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