)PAGE_CACHE_SIZE - 1);
4938
4939         while (len > 0) {
4940                 page = extent_buffer_page(dst, i);
4941                 WARN_ON(!PageUptodate(page));
4942
4943                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
4944
4945                 kaddr = page_address(page);
4946                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4947
4948                 src_offset += cur;
4949                 len -= cur;
4950                 offset = 0;
4951                 i++;
4952         }
4953 }
4954
4955 static void move_pages(struct page *dst_page, struct page *src_page,
4956                        unsigned long dst_off, unsigned long src_off,
4957                        unsigned long len)
4958 {
4959         char *dst_kaddr = page_address(dst_page);
4960         if (dst_page == src_page) {
4961                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
4962         } else {
4963                 char *src_kaddr = page_address(src_page);
4964                 char *p = dst_kaddr + dst_off + len;
4965                 char *s = src_kaddr + src_off + len;
4966
4967                 while (len--)
4968                         *--p = *--s;
4969         }
4970 }
4971
4972 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4973 {
4974         unsigned long distance = (src > dst) ? src - dst : dst - src;
4975         return distance < len;
4976 }
4977
4978 static void copy_pages(struct page *dst_page, struct page *src_page,
4979                        unsigned long dst_off, unsigned long src_off,
4980                        unsigned long len)
4981 {
4982         char *dst_kaddr = page_address(dst_page);
4983         char *src_kaddr;
4984         int must_memmove = 0;
4985
4986         if (dst_page != src_page) {
4987                 src_kaddr = page_address(src_page);
4988         } else {
4989                 src_kaddr = dst_kaddr;
4990                 if (areas_overlap(src_off, dst_off, len))
4991                         must_memmove = 1;
4992         }
4993
4994         if (must_memmove)
4995                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
4996         else
4997                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
4998 }
4999
5000 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5001                            unsigned long src_offset, unsigned long len)
5002 {
5003         size_t cur;
5004         size_t dst_off_in_page;
5005         size_t src_off_in_page;
5006         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5007         unsigned long dst_i;
5008         unsigned long src_i;
5009
5010         if (src_offset + len > dst->len) {
5011                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
5012                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5013                 BUG_ON(1);
5014         }
5015         if (dst_offset + len > dst->len) {
5016                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
5017                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5018                 BUG_ON(1);
5019         }
5020
5021         while (len > 0) {
5022                 dst_off_in_page = (start_offset + dst_offset) &
5023                         ((unsigned long)PAGE_CACHE_SIZE - 1);
5024                 src_off_in_page = (start_offset + src_offset) &
5025                         ((unsigned long)PAGE_CACHE_SIZE - 1);
5026
5027                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5028                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5029
5030                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5031                                                src_off_in_page));
5032                 cur = min_t(unsigned long, cur,
5033                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5034
5035                 copy_pages(extent_buffer_page(dst, dst_i),
5036                            extent_buffer_page(dst, src_i),
5037                            dst_off_in_page, src_off_in_page, cur);
5038
5039                 src_offset += cur;
5040                 dst_offset += cur;
5041                 len -= cur;
5042         }
5043 }
5044
5045 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5046                            unsigned long src_offset, unsigned long len)
5047 {
5048         size_t cur;
5049         size_t dst_off_in_page;
5050         size_t src_off_in_page;
5051         unsigned long dst_end = dst_offset + len - 1;
5052         unsigned long src_end = src_offset + len - 1;
5053         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5054         unsigned long dst_i;
5055         unsigned long src_i;
5056
5057         if (src_offset + len > dst->len) {
5058                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
5059                        "len %lu len %lu\n", src_offset, len, dst->len);
5060                 BUG_ON(1);
5061         }
5062         if (dst_offset + len > dst->len) {
5063                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
5064                        "len %lu len %lu\n", dst_offset, len, dst->len);
5065                 BUG_ON(1);
5066         }
5067         if (dst_offset < src_offset) {
5068                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5069                 return;
5070         }
5071         while (len > 0) {
5072                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5073                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5074
5075                 dst_off_in_page = (start_offset + dst_end) &
5076                         ((unsigned long)PAGE_CACHE_SIZE - 1);
5077                 src_off_in_page = (start_offset + src_end) &
5078                         ((unsigned long)PAGE_CACHE_SIZE - 1);
5079
5080                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5081                 cur = min(cur, dst_off_in_page + 1);
5082                 move_pages(extent_buffer_page(dst, dst_i),
5083                            extent_buffer_page(dst, src_i),
5084                            dst_off_in_page - cur + 1,
5085                            src_off_in_page - cur + 1, cur);
5086
5087                 dst_end -= cur;
5088                 src_end -= cur;
5089                 len -= cur;
5090         }
5091 }
5092
5093 int try_release_extent_buffer(struct page *page)
5094 {
5095         struct extent_buffer *eb;
5096
5097         /*
5098          * We need to make sure noboody is attaching this page to an eb right
5099          * now.
5100          */
5101         spin_lock(&page->mapping->private_lock);
5102         if (!PagePrivate(page)) {
5103                 spin_unlock(&page->mapping->private_lock);
5104                 return 1;
5105         }
5106
5107         eb = (struct extent_buffer *)page->private;
5108         BUG_ON(!eb);
5109
5110         /*
5111          * This is a little awful but should be ok, we need to make sure that
5112          * the eb doesn't disappear out from under us while we're looking at
5113          * this page.
5114          */
5115         spin_lock(&eb->refs_lock);
5116         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5117                 spin_unlock(&eb->refs_lock);
5118                 spin_unlock(&page->mapping->private_lock);
5119                 return 0;
5120         }
5121         spin_unlock(&page->mapping->private_lock);
5122
5123         /*
5124          * If tree ref isn't set then we know the ref on this eb is a real ref,
5125          * so just return, this page will likely be freed soon anyway.
5126          */
5127         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5128                 spin_unlock(&eb->refs_lock);
5129                 return 0;
5130         }
5131
5132         return release_extent_buffer(eb);
5133 }