fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/module.h>
   8 #include <linux/spinlock.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/swap.h>
  11 #include <linux/writeback.h>
  12 #include <linux/pagevec.h>
  13 #include <linux/prefetch.h>
  14 #include <linux/cleancache.h>
  15 #include "extent_io.h"
  16 #include "extent_map.h"
  17 #include "compat.h"
  18 #include "ctree.h"
  19 #include "btrfs_inode.h"
  20 #include "volumes.h"
  21 #include "check-integrity.h"
  22
  23 static struct kmem_cache *extent_state_cache;
  24 static struct kmem_cache *extent_buffer_cache;
  25
  26 static LIST_HEAD(buffers);
  27 static LIST_HEAD(states);
  28
  29 #define LEAK_DEBUG 0
  30 #if LEAK_DEBUG
  31 static DEFINE_SPINLOCK(leak_lock);
  32 #endif
  33
  34 #define BUFFER_LRU_MAX 64
  35
  36 struct tree_entry {
  37         u64 start;
  38         u64 end;
  39         struct rb_node rb_node;
  40 };
  41
  42 struct extent_page_data {
  43         struct bio *bio;
  44         struct extent_io_tree *tree;
  45         get_extent_t *get_extent;
  46
  47         /* tells writepage not to lock the state bits for this range
  48          * it still does the unlocking
  49          */
  50         unsigned int extent_locked:1;
  51
  52         /* tells the submit_bio code to use a WRITE_SYNC */
  53         unsigned int sync_io:1;
  54 };
  55
  56 int __init extent_io_init(void)
  57 {
  58         extent_state_cache = kmem_cache_create("extent_state",
  59                         sizeof(struct extent_state), 0,
  60                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  61         if (!extent_state_cache)
  62                 return -ENOMEM;
  63
  64         extent_buffer_cache = kmem_cache_create("extent_buffers",
  65                         sizeof(struct extent_buffer), 0,
  66                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  67         if (!extent_buffer_cache)
  68                 goto free_state_cache;
  69         return 0;
  70
  71 free_state_cache:
  72         kmem_cache_destroy(extent_state_cache);
  73         return -ENOMEM;
  74 }
  75
  76 void extent_io_exit(void)
  77 {
  78         struct extent_state *state;
  79         struct extent_buffer *eb;
  80
  81         while (!list_empty(&states)) {
  82                 state = list_entry(states.next, struct extent_state, leak_list);
  83                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  84                        "state %lu in tree %p refs %d\n",
  85                        (unsigned long long)state->start,
  86                        (unsigned long long)state->end,
  87                        state->state, state->tree, atomic_read(&state->refs));
  88                 list_del(&state->leak_list);
  89                 kmem_cache_free(extent_state_cache, state);
  90
  91         }
  92
  93         while (!list_empty(&buffers)) {
  94                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  95                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  96                        "refs %d\n", (unsigned long long)eb->start,
  97                        eb->len, atomic_read(&eb->refs));
  98                 list_del(&eb->leak_list);
  99                 kmem_cache_free(extent_buffer_cache, eb);
 100         }
 101         if (extent_state_cache)
 102                 kmem_cache_destroy(extent_state_cache);
 103         if (extent_buffer_cache)
 104                 kmem_cache_destroy(extent_buffer_cache);
 105 }
 106
 107 void extent_io_tree_init(struct extent_io_tree *tree,
 108                          struct address_space *mapping)
 109 {
 110         tree->state = RB_ROOT;
 111         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 112         tree->ops = NULL;
 113         tree->dirty_bytes = 0;
 114         spin_lock_init(&tree->lock);
 115         spin_lock_init(&tree->buffer_lock);
 116         tree->mapping = mapping;
 117 }
 118
 119 static struct extent_state *alloc_extent_state(gfp_t mask)
 120 {
 121         struct extent_state *state;
 122 #if LEAK_DEBUG
 123         unsigned long flags;
 124 #endif
 125
 126         state = kmem_cache_alloc(extent_state_cache, mask);
 127         if (!state)
 128                 return state;
 129         state->state = 0;
 130         state->private = 0;
 131         state->tree = NULL;
 132 #if LEAK_DEBUG
 133         spin_lock_irqsave(&leak_lock, flags);
 134         list_add(&state->leak_list, &states);
 135         spin_unlock_irqrestore(&leak_lock, flags);
 136 #endif
 137         atomic_set(&state->refs, 1);
 138         init_waitqueue_head(&state->wq);
 139         return state;
 140 }
 141
 142 void free_extent_state(struct extent_state *state)
 143 {
 144         if (!state)
 145                 return;
 146         if (atomic_dec_and_test(&state->refs)) {
 147 #if LEAK_DEBUG
 148                 unsigned long flags;
 149 #endif
 150                 WARN_ON(state->tree);
 151 #if LEAK_DEBUG
 152                 spin_lock_irqsave(&leak_lock, flags);
 153                 list_del(&state->leak_list);
 154                 spin_unlock_irqrestore(&leak_lock, flags);
 155 #endif
 156                 kmem_cache_free(extent_state_cache, state);
 157         }
 158 }
 159
 160 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 161                                    struct rb_node *node)
 162 {
 163         struct rb_node **p = &root->rb_node;
 164         struct rb_node *parent = NULL;
 165         struct tree_entry *entry;
 166
 167         while (*p) {
 168                 parent = *p;
 169                 entry = rb_entry(parent, struct tree_entry, rb_node);
 170
 171                 if (offset < entry->start)
 172                         p = &(*p)->rb_left;
 173                 else if (offset > entry->end)
 174                         p = &(*p)->rb_right;
 175                 else
 176                         return parent;
 177         }
 178
 179         entry = rb_entry(node, struct tree_entry, rb_node);
 180         rb_link_node(node, parent, p);
 181         rb_insert_color(node, root);
 182         return NULL;
 183 }
 184
 185 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 186                                      struct rb_node **prev_ret,
 187                                      struct rb_node **next_ret)
 188 {
 189         struct rb_root *root = &tree->state;
 190         struct rb_node *n = root->rb_node;
 191         struct rb_node *prev = NULL;
 192         struct rb_node *orig_prev = NULL;
 193         struct tree_entry *entry;
 194         struct tree_entry *prev_entry = NULL;
 195
 196         while (n) {
 197                 entry = rb_entry(n, struct tree_entry, rb_node);
 198                 prev = n;
 199                 prev_entry = entry;
 200
 201                 if (offset < entry->start)
 202                         n = n->rb_left;
 203                 else if (offset > entry->end)
 204                         n = n->rb_right;
 205                 else
 206                         return n;
 207         }
 208
 209         if (prev_ret) {
 210                 orig_prev = prev;
 211                 while (prev && offset > prev_entry->end) {
 212                         prev = rb_next(prev);
 213                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 214                 }
 215                 *prev_ret = prev;
 216                 prev = orig_prev;
 217         }
 218
 219         if (next_ret) {
 220                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 221                 while (prev && offset < prev_entry->start) {
 222                         prev = rb_prev(prev);
 223                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 224                 }
 225                 *next_ret = prev;
 226         }
 227         return NULL;
 228 }
 229
 230 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 231                                           u64 offset)
 232 {
 233         struct rb_node *prev = NULL;
 234         struct rb_node *ret;
 235
 236         ret = __etree_search(tree, offset, &prev, NULL);
 237         if (!ret)
 238                 return prev;
 239         return ret;
 240 }
 241
 242 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 243                      struct extent_state *other)
 244 {
 245         if (tree->ops && tree->ops->merge_extent_hook)
 246                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 247                                              other);
 248 }
 249
 250 /*
 251  * utility function to look for merge candidates inside a given range.
 252  * Any extents with matching state are merged together into a single
 253  * extent in the tree.  Extents with EXTENT_IO in their state field
 254  * are not merged because the end_io handlers need to be able to do
 255  * operations on them without sleeping (or doing allocations/splits).
 256  *
 257  * This should be called with the tree lock held.
 258  */
 259 static void merge_state(struct extent_io_tree *tree,
 260                         struct extent_state *state)
 261 {
 262         struct extent_state *other;
 263         struct rb_node *other_node;
 264
 265         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 266                 return;
 267
 268         other_node = rb_prev(&state->rb_node);
 269         if (other_node) {
 270                 other = rb_entry(other_node, struct extent_state, rb_node);
 271                 if (other->end == state->start - 1 &&
 272                     other->state == state->state) {
 273                         merge_cb(tree, state, other);
 274                         state->start = other->start;
 275                         other->tree = NULL;
 276                         rb_erase(&other->rb_node, &tree->state);
 277                         free_extent_state(other);
 278                 }
 279         }
 280         other_node = rb_next(&state->rb_node);
 281         if (other_node) {
 282                 other = rb_entry(other_node, struct extent_state, rb_node);
 283                 if (other->start == state->end + 1 &&
 284                     other->state == state->state) {
 285                         merge_cb(tree, state, other);
 286                         state->end = other->end;
 287                         other->tree = NULL;
 288                         rb_erase(&other->rb_node, &tree->state);
 289                         free_extent_state(other);
 290                 }
 291         }
 292 }
 293
 294 static void set_state_cb(struct extent_io_tree *tree,
 295                          struct extent_state *state, int *bits)
 296 {
 297         if (tree->ops && tree->ops->set_bit_hook)
 298                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 299 }
 300
 301 static void clear_state_cb(struct extent_io_tree *tree,
 302                            struct extent_state *state, int *bits)
 303 {
 304         if (tree->ops && tree->ops->clear_bit_hook)
 305                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 306 }
 307
 308 static void set_state_bits(struct extent_io_tree *tree,
 309                            struct extent_state *state, int *bits);
 310
 311 /*
 312  * insert an extent_state struct into the tree.  'bits' are set on the
 313  * struct before it is inserted.
 314  *
 315  * This may return -EEXIST if the extent is already there, in which case the
 316  * state struct is freed.
 317  *
 318  * The tree lock is not taken internally.  This is a utility function and
 319  * probably isn't what you want to call (see set/clear_extent_bit).
 320  */
 321 static int insert_state(struct extent_io_tree *tree,
 322                         struct extent_state *state, u64 start, u64 end,
 323                         int *bits)
 324 {
 325         struct rb_node *node;
 326
 327         if (end < start) {
 328                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
 329                        (unsigned long long)end,
 330                        (unsigned long long)start);
 331                 WARN_ON(1);
 332         }
 333         state->start = start;
 334         state->end = end;
 335
 336         set_state_bits(tree, state, bits);
 337
 338         node = tree_insert(&tree->state, end, &state->rb_node);
 339         if (node) {
 340                 struct extent_state *found;
 341                 found = rb_entry(node, struct extent_state, rb_node);
 342                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 343                        "%llu %llu\n", (unsigned long long)found->start,
 344                        (unsigned long long)found->end,
 345                        (unsigned long long)start, (unsigned long long)end);
 346                 return -EEXIST;
 347         }
 348         state->tree = tree;
 349         merge_state(tree, state);
 350         return 0;
 351 }
 352
 353 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 354                      u64 split)
 355 {
 356         if (tree->ops && tree->ops->split_extent_hook)
 357                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 358 }
 359
 360 /*
 361  * split a given extent state struct in two, inserting the preallocated
 362  * struct 'prealloc' as the newly created second half.  'split' indicates an
 363  * offset inside 'orig' where it should be split.
 364  *
 365  * Before calling,
 366  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 367  * are two extent state structs in the tree:
 368  * prealloc: [orig->start, split - 1]
 369  * orig: [ split, orig->end ]
 370  *
 371  * The tree locks are not taken by this function. They need to be held
 372  * by the caller.
 373  */
 374 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 375                        struct extent_state *prealloc, u64 split)
 376 {
 377         struct rb_node *node;
 378
 379         split_cb(tree, orig, split);
 380
 381         prealloc->start = orig->start;
 382         prealloc->end = split - 1;
 383         prealloc->state = orig->state;
 384         orig->start = split;
 385
 386         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 387         if (node) {
 388                 free_extent_state(prealloc);
 389                 return -EEXIST;
 390         }
 391         prealloc->tree = tree;
 392         return 0;
 393 }
 394
 395 /*
 396  * utility function to clear some bits in an extent state struct.
 397  * it will optionally wake up any one waiting on this state (wake == 1), or
 398  * forcibly remove the state from the tree (delete == 1).
 399  *
 400  * If no bits are set on the state struct after clearing things, the
 401  * struct is freed and removed from the tree
 402  */
 403 static int clear_state_bit(struct extent_io_tree *tree,
 404                             struct extent_state *state,
 405                             int *bits, int wake)
 406 {
 407         int bits_to_clear = *bits & ~EXTENT_CTLBITS;
 408         int ret = state->state & bits_to_clear;
 409
 410         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 411                 u64 range = state->end - state->start + 1;
 412                 WARN_ON(range > tree->dirty_bytes);
 413                 tree->dirty_bytes -= range;
 414         }
 415         clear_state_cb(tree, state, bits);
 416         state->state &= ~bits_to_clear;
 417         if (wake)
 418                 wake_up(&state->wq);
 419         if (state->state == 0) {
 420                 if (state->tree) {
 421                         rb_erase(&state->rb_node, &tree->state);
 422                         state->tree = NULL;
 423                         free_extent_state(state);
 424                 } else {
 425                         WARN_ON(1);
 426                 }
 427         } else {
 428                 merge_state(tree, state);
 429         }
 430         return ret;
 431 }
 432
 433 static struct extent_state *
 434 alloc_extent_state_atomic(struct extent_state *prealloc)
 435 {
 436         if (!prealloc)
 437                 prealloc = alloc_extent_state(GFP_ATOMIC);
 438
 439         return prealloc;
 440 }
 441
 442 /*
 443  * clear some bits on a range in the tree.  This may require splitting
 444  * or inserting elements in the tree, so the gfp mask is used to
 445  * indicate which allocations or sleeping are allowed.
 446  *
 447  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 448  * the given range from the tree regardless of state (ie for truncate).
 449  *
 450  * the range [start, end] is inclusive.
 451  *
 452  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 453  * bits were already set, or zero if none of the bits were already set.
 454  */
 455 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 456                      int bits, int wake, int delete,
 457                      struct extent_state **cached_state,
 458                      gfp_t mask)
 459 {
 460         struct extent_state *state;
 461         struct extent_state *cached;
 462         struct extent_state *prealloc = NULL;
 463         struct rb_node *next_node;
 464         struct rb_node *node;
 465         u64 last_end;
 466         int err;
 467         int set = 0;
 468         int clear = 0;
 469
 470         if (delete)
 471                 bits |= ~EXTENT_CTLBITS;
 472         bits |= EXTENT_FIRST_DELALLOC;
 473
 474         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 475                 clear = 1;
 476 again:
 477         if (!prealloc && (mask & __GFP_WAIT)) {
 478                 prealloc = alloc_extent_state(mask);
 479                 if (!prealloc)
 480                         return -ENOMEM;
 481         }
 482
 483         spin_lock(&tree->lock);
 484         if (cached_state) {
 485                 cached = *cached_state;
 486
 487                 if (clear) {
 488                         *cached_state = NULL;
 489                         cached_state = NULL;
 490                 }
 491
 492                 if (cached && cached->tree && cached->start <= start &&
 493                     cached->end > start) {
 494                         if (clear)
 495                                 atomic_dec(&cached->refs);
 496                         state = cached;
 497                         goto hit_next;
 498                 }
 499                 if (clear)
 500                         free_extent_state(cached);
 501         }
 502         /*
 503          * this search will find the extents that end after
 504          * our range starts
 505          */
 506         node = tree_search(tree, start);
 507         if (!node)
 508                 goto out;
 509         state = rb_entry(node, struct extent_state, rb_node);
 510 hit_next:
 511         if (state->start > end)
 512                 goto out;
 513         WARN_ON(state->end < start);
 514         last_end = state->end;
 515
 516         /*
 517          *     | ---- desired range ---- |
 518          *  | state | or
 519          *  | ------------- state -------------- |
 520          *
 521          * We need to split the extent we found, and may flip
 522          * bits on second half.
 523          *
 524          * If the extent we found extends past our range, we
 525          * just split and search again.  It'll get split again
 526          * the next time though.
 527          *
 528          * If the extent we found is inside our range, we clear
 529          * the desired bit on it.
 530          */
 531
 532         if (state->start < start) {
 533                 prealloc = alloc_extent_state_atomic(prealloc);
 534                 BUG_ON(!prealloc);
 535                 err = split_state(tree, state, prealloc, start);
 536                 BUG_ON(err == -EEXIST);
 537                 prealloc = NULL;
 538                 if (err)
 539                         goto out;
 540                 if (state->end <= end) {
 541                         set |= clear_state_bit(tree, state, &bits, wake);
 542                         if (last_end == (u64)-1)
 543                                 goto out;
 544                         start = last_end + 1;
 545                 }
 546                 goto search_again;
 547         }
 548         /*
 549          * | ---- desired range ---- |
 550          *                        | state |
 551          * We need to split the extent, and clear the bit
 552          * on the first half
 553          */
 554         if (state->start <= end && state->end > end) {
 555                 prealloc = alloc_extent_state_atomic(prealloc);
 556                 BUG_ON(!prealloc);
 557                 err = split_state(tree, state, prealloc, end + 1);
 558                 BUG_ON(err == -EEXIST);
 559                 if (wake)
 560                         wake_up(&state->wq);
 561
 562                 set |= clear_state_bit(tree, prealloc, &bits, wake);
 563
 564                 prealloc = NULL;
 565                 goto out;
 566         }
 567
 568         if (state->end < end && prealloc && !need_resched())
 569                 next_node = rb_next(&state->rb_node);
 570         else
 571                 next_node = NULL;
 572
 573         set |= clear_state_bit(tree, state, &bits, wake);
 574         if (last_end == (u64)-1)
 575                 goto out;
 576         start = last_end + 1;
 577         if (start <= end && next_node) {
 578                 state = rb_entry(next_node, struct extent_state,
 579                                  rb_node);
 580                 if (state->start == start)
 581                         goto hit_next;
 582         }
 583         goto search_again;
 584
 585 out:
 586         spin_unlock(&tree->lock);
 587         if (prealloc)
 588                 free_extent_state(prealloc);
 589
 590         return set;
 591
 592 search_again:
 593         if (start > end)
 594                 goto out;
 595         spin_unlock(&tree->lock);
 596         if (mask & __GFP_WAIT)
 597                 cond_resched();
 598         goto again;
 599 }
 600
 601 static int wait_on_state(struct extent_io_tree *tree,
 602                          struct extent_state *state)
 603                 __releases(tree->lock)
 604                 __acquires(tree->lock)
 605 {
 606         DEFINE_WAIT(wait);
 607         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 608         spin_unlock(&tree->lock);
 609         schedule();
 610         spin_lock(&tree->lock);
 611         finish_wait(&state->wq, &wait);
 612         return 0;
 613 }
 614
 615 /*
 616  * waits for one or more bits to clear on a range in the state tree.
 617  * The range [start, end] is inclusive.
 618  * The tree lock is taken by this function
 619  */
 620 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 621 {
 622         struct extent_state *state;
 623         struct rb_node *node;
 624
 625         spin_lock(&tree->lock);
 626 again:
 627         while (1) {
 628                 /*
 629                  * this search will find all the extents that end after
 630                  * our range starts
 631                  */
 632                 node = tree_search(tree, start);
 633                 if (!node)
 634                         break;
 635
 636                 state = rb_entry(node, struct extent_state, rb_node);
 637
 638                 if (state->start > end)
 639                         goto out;
 640
 641                 if (state->state & bits) {
 642                         start = state->start;
 643                         atomic_inc(&state->refs);
 644                         wait_on_state(tree, state);
 645                         free_extent_state(state);
 646                         goto again;
 647                 }
 648                 start = state->end + 1;
 649
 650                 if (start > end)
 651                         break;
 652
 653                 cond_resched_lock(&tree->lock);
 654         }
 655 out:
 656         spin_unlock(&tree->lock);
 657         return 0;
 658 }
 659
 660 static void set_state_bits(struct extent_io_tree *tree,
 661                            struct extent_state *state,
 662                            int *bits)
 663 {
 664         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 665
 666         set_state_cb(tree, state, bits);
 667         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 668                 u64 range = state->end - state->start + 1;
 669                 tree->dirty_bytes += range;
 670         }
 671         state->state |= bits_to_set;
 672 }
 673
 674 static void cache_state(struct extent_state *state,
 675                         struct extent_state **cached_ptr)
 676 {
 677         if (cached_ptr && !(*cached_ptr)) {
 678                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 679                         *cached_ptr = state;
 680                         atomic_inc(&state->refs);
 681                 }
 682         }
 683 }
 684
 685 static void uncache_state(struct extent_state **cached_ptr)
 686 {
 687         if (cached_ptr && (*cached_ptr)) {
 688                 struct extent_state *state = *cached_ptr;
 689                 *cached_ptr = NULL;
 690                 free_extent_state(state);
 691         }
 692 }
 693
 694 /*
 695  * set some bits on a range in the tree.  This may require allocations or
 696  * sleeping, so the gfp mask is used to indicate what is allowed.
 697  *
 698  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 699  * part of the range already has the desired bits set.  The start of the
 700  * existing range is returned in failed_start in this case.
 701  *
 702  * [start, end] is inclusive This takes the tree lock.
 703  */
 704
 705 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 706                    int bits, int exclusive_bits, u64 *failed_start,
 707                    struct extent_state **cached_state, gfp_t mask)
 708 {
 709         struct extent_state *state;
 710         struct extent_state *prealloc = NULL;
 711         struct rb_node *node;
 712         int err = 0;
 713         u64 last_start;
 714         u64 last_end;
 715
 716         bits |= EXTENT_FIRST_DELALLOC;
 717 again:
 718         if (!prealloc && (mask & __GFP_WAIT)) {
 719                 prealloc = alloc_extent_state(mask);
 720                 BUG_ON(!prealloc);
 721         }
 722
 723         spin_lock(&tree->lock);
 724         if (cached_state && *cached_state) {
 725                 state = *cached_state;
 726                 if (state->start <= start && state->end > start &&
 727                     state->tree) {
 728                         node = &state->rb_node;
 729                         goto hit_next;
 730                 }
 731         }
 732         /*
 733          * this search will find all the extents that end after
 734          * our range starts.
 735          */
 736         node = tree_search(tree, start);
 737         if (!node) {
 738                 prealloc = alloc_extent_state_atomic(prealloc);
 739                 BUG_ON(!prealloc);
 740                 err = insert_state(tree, prealloc, start, end, &bits);
 741                 prealloc = NULL;
 742                 BUG_ON(err == -EEXIST);
 743                 goto out;
 744         }
 745         state = rb_entry(node, struct extent_state, rb_node);
 746 hit_next:
 747         last_start = state->start;
 748         last_end = state->end;
 749
 750         /*
 751          * | ---- desired range ---- |
 752          * | state |
 753          *
 754          * Just lock what we found and keep going
 755          */
 756         if (state->start == start && state->end <= end) {
 757                 struct rb_node *next_node;
 758                 if (state->state & exclusive_bits) {
 759                         *failed_start = state->start;
 760                         err = -EEXIST;
 761                         goto out;
 762                 }
 763
 764                 set_state_bits(tree, state, &bits);
 765
 766                 cache_state(state, cached_state);
 767                 merge_state(tree, state);
 768                 if (last_end == (u64)-1)
 769                         goto out;
 770
 771                 start = last_end + 1;
 772                 next_node = rb_next(&state->rb_node);
 773                 if (next_node && start < end && prealloc && !need_resched()) {
 774                         state = rb_entry(next_node, struct extent_state,
 775                                          rb_node);
 776                         if (state->start == start)
 777                                 goto hit_next;
 778                 }
 779                 goto search_again;
 780         }
 781
 782         /*
 783          *     | ---- desired range ---- |
 784          * | state |
 785          *   or
 786          * | ------------- state -------------- |
 787          *
 788          * We need to split the extent we found, and may flip bits on
 789          * second half.
 790          *
 791          * If the extent we found extends past our
 792          * range, we just split and search again.  It'll get split
 793          * again the next time though.
 794          *
 795          * If the extent we found is inside our range, we set the
 796          * desired bit on it.
 797          */
 798         if (state->start < start) {
 799                 if (state->state & exclusive_bits) {
 800                         *failed_start = start;
 801                         err = -EEXIST;
 802                         goto out;
 803                 }
 804
 805                 prealloc = alloc_extent_state_atomic(prealloc);
 806                 BUG_ON(!prealloc);
 807                 err = split_state(tree, state, prealloc, start);
 808                 BUG_ON(err == -EEXIST);
 809                 prealloc = NULL;
 810                 if (err)
 811                         goto out;
 812                 if (state->end <= end) {
 813                         set_state_bits(tree, state, &bits);
 814                         cache_state(state, cached_state);
 815                         merge_state(tree, state);
 816                         if (last_end == (u64)-1)
 817                                 goto out;
 818                         start = last_end + 1;
 819                 }
 820                 goto search_again;
 821         }
 822         /*
 823          * | ---- desired range ---- |
 824          *     | state | or               | state |
 825          *
 826          * There's a hole, we need to insert something in it and
 827          * ignore the extent we found.
 828          */
 829         if (state->start > start) {
 830                 u64 this_end;
 831                 if (end < last_start)
 832                         this_end = end;
 833                 else
 834                         this_end = last_start - 1;
 835
 836                 prealloc = alloc_extent_state_atomic(prealloc);
 837                 BUG_ON(!prealloc);
 838
 839                 /*
 840                  * Avoid to free 'prealloc' if it can be merged with
 841                  * the later extent.
 842                  */
 843                 err = insert_state(tree, prealloc, start, this_end,
 844                                    &bits);
 845                 BUG_ON(err == -EEXIST);
 846                 if (err) {
 847                         free_extent_state(prealloc);
 848                         prealloc = NULL;
 849                         goto out;
 850                 }
 851                 cache_state(prealloc, cached_state);
 852                 prealloc = NULL;
 853                 start = this_end + 1;
 854                 goto search_again;
 855         }
 856         /*
 857          * | ---- desired range ---- |
 858          *                        | state |
 859          * We need to split the extent, and set the bit
 860          * on the first half
 861          */
 862         if (state->start <= end && state->end > end) {
 863                 if (state->state & exclusive_bits) {
 864                         *failed_start = start;
 865                         err = -EEXIST;
 866                         goto out;
 867                 }
 868
 869                 prealloc = alloc_extent_state_atomic(prealloc);
 870                 BUG_ON(!prealloc);
 871                 err = split_state(tree, state, prealloc, end + 1);
 872                 BUG_ON(err == -EEXIST);
 873
 874                 set_state_bits(tree, prealloc, &bits);
 875                 cache_state(prealloc, cached_state);
 876                 merge_state(tree, prealloc);
 877                 prealloc = NULL;
 878                 goto out;
 879         }
 880
 881         goto search_again;
 882
 883 out:
 884         spin_unlock(&tree->lock);
 885         if (prealloc)
 886                 free_extent_state(prealloc);
 887
 888         return err;
 889
 890 search_again:
 891         if (start > end)
 892                 goto out;
 893         spin_unlock(&tree->lock);
 894         if (mask & __GFP_WAIT)
 895                 cond_resched();
 896         goto again;
 897 }
 898
 899 /**
 900  * convert_extent - convert all bits in a given range from one bit to another
 901  * @tree:       the io tree to search
 902  * @start:      the start offset in bytes
 903  * @end:        the end offset in bytes (inclusive)
 904  * @bits:       the bits to set in this range
 905  * @clear_bits: the bits to clear in this range
 906  * @mask:       the allocation mask
 907  *
 908  * This will go through and set bits for the given range.  If any states exist
 909  * already in this range they are set with the given bit and cleared of the
 910  * clear_bits.  This is only meant to be used by things that are mergeable, ie
 911  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
 912  * boundary bits like LOCK.
 913  */
 914 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 915                        int bits, int clear_bits, gfp_t mask)
 916 {
 917         struct extent_state *state;
 918         struct extent_state *prealloc = NULL;
 919         struct rb_node *node;
 920         int err = 0;
 921         u64 last_start;
 922         u64 last_end;
 923
 924 again:
 925         if (!prealloc && (mask & __GFP_WAIT)) {
 926                 prealloc = alloc_extent_state(mask);
 927                 if (!prealloc)
 928                         return -ENOMEM;
 929         }
 930
 931         spin_lock(&tree->lock);
 932         /*
 933          * this search will find all the extents that end after
 934          * our range starts.
 935          */
 936         node = tree_search(tree, start);
 937         if (!node) {
 938                 prealloc = alloc_extent_state_atomic(prealloc);
 939                 if (!prealloc) {
 940                         err = -ENOMEM;
 941                         goto out;
 942                 }
 943                 err = insert_state(tree, prealloc, start, end, &bits);
 944                 prealloc = NULL;
 945                 BUG_ON(err == -EEXIST);
 946                 goto out;
 947         }
 948         state = rb_entry(node, struct extent_state, rb_node);
 949 hit_next:
 950         last_start = state->start;
 951         last_end = state->end;
 952
 953         /*
 954          * | ---- desired range ---- |
 955          * | state |
 956          *
 957          * Just lock what we found and keep going
 958          */
 959         if (state->start == start && state->end <= end) {
 960                 struct rb_node *next_node;
 961
 962                 set_state_bits(tree, state, &bits);
 963                 clear_state_bit(tree, state, &clear_bits, 0);
 964
 965                 merge_state(tree, state);
 966                 if (last_end == (u64)-1)
 967                         goto out;
 968
 969                 start = last_end + 1;
 970                 next_node = rb_next(&state->rb_node);
 971                 if (next_node && start < end && prealloc && !need_resched()) {
 972                         state = rb_entry(next_node, struct extent_state,
 973                                          rb_node);
 974                         if (state->start == start)
 975                                 goto hit_next;
 976                 }
 977                 goto search_again;
 978         }
 979
 980         /*
 981          *     | ---- desired range ---- |
 982          * | state |
 983          *   or
 984          * | ------------- state -------------- |
 985          *
 986          * We need to split the extent we found, and may flip bits on
 987          * second half.
 988          *
 989          * If the extent we found extends past our
 990          * range, we just split and search again.  It'll get split
 991          * again the next time though.
 992          *
 993          * If the extent we found is inside our range, we set the
 994          * desired bit on it.
 995          */
 996         if (state->start < start) {
 997                 prealloc = alloc_extent_state_atomic(prealloc);
 998                 if (!prealloc) {
 999                         err = -ENOMEM;
1000                         goto out;
1001                 }
1002                 err = split_state(tree, state, prealloc, start);
1003                 BUG_ON(err == -EEXIST);
1004                 prealloc = NULL;
1005                 if (err)
1006                         goto out;
1007                 if (state->end <= end) {
1008                         set_state_bits(tree, state, &bits);
1009                         clear_state_bit(tree, state, &clear_bits, 0);
1010                         merge_state(tree, state);
1011                         if (last_end == (u64)-1)
1012                                 goto out;
1013                         start = last_end + 1;
1014                 }
1015                 goto search_again;
1016         }
1017         /*
1018          * | ---- desired range ---- |
1019          *     | state | or               | state |
1020          *
1021          * There's a hole, we need to insert something in it and
1022          * ignore the extent we found.
1023          */
1024         if (state->start > start) {
1025                 u64 this_end;
1026                 if (end < last_start)
1027                         this_end = end;
1028                 else
1029                         this_end = last_start - 1;
1030
1031                 prealloc = alloc_extent_state_atomic(prealloc);
1032                 if (!prealloc) {
1033                         err = -ENOMEM;
1034                         goto out;
1035                 }
1036
1037                 /*
1038                  * Avoid to free 'prealloc' if it can be merged with
1039                  * the later extent.
1040                  */
1041                 err = insert_state(tree, prealloc, start, this_end,
1042                                    &bits);
1043                 BUG_ON(err == -EEXIST);
1044                 if (err) {
1045                         free_extent_state(prealloc);
1046                         prealloc = NULL;
1047                         goto out;
1048                 }
1049                 prealloc = NULL;
1050                 start = this_end + 1;
1051                 goto search_again;
1052         }
1053         /*
1054          * | ---- desired range ---- |
1055          *                        | state |
1056          * We need to split the extent, and set the bit
1057          * on the first half
1058          */
1059         if (state->start <= end && state->end > end) {
1060                 prealloc = alloc_extent_state_atomic(prealloc);
1061                 if (!prealloc) {
1062                         err = -ENOMEM;
1063                         goto out;
1064                 }
1065
1066                 err = split_state(tree, state, prealloc, end + 1);
1067                 BUG_ON(err == -EEXIST);
1068
1069                 set_state_bits(tree, prealloc, &bits);
1070                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1071
1072                 merge_state(tree, prealloc);
1073                 prealloc = NULL;
1074                 goto out;
1075         }
1076
1077         goto search_again;
1078
1079 out:
1080         spin_unlock(&tree->lock);
1081         if (prealloc)
1082                 free_extent_state(prealloc);
1083
1084         return err;
1085
1086 search_again:
1087         if (start > end)
1088                 goto out;
1089         spin_unlock(&tree->lock);
1090         if (mask & __GFP_WAIT)
1091                 cond_resched();
1092         goto again;
1093 }
1094
1095 /* wrappers around set/clear extent bit */
1096 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1097                      gfp_t mask)
1098 {
1099         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
1100                               NULL, mask);
1101 }
1102
1103 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1104                     int bits, gfp_t mask)
1105 {
1106         return set_extent_bit(tree, start, end, bits, 0, NULL,
1107                               NULL, mask);
1108 }
1109
1110 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1111                       int bits, gfp_t mask)
1112 {
1113         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1114 }
1115
1116 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1117                         struct extent_state **cached_state, gfp_t mask)
1118 {
1119         return set_extent_bit(tree, start, end,
1120                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1121                               0, NULL, cached_state, mask);
1122 }
1123
1124 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1125                        gfp_t mask)
1126 {
1127         return clear_extent_bit(tree, start, end,
1128                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1129                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1130 }
1131
1132 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1133                      gfp_t mask)
1134 {
1135         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
1136                               NULL, mask);
1137 }
1138
1139 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1140                         struct extent_state **cached_state, gfp_t mask)
1141 {
1142         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
1143                               NULL, cached_state, mask);
1144 }
1145
1146 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
1147                                  u64 end, struct extent_state **cached_state,
1148                                  gfp_t mask)
1149 {
1150         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1151                                 cached_state, mask);
1152 }
1153
1154 /*
1155  * either insert or lock state struct between start and end use mask to tell
1156  * us if waiting is desired.
1157  */
1158 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1159                      int bits, struct extent_state **cached_state, gfp_t mask)
1160 {
1161         int err;
1162         u64 failed_start;
1163         while (1) {
1164                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1165                                      EXTENT_LOCKED, &failed_start,
1166                                      cached_state, mask);
1167                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
1168                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1169                         start = failed_start;
1170                 } else {
1171                         break;
1172                 }
1173                 WARN_ON(start > end);
1174         }
1175         return err;
1176 }
1177
1178 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1179 {
1180         return lock_extent_bits(tree, start, end, 0, NULL, mask);
1181 }
1182
1183 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1184                     gfp_t mask)
1185 {
1186         int err;
1187         u64 failed_start;
1188
1189         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1190                              &failed_start, NULL, mask);
1191         if (err == -EEXIST) {
1192                 if (failed_start > start)
1193                         clear_extent_bit(tree, start, failed_start - 1,
1194                                          EXTENT_LOCKED, 1, 0, NULL, mask);
1195                 return 0;
1196         }
1197         return 1;
1198 }
1199
1200 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1201                          struct extent_state **cached, gfp_t mask)
1202 {
1203         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1204                                 mask);
1205 }
1206
1207 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1208 {
1209         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1210                                 mask);
1211 }
1212
1213 /*
1214  * helper function to set both pages and extents in the tree writeback
1215  */
1216 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1217 {
1218         unsigned long index = start >> PAGE_CACHE_SHIFT;
1219         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1220         struct page *page;
1221
1222         while (index <= end_index) {
1223                 page = find_get_page(tree->mapping, index);
1224                 BUG_ON(!page);
1225                 set_page_writeback(page);
1226                 page_cache_release(page);
1227                 index++;
1228         }
1229         return 0;
1230 }
1231
1232 /* find the first state struct with 'bits' set after 'start', and
1233  * return it.  tree->lock must be held.  NULL will returned if
1234  * nothing was found after 'start'
1235  */
1236 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1237                                                  u64 start, int bits)
1238 {
1239         struct rb_node *node;
1240         struct extent_state *state;
1241
1242         /*
1243          * this search will find all the extents that end after
1244          * our range starts.
1245          */
1246         node = tree_search(tree, start);
1247         if (!node)
1248                 goto out;
1249
1250         while (1) {
1251                 state = rb_entry(node, struct extent_state, rb_node);
1252                 if (state->end >= start && (state->state & bits))
1253                         return state;
1254
1255                 node = rb_next(node);
1256                 if (!node)
1257                         break;
1258         }
1259 out:
1260         return NULL;
1261 }
1262
1263 /*
1264  * find the first offset in the io tree with 'bits' set. zero is
1265  * returned if we find something, and *start_ret and *end_ret are
1266  * set to reflect the state struct that was found.
1267  *
1268  * If nothing was found, 1 is returned, < 0 on error
1269  */
1270 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1271                           u64 *start_ret, u64 *end_ret, int bits)
1272 {
1273         struct extent_state *state;
1274         int ret = 1;
1275
1276         spin_lock(&tree->lock);
1277         state = find_first_extent_bit_state(tree, start, bits);
1278         if (state) {
1279                 *start_ret = state->start;
1280                 *end_ret = state->end;
1281                 ret = 0;
1282         }
1283         spin_unlock(&tree->lock);
1284         return ret;
1285 }
1286
1287 /*
1288  * find a contiguous range of bytes in the file marked as delalloc, not
1289  * more than 'max_bytes'.  start and end are used to return the range,
1290  *
1291  * 1 is returned if we find something, 0 if nothing was in the tree
1292  */
1293 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1294                                         u64 *start, u64 *end, u64 max_bytes,
1295                                         struct extent_state **cached_state)
1296 {
1297         struct rb_node *node;
1298         struct extent_state *state;
1299         u64 cur_start = *start;
1300         u64 found = 0;
1301         u64 total_bytes = 0;
1302
1303         spin_lock(&tree->lock);
1304
1305         /*
1306          * this search will find all the extents that end after
1307          * our range starts.
1308          */
1309         node = tree_search(tree, cur_start);
1310         if (!node) {
1311                 if (!found)
1312                         *end = (u64)-1;
1313                 goto out;
1314         }
1315
1316         while (1) {
1317                 state = rb_entry(node, struct extent_state, rb_node);
1318                 if (found && (state->start != cur_start ||
1319                               (state->state & EXTENT_BOUNDARY))) {
1320                         goto out;
1321                 }
1322                 if (!(state->state & EXTENT_DELALLOC)) {
1323                         if (!found)
1324                                 *end = state->end;
1325                         goto out;
1326                 }
1327                 if (!found) {
1328                         *start = state->start;
1329                         *cached_state = state;
1330                         atomic_inc(&state->refs);
1331                 }
1332                 found++;
1333                 *end = state->end;
1334                 cur_start = state->end + 1;
1335                 node = rb_next(node);
1336                 if (!node)
1337                         break;
1338                 total_bytes += state->end - state->start + 1;
1339                 if (total_bytes >= max_bytes)
1340                         break;
1341         }
1342 out:
1343         spin_unlock(&tree->lock);
1344         return found;
1345 }
1346
1347 static noinline int __unlock_for_delalloc(struct inode *inode,
1348                                           struct page *locked_page,
1349                                           u64 start, u64 end)
1350 {
1351         int ret;
1352         struct page *pages[16];
1353         unsigned long index = start >> PAGE_CACHE_SHIFT;
1354         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1355         unsigned long nr_pages = end_index - index + 1;
1356         int i;
1357
1358         if (index == locked_page->index && end_index == index)
1359                 return 0;
1360
1361         while (nr_pages > 0) {
1362                 ret = find_get_pages_contig(inode->i_mapping, index,
1363                                      min_t(unsigned long, nr_pages,
1364                                      ARRAY_SIZE(pages)), pages);
1365                 for (i = 0; i < ret; i++) {
1366                         if (pages[i] != locked_page)
1367                                 unlock_page(pages[i]);
1368                         page_cache_release(pages[i]);
1369                 }
1370                 nr_pages -= ret;
1371                 index += ret;
1372                 cond_resched();
1373         }
1374         return 0;
1375 }
1376
1377 static noinline int lock_delalloc_pages(struct inode *inode,
1378                                         struct page *locked_page,
1379                                         u64 delalloc_start,
1380                                         u64 delalloc_end)
1381 {
1382         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1383         unsigned long start_index = index;
1384         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1385         unsigned long pages_locked = 0;
1386         struct page *pages[16];
1387         unsigned long nrpages;
1388         int ret;
1389         int i;
1390
1391         /* the caller is responsible for locking the start index */
1392         if (index == locked_page->index && index == end_index)
1393                 return 0;
1394
1395         /* skip the page at the start index */
1396         nrpages = end_index - index + 1;
1397         while (nrpages > 0) {
1398                 ret = find_get_pages_contig(inode->i_mapping, index,
1399                                      min_t(unsigned long,
1400                                      nrpages, ARRAY_SIZE(pages)), pages);
1401                 if (ret == 0) {
1402                         ret = -EAGAIN;
1403                         goto done;
1404                 }
1405                 /* now we have an array of pages, lock them all */
1406                 for (i = 0; i < ret; i++) {
1407                         /*
1408                          * the caller is taking responsibility for
1409                          * locked_page
1410                          */
1411                         if (pages[i] != locked_page) {
1412                                 lock_page(pages[i]);
1413                                 if (!PageDirty(pages[i]) ||
1414                                     pages[i]->mapping != inode->i_mapping) {
1415                                         ret = -EAGAIN;
1416                                         unlock_page(pages[i]);
1417                                         page_cache_release(pages[i]);
1418                                         goto done;
1419                                 }
1420                         }
1421                         page_cache_release(pages[i]);
1422                         pages_locked++;
1423                 }
1424                 nrpages -= ret;
1425                 index += ret;
1426                 cond_resched();
1427         }
1428         ret = 0;
1429 done:
1430         if (ret && pages_locked) {
1431                 __unlock_for_delalloc(inode, locked_page,
1432                               delalloc_start,
1433                               ((u64)(start_index + pages_locked - 1)) <<
1434                               PAGE_CACHE_SHIFT);
1435         }
1436         return ret;
1437 }
1438
1439 /*
1440  * find a contiguous range of bytes in the file marked as delalloc, not
1441  * more than 'max_bytes'.  start and end are used to return the range,
1442  *
1443  * 1 is returned if we find something, 0 if nothing was in the tree
1444  */
1445 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1446                                              struct extent_io_tree *tree,
1447                                              struct page *locked_page,
1448                                              u64 *start, u64 *end,
1449                                              u64 max_bytes)
1450 {
1451         u64 delalloc_start;
1452         u64 delalloc_end;
1453         u64 found;
1454         struct extent_state *cached_state = NULL;
1455         int ret;
1456         int loops = 0;
1457
1458 again:
1459         /* step one, find a bunch of delalloc bytes starting at start */
1460         delalloc_start = *start;
1461         delalloc_end = 0;
1462         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1463                                     max_bytes, &cached_state);
1464         if (!found || delalloc_end <= *start) {
1465                 *start = delalloc_start;
1466                 *end = delalloc_end;
1467                 free_extent_state(cached_state);
1468                 return found;
1469         }
1470
1471         /*
1472          * start comes from the offset of locked_page.  We have to lock
1473          * pages in order, so we can't process delalloc bytes before
1474          * locked_page
1475          */
1476         if (delalloc_start < *start)
1477                 delalloc_start = *start;
1478
1479         /*
1480          * make sure to limit the number of pages we try to lock down
1481          * if we're looping.
1482          */
1483         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1484                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1485
1486         /* step two, lock all the pages after the page that has start */
1487         ret = lock_delalloc_pages(inode, locked_page,
1488                                   delalloc_start, delalloc_end);
1489         if (ret == -EAGAIN) {
1490                 /* some of the pages are gone, lets avoid looping by
1491                  * shortening the size of the delalloc range we're searching
1492                  */
1493                 free_extent_state(cached_state);
1494                 if (!loops) {
1495                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1496                         max_bytes = PAGE_CACHE_SIZE - offset;
1497                         loops = 1;
1498                         goto again;
1499                 } else {
1500                         found = 0;
1501                         goto out_failed;
1502                 }
1503         }
1504         BUG_ON(ret);
1505
1506         /* step three, lock the state bits for the whole range */
1507         lock_extent_bits(tree, delalloc_start, delalloc_end,
1508                          0, &cached_state, GFP_NOFS);
1509
1510         /* then test to make sure it is all still delalloc */
1511         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1512                              EXTENT_DELALLOC, 1, cached_state);
1513         if (!ret) {
1514                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1515                                      &cached_state, GFP_NOFS);
1516                 __unlock_for_delalloc(inode, locked_page,
1517                               delalloc_start, delalloc_end);
1518                 cond_resched();
1519                 goto again;
1520         }
1521         free_extent_state(cached_state);
1522         *start = delalloc_start;
1523         *end = delalloc_end;
1524 out_failed:
1525         return found;
1526 }
1527
1528 int extent_clear_unlock_delalloc(struct inode *inode,
1529                                 struct extent_io_tree *tree,
1530                                 u64 start, u64 end, struct page *locked_page,
1531                                 unsigned long op)
1532 {
1533         int ret;
1534         struct page *pages[16];
1535         unsigned long index = start >> PAGE_CACHE_SHIFT;
1536         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1537         unsigned long nr_pages = end_index - index + 1;
1538         int i;
1539         int clear_bits = 0;
1540
1541         if (op & EXTENT_CLEAR_UNLOCK)
1542                 clear_bits |= EXTENT_LOCKED;
1543         if (op & EXTENT_CLEAR_DIRTY)
1544                 clear_bits |= EXTENT_DIRTY;
1545
1546         if (op & EXTENT_CLEAR_DELALLOC)
1547                 clear_bits |= EXTENT_DELALLOC;
1548
1549         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1550         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1551                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1552                     EXTENT_SET_PRIVATE2)))
1553                 return 0;
1554
1555         while (nr_pages > 0) {
1556                 ret = find_get_pages_contig(inode->i_mapping, index,
1557                                      min_t(unsigned long,
1558                                      nr_pages, ARRAY_SIZE(pages)), pages);
1559                 for (i = 0; i < ret; i++) {
1560
1561                         if (op & EXTENT_SET_PRIVATE2)
1562                                 SetPagePrivate2(pages[i]);
1563
1564                         if (pages[i] == locked_page) {
1565                                 page_cache_release(pages[i]);
1566                                 continue;
1567                         }
1568                         if (op & EXTENT_CLEAR_DIRTY)
1569                                 clear_page_dirty_for_io(pages[i]);
1570                         if (op & EXTENT_SET_WRITEBACK)
1571                                 set_page_writeback(pages[i]);
1572                         if (op & EXTENT_END_WRITEBACK)
1573                                 end_page_writeback(pages[i]);
1574                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1575                                 unlock_page(pages[i]);
1576                         page_cache_release(pages[i]);
1577                 }
1578                 nr_pages -= ret;
1579                 index += ret;
1580                 cond_resched();
1581         }
1582         return 0;
1583 }
1584
1585 /*
1586  * count the number of bytes in the tree that have a given bit(s)
1587  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1588  * cached.  The total number found is returned.
1589  */
1590 u64 count_range_bits(struct extent_io_tree *tree,
1591                      u64 *start, u64 search_end, u64 max_bytes,
1592                      unsigned long bits, int contig)
1593 {
1594         struct rb_node *node;
1595         struct extent_state *state;
1596         u64 cur_start = *start;
1597         u64 total_bytes = 0;
1598         u64 last = 0;
1599         int found = 0;
1600
1601         if (search_end <= cur_start) {
1602                 WARN_ON(1);
1603                 return 0;
1604         }
1605
1606         spin_lock(&tree->lock);
1607         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1608                 total_bytes = tree->dirty_bytes;
1609                 goto out;
1610         }
1611         /*
1612          * this search will find all the extents that end after
1613          * our range starts.
1614          */
1615         node = tree_search(tree, cur_start);
1616         if (!node)
1617                 goto out;
1618
1619         while (1) {
1620                 state = rb_entry(node, struct extent_state, rb_node);
1621                 if (state->start > search_end)
1622                         break;
1623                 if (contig && found && state->start > last + 1)
1624                         break;
1625                 if (state->end >= cur_start && (state->state & bits) == bits) {
1626                         total_bytes += min(search_end, state->end) + 1 -
1627                                        max(cur_start, state->start);
1628                         if (total_bytes >= max_bytes)
1629                                 break;
1630                         if (!found) {
1631                                 *start = max(cur_start, state->start);
1632                                 found = 1;
1633                         }
1634                         last = state->end;
1635                 } else if (contig && found) {
1636                         break;
1637                 }
1638                 node = rb_next(node);
1639                 if (!node)
1640                         break;
1641         }
1642 out:
1643         spin_unlock(&tree->lock);
1644         return total_bytes;
1645 }
1646
1647 /*
1648  * set the private field for a given byte offset in the tree.  If there isn't
1649  * an extent_state there already, this does nothing.
1650  */
1651 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1652 {
1653         struct rb_node *node;
1654         struct extent_state *state;
1655         int ret = 0;
1656
1657         spin_lock(&tree->lock);
1658         /*
1659          * this search will find all the extents that end after
1660          * our range starts.
1661          */
1662         node = tree_search(tree, start);
1663         if (!node) {
1664                 ret = -ENOENT;
1665                 goto out;
1666         }
1667         state = rb_entry(node, struct extent_state, rb_node);
1668         if (state->start != start) {
1669                 ret = -ENOENT;
1670                 goto out;
1671         }
1672         state->private = private;
1673 out:
1674         spin_unlock(&tree->lock);
1675         return ret;
1676 }
1677
1678 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1679 {
1680         struct rb_node *node;
1681         struct extent_state *state;
1682         int ret = 0;
1683
1684         spin_lock(&tree->lock);
1685         /*
1686          * this search will find all the extents that end after
1687          * our range starts.
1688          */
1689         node = tree_search(tree, start);
1690         if (!node) {
1691                 ret = -ENOENT;
1692                 goto out;
1693         }
1694         state = rb_entry(node, struct extent_state, rb_node);
1695         if (state->start != start) {
1696                 ret = -ENOENT;
1697                 goto out;
1698         }
1699         *private = state->private;
1700 out:
1701         spin_unlock(&tree->lock);
1702         return ret;
1703 }
1704
1705 /*
1706  * searches a range in the state tree for a given mask.
1707  * If 'filled' == 1, this returns 1 only if every extent in the tree
1708  * has the bits set.  Otherwise, 1 is returned if any bit in the
1709  * range is found set.
1710  */
1711 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1712                    int bits, int filled, struct extent_state *cached)
1713 {
1714         struct extent_state *state = NULL;
1715         struct rb_node *node;
1716         int bitset = 0;
1717
1718         spin_lock(&tree->lock);
1719         if (cached && cached->tree && cached->start <= start &&
1720             cached->end > start)
1721                 node = &cached->rb_node;
1722         else
1723                 node = tree_search(tree, start);
1724         while (node && start <= end) {
1725                 state = rb_entry(node, struct extent_state, rb_node);
1726
1727                 if (filled && state->start > start) {
1728                         bitset = 0;
1729                         break;
1730                 }
1731
1732                 if (state->start > end)
1733                         break;
1734
1735                 if (state->state & bits) {
1736                         bitset = 1;
1737                         if (!filled)
1738                                 break;
1739                 } else if (filled) {
1740                         bitset = 0;
1741                         break;
1742                 }
1743
1744                 if (state->end == (u64)-1)
1745                         break;
1746
1747                 start = state->end + 1;
1748                 if (start > end)
1749                         break;
1750                 node = rb_next(node);
1751                 if (!node) {
1752                         if (filled)
1753                                 bitset = 0;
1754                         break;
1755                 }
1756         }
1757         spin_unlock(&tree->lock);
1758         return bitset;
1759 }
1760
1761 /*
1762  * helper function to set a given page up to date if all the
1763  * extents in the tree for that page are up to date
1764  */
1765 static int check_page_uptodate(struct extent_io_tree *tree,
1766                                struct page *page)
1767 {
1768         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1769         u64 end = start + PAGE_CACHE_SIZE - 1;
1770         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1771                 SetPageUptodate(page);
1772         return 0;
1773 }
1774
1775 /*
1776  * helper function to unlock a page if all the extents in the tree
1777  * for that page are unlocked
1778  */
1779 static int check_page_locked(struct extent_io_tree *tree,
1780                              struct page *page)
1781 {
1782         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1783         u64 end = start + PAGE_CACHE_SIZE - 1;
1784         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1785                 unlock_page(page);
1786         return 0;
1787 }
1788
1789 /*
1790  * helper function to end page writeback if all the extents
1791  * in the tree for that page are done with writeback
1792  */
1793 static int check_page_writeback(struct extent_io_tree *tree,
1794                              struct page *page)
1795 {
1796         end_page_writeback(page);
1797         return 0;
1798 }
1799
1800 /*
1801  * When IO fails, either with EIO or csum verification fails, we
1802  * try other mirrors that might have a good copy of the data.  This
1803  * io_failure_record is used to record state as we go through all the
1804  * mirrors.  If another mirror has good data, the page is set up to date
1805  * and things continue.  If a good mirror can't be found, the original
1806  * bio end_io callback is called to indicate things have failed.
1807  */
1808 struct io_failure_record {
1809         struct page *page;
1810         u64 start;
1811         u64 len;
1812         u64 logical;
1813         unsigned long bio_flags;
1814         int this_mirror;
1815         int failed_mirror;
1816         int in_validation;
1817 };
1818
1819 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1820                                 int did_repair)
1821 {
1822         int ret;
1823         int err = 0;
1824         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1825
1826         set_state_private(failure_tree, rec->start, 0);
1827         ret = clear_extent_bits(failure_tree, rec->start,
1828                                 rec->start + rec->len - 1,
1829                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1830         if (ret)
1831                 err = ret;
1832
1833         if (did_repair) {
1834                 ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1835                                         rec->start + rec->len - 1,
1836                                         EXTENT_DAMAGED, GFP_NOFS);
1837                 if (ret && !err)
1838                         err = ret;
1839         }
1840
1841         kfree(rec);
1842         return err;
1843 }
1844
1845 static void repair_io_failure_callback(struct bio *bio, int err)
1846 {
1847         complete(bio->bi_private);
1848 }
1849
1850 /*
1851  * this bypasses the standard btrfs submit functions deliberately, as
1852  * the standard behavior is to write all copies in a raid setup. here we only
1853  * want to write the one bad copy. so we do the mapping for ourselves and issue
1854  * submit_bio directly.
1855  * to avoid any synchonization issues, wait for the data after writing, which
1856  * actually prevents the read that triggered the error from finishing.
1857  * currently, there can be no more than two copies of every data bit. thus,
1858  * exactly one rewrite is required.
1859  */
1860 int repair_io_failure(struct btrfs_mapping_tree *map_tree, u64 start,
1861                         u64 length, u64 logical, struct page *page,
1862                         int mirror_num)
1863 {
1864         struct bio *bio;
1865         struct btrfs_device *dev;
1866         DECLARE_COMPLETION_ONSTACK(compl);
1867         u64 map_length = 0;
1868         u64 sector;
1869         struct btrfs_bio *bbio = NULL;
1870         int ret;
1871
1872         BUG_ON(!mirror_num);
1873
1874         bio = bio_alloc(GFP_NOFS, 1);
1875         if (!bio)
1876                 return -EIO;
1877         bio->bi_private = &compl;
1878         bio->bi_end_io = repair_io_failure_callback;
1879         bio->bi_size = 0;
1880         map_length = length;
1881
1882         ret = btrfs_map_block(map_tree, WRITE, logical,
1883                               &map_length, &bbio, mirror_num);
1884         if (ret) {
1885                 bio_put(bio);
1886                 return -EIO;
1887         }
1888         BUG_ON(mirror_num != bbio->mirror_num);
1889         sector = bbio->stripes[mirror_num-1].physical >> 9;
1890         bio->bi_sector = sector;
1891         dev = bbio->stripes[mirror_num-1].dev;
1892         kfree(bbio);
1893         if (!dev || !dev->bdev || !dev->writeable) {
1894                 bio_put(bio);
1895                 return -EIO;
1896         }
1897         bio->bi_bdev = dev->bdev;
1898         bio_add_page(bio, page, length, start-page_offset(page));
1899         btrfsic_submit_bio(WRITE_SYNC, bio);
1900         wait_for_completion(&compl);
1901
1902         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
1903                 /* try to remap that extent elsewhere? */
1904                 bio_put(bio);
1905                 return -EIO;
1906         }
1907
1908         printk(KERN_INFO "btrfs read error corrected: ino %lu off %llu (dev %s "
1909                         "sector %llu)\n", page->mapping->host->i_ino, start,
1910                         dev->name, sector);
1911
1912         bio_put(bio);
1913         return 0;
1914 }
1915
1916 /*
1917  * each time an IO finishes, we do a fast check in the IO failure tree
1918  * to see if we need to process or clean up an io_failure_record
1919  */
1920 static int clean_io_failure(u64 start, struct page *page)
1921 {
1922         u64 private;
1923         u64 private_failure;
1924         struct io_failure_record *failrec;
1925         struct btrfs_mapping_tree *map_tree;
1926         struct extent_state *state;
1927         int num_copies;
1928         int did_repair = 0;
1929         int ret;
1930         struct inode *inode = page->mapping->host;
1931
1932         private = 0;
1933         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
1934                                 (u64)-1, 1, EXTENT_DIRTY, 0);
1935         if (!ret)
1936                 return 0;
1937
1938         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
1939                                 &private_failure);
1940         if (ret)
1941                 return 0;
1942
1943         failrec = (struct io_failure_record *)(unsigned long) private_failure;
1944         BUG_ON(!failrec->this_mirror);
1945
1946         if (failrec->in_validation) {
1947                 /* there was no real error, just free the record */
1948                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
1949                          failrec->start);
1950                 did_repair = 1;
1951                 goto out;
1952         }
1953
1954         spin_lock(&BTRFS_I(inode)->io_tree.lock);
1955         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
1956                                             failrec->start,
1957                                             EXTENT_LOCKED);
1958         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
1959
1960         if (state && state->start == failrec->start) {
1961                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
1962                 num_copies = btrfs_num_copies(map_tree, failrec->logical,
1963                                                 failrec->len);
1964                 if (num_copies > 1)  {
1965                         ret = repair_io_failure(map_tree, start, failrec->len,
1966                                                 failrec->logical, page,
1967                                                 failrec->failed_mirror);
1968                         did_repair = !ret;
1969                 }
1970         }
1971
1972 out:
1973         if (!ret)
1974                 ret = free_io_failure(inode, failrec, did_repair);
1975
1976         return ret;
1977 }
1978
1979 /*
1980  * this is a generic handler for readpage errors (default
1981  * readpage_io_failed_hook). if other copies exist, read those and write back
1982  * good data to the failed position. does not investigate in remapping the
1983  * failed extent elsewhere, hoping the device will be smart enough to do this as
1984  * needed
1985  */
1986
1987 static int bio_readpage_error(struct bio *failed_bio, struct page *page,
1988                                 u64 start, u64 end, int failed_mirror,
1989                                 struct extent_state *state)
1990 {
1991         struct io_failure_record *failrec = NULL;
1992         u64 private;
1993         struct extent_map *em;
1994         struct inode *inode = page->mapping->host;
1995         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1996         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1997         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
1998         struct bio *bio;
1999         int num_copies;
2000         int ret;
2001         int read_mode;
2002         u64 logical;
2003
2004         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2005
2006         ret = get_state_private(failure_tree, start, &private);
2007         if (ret) {
2008                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2009                 if (!failrec)
2010                         return -ENOMEM;
2011                 failrec->start = start;
2012                 failrec->len = end - start + 1;
2013                 failrec->this_mirror = 0;
2014                 failrec->bio_flags = 0;
2015                 failrec->in_validation = 0;
2016
2017                 read_lock(&em_tree->lock);
2018                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2019                 if (!em) {
2020                         read_unlock(&em_tree->lock);
2021                         kfree(failrec);
2022                         return -EIO;
2023                 }
2024
2025                 if (em->start > start || em->start + em->len < start) {
2026                         free_extent_map(em);
2027                         em = NULL;
2028                 }
2029                 read_unlock(&em_tree->lock);
2030
2031                 if (!em || IS_ERR(em)) {
2032                         kfree(failrec);
2033                         return -EIO;
2034                 }
2035                 logical = start - em->start;
2036                 logical = em->block_start + logical;
2037                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2038                         logical = em->block_start;
2039                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2040                         extent_set_compress_type(&failrec->bio_flags,
2041                                                  em->compress_type);
2042                 }
2043                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2044                          "len=%llu\n", logical, start, failrec->len);
2045                 failrec->logical = logical;
2046                 free_extent_map(em);
2047
2048                 /* set the bits in the private failure tree */
2049                 ret = set_extent_bits(failure_tree, start, end,
2050                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2051                 if (ret >= 0)
2052                         ret = set_state_private(failure_tree, start,
2053                                                 (u64)(unsigned long)failrec);
2054                 /* set the bits in the inode's tree */
2055                 if (ret >= 0)
2056                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2057                                                 GFP_NOFS);
2058                 if (ret < 0) {
2059                         kfree(failrec);
2060                         return ret;
2061                 }
2062         } else {
2063                 failrec = (struct io_failure_record *)(unsigned long)private;
2064                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2065                          "start=%llu, len=%llu, validation=%d\n",
2066                          failrec->logical, failrec->start, failrec->len,
2067                          failrec->in_validation);
2068                 /*
2069                  * when data can be on disk more than twice, add to failrec here
2070                  * (e.g. with a list for failed_mirror) to make
2071                  * clean_io_failure() clean all those errors at once.
2072                  */
2073         }
2074         num_copies = btrfs_num_copies(
2075                               &BTRFS_I(inode)->root->fs_info->mapping_tree,
2076                               failrec->logical, failrec->len);
2077         if (num_copies == 1) {
2078                 /*
2079                  * we only have a single copy of the data, so don't bother with
2080                  * all the retry and error correction code that follows. no
2081                  * matter what the error is, it is very likely to persist.
2082                  */
2083                 pr_debug("bio_readpage_error: cannot repair, num_copies == 1. "
2084                          "state=%p, num_copies=%d, next_mirror %d, "
2085                          "failed_mirror %d\n", state, num_copies,
2086                          failrec->this_mirror, failed_mirror);
2087                 free_io_failure(inode, failrec, 0);
2088                 return -EIO;
2089         }
2090
2091         if (!state) {
2092                 spin_lock(&tree->lock);
2093                 state = find_first_extent_bit_state(tree, failrec->start,
2094                                                     EXTENT_LOCKED);
2095                 if (state && state->start != failrec->start)
2096                         state = NULL;
2097                 spin_unlock(&tree->lock);
2098         }
2099
2100         /*
2101          * there are two premises:
2102          *      a) deliver good data to the caller
2103          *      b) correct the bad sectors on disk
2104          */
2105         if (failed_bio->bi_vcnt > 1) {
2106                 /*
2107                  * to fulfill b), we need to know the exact failing sectors, as
2108                  * we don't want to rewrite any more than the failed ones. thus,
2109                  * we need separate read requests for the failed bio
2110                  *
2111                  * if the following BUG_ON triggers, our validation request got
2112                  * merged. we need separate requests for our algorithm to work.
2113                  */
2114                 BUG_ON(failrec->in_validation);
2115                 failrec->in_validation = 1;
2116                 failrec->this_mirror = failed_mirror;
2117                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2118         } else {
2119                 /*
2120                  * we're ready to fulfill a) and b) alongside. get a good copy
2121                  * of the failed sector and if we succeed, we have setup
2122                  * everything for repair_io_failure to do the rest for us.
2123                  */
2124                 if (failrec->in_validation) {
2125                         BUG_ON(failrec->this_mirror != failed_mirror);
2126                         failrec->in_validation = 0;
2127                         failrec->this_mirror = 0;
2128                 }
2129                 failrec->failed_mirror = failed_mirror;
2130                 failrec->this_mirror++;
2131                 if (failrec->this_mirror == failed_mirror)
2132                         failrec->this_mirror++;
2133                 read_mode = READ_SYNC;
2134         }
2135
2136         if (!state || failrec->this_mirror > num_copies) {
2137                 pr_debug("bio_readpage_error: (fail) state=%p, num_copies=%d, "
2138                          "next_mirror %d, failed_mirror %d\n", state,
2139                          num_copies, failrec->this_mirror, failed_mirror);
2140                 free_io_failure(inode, failrec, 0);
2141                 return -EIO;
2142         }
2143
2144         bio = bio_alloc(GFP_NOFS, 1);
2145         bio->bi_private = state;
2146         bio->bi_end_io = failed_bio->bi_end_io;
2147         bio->bi_sector = failrec->logical >> 9;
2148         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2149         bio->bi_size = 0;
2150
2151         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2152
2153         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2154                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2155                  failrec->this_mirror, num_copies, failrec->in_validation);
2156
2157         tree->ops->submit_bio_hook(inode, read_mode, bio, failrec->this_mirror,
2158                                         failrec->bio_flags, 0);
2159         return 0;
2160 }
2161
2162 /* lots and lots of room for performance fixes in the end_bio funcs */
2163
2164 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2165 {
2166         int uptodate = (err == 0);
2167         struct extent_io_tree *tree;
2168         int ret;
2169
2170         tree = &BTRFS_I(page->mapping->host)->io_tree;
2171
2172         if (tree->ops && tree->ops->writepage_end_io_hook) {
2173                 ret = tree->ops->writepage_end_io_hook(page, start,
2174                                                end, NULL, uptodate);
2175                 if (ret)
2176                         uptodate = 0;
2177         }
2178
2179         if (!uptodate && tree->ops &&
2180             tree->ops->writepage_io_failed_hook) {
2181                 ret = tree->ops->writepage_io_failed_hook(NULL, page,
2182                                                  start, end, NULL);
2183                 /* Writeback already completed */
2184                 if (ret == 0)
2185                         return 1;
2186         }
2187
2188         if (!uptodate) {
2189                 clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
2190                 ClearPageUptodate(page);
2191                 SetPageError(page);
2192         }
2193         return 0;
2194 }
2195
2196 /*
2197  * after a writepage IO is done, we need to:
2198  * clear the uptodate bits on error
2199  * clear the writeback bits in the extent tree for this IO
2200  * end_page_writeback if the page has no more pending IO
2201  *
2202  * Scheduling is not allowed, so the extent state tree is expected
2203  * to have one and only one object corresponding to this IO.
2204  */
2205 static void end_bio_extent_writepage(struct bio *bio, int err)
2206 {
2207         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2208         struct extent_io_tree *tree;
2209         u64 start;
2210         u64 end;
2211         int whole_page;
2212
2213         do {
2214                 struct page *page = bvec->bv_page;
2215                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2216
2217                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2218                          bvec->bv_offset;
2219                 end = start + bvec->bv_len - 1;
2220
2221                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2222                         whole_page = 1;
2223                 else
2224                         whole_page = 0;
2225
2226                 if (--bvec >= bio->bi_io_vec)
2227                         prefetchw(&bvec->bv_page->flags);
2228
2229                 if (end_extent_writepage(page, err, start, end))
2230                         continue;
2231
2232                 if (whole_page)
2233                         end_page_writeback(page);
2234                 else
2235                         check_page_writeback(tree, page);
2236         } while (bvec >= bio->bi_io_vec);
2237
2238         bio_put(bio);
2239 }
2240
2241 /*
2242  * after a readpage IO is done, we need to:
2243  * clear the uptodate bits on error
2244  * set the uptodate bits if things worked
2245  * set the page up to date if all extents in the tree are uptodate
2246  * clear the lock bit in the extent tree
2247  * unlock the page if there are no other extents locked for it
2248  *
2249  * Scheduling is not allowed, so the extent state tree is expected
2250  * to have one and only one object corresponding to this IO.
2251  */
2252 static void end_bio_extent_readpage(struct bio *bio, int err)
2253 {
2254         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2255         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2256         struct bio_vec *bvec = bio->bi_io_vec;
2257         struct extent_io_tree *tree;
2258         u64 start;
2259         u64 end;
2260         int whole_page;
2261         int ret;
2262
2263         if (err)
2264                 uptodate = 0;
2265
2266         do {
2267                 struct page *page = bvec->bv_page;
2268                 struct extent_state *cached = NULL;
2269                 struct extent_state *state;
2270
2271                 pr_debug("end_bio_extent_readpage: bi_vcnt=%d, idx=%d, err=%d, "
2272                          "mirror=%ld\n", bio->bi_vcnt, bio->bi_idx, err,
2273                          (long int)bio->bi_bdev);
2274                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2275
2276                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2277                         bvec->bv_offset;
2278                 end = start + bvec->bv_len - 1;
2279
2280                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2281                         whole_page = 1;
2282                 else
2283                         whole_page = 0;
2284
2285                 if (++bvec <= bvec_end)
2286                         prefetchw(&bvec->bv_page->flags);
2287
2288                 spin_lock(&tree->lock);
2289                 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
2290                 if (state && state->start == start) {
2291                         /*
2292                          * take a reference on the state, unlock will drop
2293                          * the ref
2294                          */
2295                         cache_state(state, &cached);
2296                 }
2297                 spin_unlock(&tree->lock);
2298
2299                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
2300                         ret = tree->ops->readpage_end_io_hook(page, start, end,
2301                                                               state);
2302                         if (ret)
2303                                 uptodate = 0;
2304                         else
2305                                 clean_io_failure(start, page);
2306                 }
2307                 if (!uptodate) {
2308                         int failed_mirror;
2309                         failed_mirror = (int)(unsigned long)bio->bi_bdev;
2310                         /*
2311                          * The generic bio_readpage_error handles errors the
2312                          * following way: If possible, new read requests are
2313                          * created and submitted and will end up in
2314                          * end_bio_extent_readpage as well (if we're lucky, not
2315                          * in the !uptodate case). In that case it returns 0 and
2316                          * we just go on with the next page in our bio. If it
2317                          * can't handle the error it will return -EIO and we
2318                          * remain responsible for that page.
2319                          */
2320                         ret = bio_readpage_error(bio, page, start, end,
2321                                                         failed_mirror, NULL);
2322                         if (ret == 0) {
2323 error_handled:
2324                                 uptodate =
2325                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2326                                 if (err)
2327                                         uptodate = 0;
2328                                 uncache_state(&cached);
2329                                 continue;
2330                         }
2331                         if (tree->ops && tree->ops->readpage_io_failed_hook) {
2332                                 ret = tree->ops->readpage_io_failed_hook(
2333                                                         bio, page, start, end,
2334                                                         failed_mirror, state);
2335                                 if (ret == 0)
2336                                         goto error_handled;
2337                         }
2338                 }
2339
2340                 if (uptodate) {
2341                         set_extent_uptodate(tree, start, end, &cached,
2342                                             GFP_ATOMIC);
2343                 }
2344                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2345
2346                 if (whole_page) {
2347                         if (uptodate) {
2348                                 SetPageUptodate(page);
2349                         } else {
2350                                 ClearPageUptodate(page);
2351                                 SetPageError(page);
2352                         }
2353                         unlock_page(page);
2354                 } else {
2355                         if (uptodate) {
2356                                 check_page_uptodate(tree, page);
2357                         } else {
2358                                 ClearPageUptodate(page);
2359                                 SetPageError(page);
2360                         }
2361                         check_page_locked(tree, page);
2362                 }
2363         } while (bvec <= bvec_end);
2364
2365         bio_put(bio);
2366 }
2367
2368 struct bio *
2369 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2370                 gfp_t gfp_flags)
2371 {
2372         struct bio *bio;
2373
2374         bio = bio_alloc(gfp_flags, nr_vecs);
2375
2376         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2377                 while (!bio && (nr_vecs /= 2))
2378                         bio = bio_alloc(gfp_flags, nr_vecs);
2379         }
2380
2381         if (bio) {
2382                 bio->bi_size = 0;
2383                 bio->bi_bdev = bdev;
2384                 bio->bi_sector = first_sector;
2385         }
2386         return bio;
2387 }
2388
2389 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
2390                           unsigned long bio_flags)
2391 {
2392         int ret = 0;
2393         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2394         struct page *page = bvec->bv_page;
2395         struct extent_io_tree *tree = bio->bi_private;
2396         u64 start;
2397
2398         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
2399
2400         bio->bi_private = NULL;
2401
2402         bio_get(bio);
2403
2404         if (tree->ops && tree->ops->submit_bio_hook)
2405                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2406                                            mirror_num, bio_flags, start);
2407         else
2408                 btrfsic_submit_bio(rw, bio);
2409
2410         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2411                 ret = -EOPNOTSUPP;
2412         bio_put(bio);
2413         return ret;
2414 }
2415
2416 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2417                               struct page *page, sector_t sector,
2418                               size_t size, unsigned long offset,
2419                               struct block_device *bdev,
2420                               struct bio **bio_ret,
2421                               unsigned long max_pages,
2422                               bio_end_io_t end_io_func,
2423                               int mirror_num,
2424                               unsigned long prev_bio_flags,
2425                               unsigned long bio_flags)
2426 {
2427         int ret = 0;
2428         struct bio *bio;
2429         int nr;
2430         int contig = 0;
2431         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2432         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2433         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2434
2435         if (bio_ret && *bio_ret) {
2436                 bio = *bio_ret;
2437                 if (old_compressed)
2438                         contig = bio->bi_sector == sector;
2439                 else
2440                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
2441                                 sector;
2442
2443                 if (prev_bio_flags != bio_flags || !contig ||
2444                     (tree->ops && tree->ops->merge_bio_hook &&
2445                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
2446                                                bio_flags)) ||
2447                     bio_add_page(bio, page, page_size, offset) < page_size) {
2448                         ret = submit_one_bio(rw, bio, mirror_num,
2449                                              prev_bio_flags);
2450                         bio = NULL;
2451                 } else {
2452                         return 0;
2453                 }
2454         }
2455         if (this_compressed)
2456                 nr = BIO_MAX_PAGES;
2457         else
2458                 nr = bio_get_nr_vecs(bdev);
2459
2460         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2461         if (!bio)
2462                 return -ENOMEM;
2463
2464         bio_add_page(bio, page, page_size, offset);
2465         bio->bi_end_io = end_io_func;
2466         bio->bi_private = tree;
2467
2468         if (bio_ret)
2469                 *bio_ret = bio;
2470         else
2471                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2472
2473         return ret;
2474 }
2475
2476 void set_page_extent_mapped(struct page *page)
2477 {
2478         if (!PagePrivate(page)) {
2479                 SetPagePrivate(page);
2480                 page_cache_get(page);
2481                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2482         }
2483 }
2484
2485 static void set_page_extent_head(struct page *page, unsigned long len)
2486 {
2487         WARN_ON(!PagePrivate(page));
2488         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
2489 }
2490
2491 /*
2492  * basic readpage implementation.  Locked extent state structs are inserted
2493  * into the tree that are removed when the IO is done (by the end_io
2494  * handlers)
2495  */
2496 static int __extent_read_full_page(struct extent_io_tree *tree,
2497                                    struct page *page,
2498                                    get_extent_t *get_extent,
2499                                    struct bio **bio, int mirror_num,
2500                                    unsigned long *bio_flags)
2501 {
2502         struct inode *inode = page->mapping->host;
2503         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2504         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2505         u64 end;
2506         u64 cur = start;
2507         u64 extent_offset;
2508         u64 last_byte = i_size_read(inode);
2509         u64 block_start;
2510         u64 cur_end;
2511         sector_t sector;
2512         struct extent_map *em;
2513         struct block_device *bdev;
2514         struct btrfs_ordered_extent *ordered;
2515         int ret;
2516         int nr = 0;
2517         size_t pg_offset = 0;
2518         size_t iosize;
2519         size_t disk_io_size;
2520         size_t blocksize = inode->i_sb->s_blocksize;
2521         unsigned long this_bio_flag = 0;
2522
2523         set_page_extent_mapped(page);
2524
2525         if (!PageUptodate(page)) {
2526                 if (cleancache_get_page(page) == 0) {
2527                         BUG_ON(blocksize != PAGE_SIZE);
2528                         goto out;
2529                 }
2530         }
2531
2532         end = page_end;
2533         while (1) {
2534                 lock_extent(tree, start, end, GFP_NOFS);
2535                 ordered = btrfs_lookup_ordered_extent(inode, start);
2536                 if (!ordered)
2537                         break;
2538                 unlock_extent(tree, start, end, GFP_NOFS);
2539                 btrfs_start_ordered_extent(inode, ordered, 1);
2540                 btrfs_put_ordered_extent(ordered);
2541         }
2542
2543         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2544                 char *userpage;
2545                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2546
2547                 if (zero_offset) {
2548                         iosize = PAGE_CACHE_SIZE - zero_offset;
2549                         userpage = kmap_atomic(page, KM_USER0);
2550                         memset(userpage + zero_offset, 0, iosize);
2551                         flush_dcache_page(page);
2552                         kunmap_atomic(userpage, KM_USER0);
2553                 }
2554         }
2555         while (cur <= end) {
2556                 if (cur >= last_byte) {
2557                         char *userpage;
2558                         struct extent_state *cached = NULL;
2559
2560                         iosize = PAGE_CACHE_SIZE - pg_offset;
2561                         userpage = kmap_atomic(page, KM_USER0);
2562                         memset(userpage + pg_offset, 0, iosize);
2563                         flush_dcache_page(page);
2564                         kunmap_atomic(userpage, KM_USER0);
2565                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2566                                             &cached, GFP_NOFS);
2567                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2568                                              &cached, GFP_NOFS);
2569                         break;
2570                 }
2571                 em = get_extent(inode, page, pg_offset, cur,
2572                                 end - cur + 1, 0);
2573                 if (IS_ERR_OR_NULL(em)) {
2574                         SetPageError(page);
2575                         unlock_extent(tree, cur, end, GFP_NOFS);
2576                         break;
2577                 }
2578                 extent_offset = cur - em->start;
2579                 BUG_ON(extent_map_end(em) <= cur);
2580                 BUG_ON(end < cur);
2581
2582                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2583                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2584                         extent_set_compress_type(&this_bio_flag,
2585                                                  em->compress_type);
2586                 }
2587
2588                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2589                 cur_end = min(extent_map_end(em) - 1, end);
2590                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2591                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2592                         disk_io_size = em->block_len;
2593                         sector = em->block_start >> 9;
2594                 } else {
2595                         sector = (em->block_start + extent_offset) >> 9;
2596                         disk_io_size = iosize;
2597                 }
2598                 bdev = em->bdev;
2599                 block_start = em->block_start;
2600                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2601                         block_start = EXTENT_MAP_HOLE;
2602                 free_extent_map(em);
2603                 em = NULL;
2604
2605                 /* we've found a hole, just zero and go on */
2606                 if (block_start == EXTENT_MAP_HOLE) {
2607                         char *userpage;
2608                         struct extent_state *cached = NULL;
2609
2610                         userpage = kmap_atomic(page, KM_USER0);
2611                         memset(userpage + pg_offset, 0, iosize);
2612                         flush_dcache_page(page);
2613                         kunmap_atomic(userpage, KM_USER0);
2614
2615                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2616                                             &cached, GFP_NOFS);
2617                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2618                                              &cached, GFP_NOFS);
2619                         cur = cur + iosize;
2620                         pg_offset += iosize;
2621                         continue;
2622                 }
2623                 /* the get_extent function already copied into the page */
2624                 if (test_range_bit(tree, cur, cur_end,
2625                                    EXTENT_UPTODATE, 1, NULL)) {
2626                         check_page_uptodate(tree, page);
2627                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2628                         cur = cur + iosize;
2629                         pg_offset += iosize;
2630                         continue;
2631                 }
2632                 /* we have an inline extent but it didn't get marked up
2633                  * to date.  Error out
2634                  */
2635                 if (block_start == EXTENT_MAP_INLINE) {
2636                         SetPageError(page);
2637                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2638                         cur = cur + iosize;
2639                         pg_offset += iosize;
2640                         continue;
2641                 }
2642
2643                 ret = 0;
2644                 if (tree->ops && tree->ops->readpage_io_hook) {
2645                         ret = tree->ops->readpage_io_hook(page, cur,
2646                                                           cur + iosize - 1);
2647                 }
2648                 if (!ret) {
2649                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2650                         pnr -= page->index;
2651                         ret = submit_extent_page(READ, tree, page,
2652                                          sector, disk_io_size, pg_offset,
2653                                          bdev, bio, pnr,
2654                                          end_bio_extent_readpage, mirror_num,
2655                                          *bio_flags,
2656                                          this_bio_flag);
2657                         nr++;
2658                         *bio_flags = this_bio_flag;
2659                 }
2660                 if (ret)
2661                         SetPageError(page);
2662                 cur = cur + iosize;
2663                 pg_offset += iosize;
2664         }
2665 out:
2666         if (!nr) {
2667                 if (!PageError(page))
2668                         SetPageUptodate(page);
2669                 unlock_page(page);
2670         }
2671         return 0;
2672 }
2673
2674 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2675                             get_extent_t *get_extent, int mirror_num)
2676 {
2677         struct bio *bio = NULL;
2678         unsigned long bio_flags = 0;
2679         int ret;
2680
2681         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
2682                                       &bio_flags);
2683         if (bio)
2684                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
2685         return ret;
2686 }
2687
2688 static noinline void update_nr_written(struct page *page,
2689                                       struct writeback_control *wbc,
2690                                       unsigned long nr_written)
2691 {
2692         wbc->nr_to_write -= nr_written;
2693         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2694             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2695                 page->mapping->writeback_index = page->index + nr_written;
2696 }
2697
2698 /*
2699  * the writepage semantics are similar to regular writepage.  extent
2700  * records are inserted to lock ranges in the tree, and as dirty areas
2701  * are found, they are marked writeback.  Then the lock bits are removed
2702  * and the end_io handler clears the writeback ranges
2703  */
2704 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2705                               void *data)
2706 {
2707         struct inode *inode = page->mapping->host;
2708         struct extent_page_data *epd = data;
2709         struct extent_io_tree *tree = epd->tree;
2710         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2711         u64 delalloc_start;
2712         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2713         u64 end;
2714         u64 cur = start;
2715         u64 extent_offset;
2716         u64 last_byte = i_size_read(inode);
2717         u64 block_start;
2718         u64 iosize;
2719         sector_t sector;
2720         struct extent_state *cached_state = NULL;
2721         struct extent_map *em;
2722         struct block_device *bdev;
2723         int ret;
2724         int nr = 0;
2725         size_t pg_offset = 0;
2726         size_t blocksize;
2727         loff_t i_size = i_size_read(inode);
2728         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2729         u64 nr_delalloc;
2730         u64 delalloc_end;
2731         int page_started;
2732         int compressed;
2733         int write_flags;
2734         unsigned long nr_written = 0;
2735         bool fill_delalloc = true;
2736
2737         if (wbc->sync_mode == WB_SYNC_ALL)
2738                 write_flags = WRITE_SYNC;
2739         else
2740                 write_flags = WRITE;
2741
2742         trace___extent_writepage(page, inode, wbc);
2743
2744         WARN_ON(!PageLocked(page));
2745
2746         ClearPageError(page);
2747
2748         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2749         if (page->index > end_index ||
2750            (page->index == end_index && !pg_offset)) {
2751                 page->mapping->a_ops->invalidatepage(page, 0);
2752                 unlock_page(page);
2753                 return 0;
2754         }
2755
2756         if (page->index == end_index) {
2757                 char *userpage;
2758
2759                 userpage = kmap_atomic(page, KM_USER0);
2760                 memset(userpage + pg_offset, 0,
2761                        PAGE_CACHE_SIZE - pg_offset);
2762                 kunmap_atomic(userpage, KM_USER0);
2763                 flush_dcache_page(page);
2764         }
2765         pg_offset = 0;
2766
2767         set_page_extent_mapped(page);
2768
2769         if (!tree->ops || !tree->ops->fill_delalloc)
2770                 fill_delalloc = false;
2771
2772         delalloc_start = start;
2773         delalloc_end = 0;
2774         page_started = 0;
2775         if (!epd->extent_locked && fill_delalloc) {
2776                 u64 delalloc_to_write = 0;
2777                 /*
2778                  * make sure the wbc mapping index is at least updated
2779                  * to this page.
2780                  */
2781                 update_nr_written(page, wbc, 0);
2782
2783                 while (delalloc_end < page_end) {
2784                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2785                                                        page,
2786                                                        &delalloc_start,
2787                                                        &delalloc_end,
2788                                                        128 * 1024 * 1024);
2789                         if (nr_delalloc == 0) {
2790                                 delalloc_start = delalloc_end + 1;
2791                                 continue;
2792                         }
2793                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2794                                                  delalloc_end, &page_started,
2795                                                  &nr_written);
2796                         /*
2797                          * delalloc_end is already one less than the total
2798                          * length, so we don't subtract one from
2799                          * PAGE_CACHE_SIZE
2800                          */
2801                         delalloc_to_write += (delalloc_end - delalloc_start +
2802                                               PAGE_CACHE_SIZE) >>
2803                                               PAGE_CACHE_SHIFT;
2804                         delalloc_start = delalloc_end + 1;
2805                 }
2806                 if (wbc->nr_to_write < delalloc_to_write) {
2807                         int thresh = 8192;
2808
2809                         if (delalloc_to_write < thresh * 2)
2810                                 thresh = delalloc_to_write;
2811                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2812                                                  thresh);
2813                 }
2814
2815                 /* did the fill delalloc function already unlock and start
2816                  * the IO?
2817                  */
2818                 if (page_started) {
2819                         ret = 0;
2820                         /*
2821                          * we've unlocked the page, so we can't update
2822                          * the mapping's writeback index, just update
2823                          * nr_to_write.
2824                          */
2825                         wbc->nr_to_write -= nr_written;
2826                         goto done_unlocked;
2827                 }
2828         }
2829         if (tree->ops && tree->ops->writepage_start_hook) {
2830                 ret = tree->ops->writepage_start_hook(page, start,
2831                                                       page_end);
2832                 if (ret) {
2833                         /* Fixup worker will requeue */
2834                         if (ret == -EBUSY)
2835                                 wbc->pages_skipped++;
2836                         else
2837                                 redirty_page_for_writepage(wbc, page);
2838                         update_nr_written(page, wbc, nr_written);
2839                         unlock_page(page);
2840                         ret = 0;
2841                         goto done_unlocked;
2842                 }
2843         }
2844
2845         /*
2846          * we don't want to touch the inode after unlocking the page,
2847          * so we update the mapping writeback index now
2848          */
2849         update_nr_written(page, wbc, nr_written + 1);
2850
2851         end = page_end;
2852         if (last_byte <= start) {
2853                 if (tree->ops && tree->ops->writepage_end_io_hook)
2854                         tree->ops->writepage_end_io_hook(page, start,
2855                                                          page_end, NULL, 1);
2856                 goto done;
2857         }
2858
2859         blocksize = inode->i_sb->s_blocksize;
2860
2861         while (cur <= end) {
2862                 if (cur >= last_byte) {
2863                         if (tree->ops && tree->ops->writepage_end_io_hook)
2864                                 tree->ops->writepage_end_io_hook(page, cur,
2865                                                          page_end, NULL, 1);
2866                         break;
2867                 }
2868                 em = epd->get_extent(inode, page, pg_offset, cur,
2869                                      end - cur + 1, 1);
2870                 if (IS_ERR_OR_NULL(em)) {
2871                         SetPageError(page);
2872                         break;
2873                 }
2874
2875                 extent_offset = cur - em->start;
2876                 BUG_ON(extent_map_end(em) <= cur);
2877                 BUG_ON(end < cur);
2878                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2879                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2880                 sector = (em->block_start + extent_offset) >> 9;
2881                 bdev = em->bdev;
2882                 block_start = em->block_start;
2883                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2884                 free_extent_map(em);
2885                 em = NULL;
2886
2887                 /*
2888                  * compressed and inline extents are written through other
2889                  * paths in the FS
2890                  */
2891                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2892                     block_start == EXTENT_MAP_INLINE) {
2893                         /*
2894                          * end_io notification does not happen here for
2895                          * compressed extents
2896                          */
2897                         if (!compressed && tree->ops &&
2898                             tree->ops->writepage_end_io_hook)
2899                                 tree->ops->writepage_end_io_hook(page, cur,
2900                                                          cur + iosize - 1,
2901                                                          NULL, 1);
2902                         else if (compressed) {
2903                                 /* we don't want to end_page_writeback on
2904                                  * a compressed extent.  this happens
2905                                  * elsewhere
2906                                  */
2907                                 nr++;
2908                         }
2909
2910                         cur += iosize;
2911                         pg_offset += iosize;
2912                         continue;
2913                 }
2914                 /* leave this out until we have a page_mkwrite call */
2915                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2916                                    EXTENT_DIRTY, 0, NULL)) {
2917                         cur = cur + iosize;
2918                         pg_offset += iosize;
2919                         continue;
2920                 }
2921
2922                 if (tree->ops && tree->ops->writepage_io_hook) {
2923                         ret = tree->ops->writepage_io_hook(page, cur,
2924                                                 cur + iosize - 1);
2925                 } else {
2926                         ret = 0;
2927                 }
2928                 if (ret) {
2929                         SetPageError(page);
2930                 } else {
2931                         unsigned long max_nr = end_index + 1;
2932
2933                         set_range_writeback(tree, cur, cur + iosize - 1);
2934                         if (!PageWriteback(page)) {
2935                                 printk(KERN_ERR "btrfs warning page %lu not "
2936                                        "writeback, cur %llu end %llu\n",
2937                                        page->index, (unsigned long long)cur,
2938                                        (unsigned long long)end);
2939                         }
2940
2941                         ret = submit_extent_page(write_flags, tree, page,
2942                                                  sector, iosize, pg_offset,
2943                                                  bdev, &epd->bio, max_nr,
2944                                                  end_bio_extent_writepage,
2945                                                  0, 0, 0);
2946                         if (ret)
2947                                 SetPageError(page);
2948                 }
2949                 cur = cur + iosize;
2950                 pg_offset += iosize;
2951                 nr++;
2952         }
2953 done:
2954         if (nr == 0) {
2955                 /* make sure the mapping tag for page dirty gets cleared */
2956                 set_page_writeback(page);
2957                 end_page_writeback(page);
2958         }
2959         unlock_page(page);
2960
2961 done_unlocked:
2962
2963         /* drop our reference on any cached states */
2964         free_extent_state(cached_state);
2965         return 0;
2966 }
2967
2968 /**
2969  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2970  * @mapping: address space structure to write
2971  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2972  * @writepage: function called for each page
2973  * @data: data passed to writepage function
2974  *
2975  * If a page is already under I/O, write_cache_pages() skips it, even
2976  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2977  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2978  * and msync() need to guarantee that all the data which was dirty at the time
2979  * the call was made get new I/O started against them.  If wbc->sync_mode is
2980  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2981  * existing IO to complete.
2982  */
2983 static int extent_write_cache_pages(struct extent_io_tree *tree,
2984                              struct address_space *mapping,
2985                              struct writeback_control *wbc,
2986                              writepage_t writepage, void *data,
2987                              void (*flush_fn)(void *))
2988 {
2989         int ret = 0;
2990         int done = 0;
2991         int nr_to_write_done = 0;
2992         struct pagevec pvec;
2993         int nr_pages;
2994         pgoff_t index;
2995         pgoff_t end;            /* Inclusive */
2996         int scanned = 0;
2997         int tag;
2998
2999         pagevec_init(&pvec, 0);
3000         if (wbc->range_cyclic) {
3001                 index = mapping->writeback_index; /* Start from prev offset */
3002                 end = -1;
3003         } else {
3004                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3005                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3006                 scanned = 1;
3007         }
3008         if (wbc->sync_mode == WB_SYNC_ALL)
3009                 tag = PAGECACHE_TAG_TOWRITE;
3010         else
3011                 tag = PAGECACHE_TAG_DIRTY;
3012 retry:
3013         if (wbc->sync_mode == WB_SYNC_ALL)
3014                 tag_pages_for_writeback(mapping, index, end);
3015         while (!done && !nr_to_write_done && (index <= end) &&
3016                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3017                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3018                 unsigned i;
3019
3020                 scanned = 1;
3021                 for (i = 0; i < nr_pages; i++) {
3022                         struct page *page = pvec.pages[i];
3023
3024                         /*
3025                          * At this point we hold neither mapping->tree_lock nor
3026                          * lock on the page itself: the page may be truncated or
3027                          * invalidated (changing page->mapping to NULL), or even
3028                          * swizzled back from swapper_space to tmpfs file
3029                          * mapping
3030                          */
3031                         if (tree->ops &&
3032                             tree->ops->write_cache_pages_lock_hook) {
3033                                 tree->ops->write_cache_pages_lock_hook(page,
3034                                                                data, flush_fn);
3035                         } else {
3036                                 if (!trylock_page(page)) {
3037                                         flush_fn(data);
3038                                         lock_page(page);
3039                                 }
3040                         }
3041
3042                         if (unlikely(page->mapping != mapping)) {
3043                                 unlock_page(page);
3044                                 continue;
3045                         }
3046
3047                         if (!wbc->range_cyclic && page->index > end) {
3048                                 done = 1;
3049                                 unlock_page(page);
3050                                 continue;
3051                         }
3052
3053                         if (wbc->sync_mode != WB_SYNC_NONE) {
3054                                 if (PageWriteback(page))
3055                                         flush_fn(data);
3056                                 wait_on_page_writeback(page);
3057                         }
3058
3059                         if (PageWriteback(page) ||
3060                             !clear_page_dirty_for_io(page)) {
3061                                 unlock_page(page);
3062                                 continue;
3063                         }
3064
3065                         ret = (*writepage)(page, wbc, data);
3066
3067                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3068                                 unlock_page(page);
3069                                 ret = 0;
3070                         }
3071                         if (ret)
3072                                 done = 1;
3073
3074                         /*
3075                          * the filesystem may choose to bump up nr_to_write.
3076                          * We have to make sure to honor the new nr_to_write
3077                          * at any time
3078                          */
3079                         nr_to_write_done = wbc->nr_to_write <= 0;
3080                 }
3081                 pagevec_release(&pvec);
3082                 cond_resched();
3083         }
3084         if (!scanned && !done) {
3085                 /*
3086                  * We hit the last page and there is more work to be done: wrap
3087                  * back to the start of the file
3088                  */
3089                 scanned = 1;
3090                 index = 0;
3091                 goto retry;
3092         }
3093         return ret;
3094 }
3095
3096 static void flush_epd_write_bio(struct extent_page_data *epd)
3097 {
3098         if (epd->bio) {
3099                 if (epd->sync_io)
3100                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
3101                 else
3102                         submit_one_bio(WRITE, epd->bio, 0, 0);
3103                 epd->bio = NULL;
3104         }
3105 }
3106
3107 static noinline void flush_write_bio(void *data)
3108 {
3109         struct extent_page_data *epd = data;
3110         flush_epd_write_bio(epd);
3111 }
3112
3113 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3114                           get_extent_t *get_extent,
3115                           struct writeback_control *wbc)
3116 {
3117         int ret;
3118         struct extent_page_data epd = {
3119                 .bio = NULL,
3120                 .tree = tree,
3121                 .get_extent = get_extent,
3122                 .extent_locked = 0,
3123                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3124         };
3125
3126         ret = __extent_writepage(page, wbc, &epd);
3127
3128         flush_epd_write_bio(&epd);
3129         return ret;
3130 }
3131
3132 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3133                               u64 start, u64 end, get_extent_t *get_extent,
3134                               int mode)
3135 {
3136         int ret = 0;
3137         struct address_space *mapping = inode->i_mapping;
3138         struct page *page;
3139         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3140                 PAGE_CACHE_SHIFT;
3141
3142         struct extent_page_data epd = {
3143                 .bio = NULL,
3144                 .tree = tree,
3145                 .get_extent = get_extent,
3146                 .extent_locked = 1,
3147                 .sync_io = mode == WB_SYNC_ALL,
3148         };
3149         struct writeback_control wbc_writepages = {
3150                 .sync_mode      = mode,
3151                 .nr_to_write    = nr_pages * 2,
3152                 .range_start    = start,
3153                 .range_end      = end + 1,
3154         };
3155
3156         while (start <= end) {
3157                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3158                 if (clear_page_dirty_for_io(page))
3159                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3160                 else {
3161                         if (tree->ops && tree->ops->writepage_end_io_hook)
3162                                 tree->ops->writepage_end_io_hook(page, start,
3163                                                  start + PAGE_CACHE_SIZE - 1,
3164                                                  NULL, 1);
3165                         unlock_page(page);
3166                 }
3167                 page_cache_release(page);
3168                 start += PAGE_CACHE_SIZE;
3169         }
3170
3171         flush_epd_write_bio(&epd);
3172         return ret;
3173 }
3174
3175 int extent_writepages(struct extent_io_tree *tree,
3176                       struct address_space *mapping,
3177                       get_extent_t *get_extent,
3178                       struct writeback_control *wbc)
3179 {
3180         int ret = 0;
3181         struct extent_page_data epd = {
3182                 .bio = NULL,
3183                 .tree = tree,
3184                 .get_extent = get_extent,
3185                 .extent_locked = 0,
3186                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3187         };
3188
3189         ret = extent_write_cache_pages(tree, mapping, wbc,
3190                                        __extent_writepage, &epd,
3191                                        flush_write_bio);
3192         flush_epd_write_bio(&epd);
3193         return ret;
3194 }
3195
3196 int extent_readpages(struct extent_io_tree *tree,
3197                      struct address_space *mapping,
3198                      struct list_head *pages, unsigned nr_pages,
3199                      get_extent_t get_extent)
3200 {
3201         struct bio *bio = NULL;
3202         unsigned page_idx;
3203         unsigned long bio_flags = 0;
3204
3205         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3206                 struct page *page = list_entry(pages->prev, struct page, lru);
3207
3208                 prefetchw(&page->flags);
3209                 list_del(&page->lru);
3210                 if (!add_to_page_cache_lru(page, mapping,
3211                                         page->index, GFP_NOFS)) {
3212                         __extent_read_full_page(tree, page, get_extent,
3213                                                 &bio, 0, &bio_flags);
3214                 }
3215                 page_cache_release(page);
3216         }
3217         BUG_ON(!list_empty(pages));
3218         if (bio)
3219                 submit_one_bio(READ, bio, 0, bio_flags);
3220         return 0;
3221 }
3222
3223 /*
3224  * basic invalidatepage code, this waits on any locked or writeback
3225  * ranges corresponding to the page, and then deletes any extent state
3226  * records from the tree
3227  */
3228 int extent_invalidatepage(struct extent_io_tree *tree,
3229                           struct page *page, unsigned long offset)
3230 {
3231         struct extent_state *cached_state = NULL;
3232         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
3233         u64 end = start + PAGE_CACHE_SIZE - 1;
3234         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3235
3236         start += (offset + blocksize - 1) & ~(blocksize - 1);
3237         if (start > end)
3238                 return 0;
3239
3240         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
3241         wait_on_page_writeback(page);
3242         clear_extent_bit(tree, start, end,
3243                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3244                          EXTENT_DO_ACCOUNTING,
3245                          1, 1, &cached_state, GFP_NOFS);
3246         return 0;
3247 }
3248
3249 /*
3250  * a helper for releasepage, this tests for areas of the page that
3251  * are locked or under IO and drops the related state bits if it is safe
3252  * to drop the page.
3253  */
3254 int try_release_extent_state(struct extent_map_tree *map,
3255                              struct extent_io_tree *tree, struct page *page,
3256                              gfp_t mask)
3257 {
3258         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3259         u64 end = start + PAGE_CACHE_SIZE - 1;
3260         int ret = 1;
3261
3262         if (test_range_bit(tree, start, end,
3263                            EXTENT_IOBITS, 0, NULL))
3264                 ret = 0;
3265         else {
3266                 if ((mask & GFP_NOFS) == GFP_NOFS)
3267                         mask = GFP_NOFS;
3268                 /*
3269                  * at this point we can safely clear everything except the
3270                  * locked bit and the nodatasum bit
3271                  */
3272                 ret = clear_extent_bit(tree, start, end,
3273                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3274                                  0, 0, NULL, mask);
3275
3276                 /* if clear_extent_bit failed for enomem reasons,
3277                  * we can't allow the release to continue.
3278                  */
3279                 if (ret < 0)
3280                         ret = 0;
3281                 else
3282                         ret = 1;
3283         }
3284         return ret;
3285 }
3286
3287 /*
3288  * a helper for releasepage.  As long as there are no locked extents
3289  * in the range corresponding to the page, both state records and extent
3290  * map records are removed
3291  */
3292 int try_release_extent_mapping(struct extent_map_tree *map,
3293                                struct extent_io_tree *tree, struct page *page,
3294                                gfp_t mask)
3295 {
3296         struct extent_map *em;
3297         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3298         u64 end = start + PAGE_CACHE_SIZE - 1;
3299
3300         if ((mask & __GFP_WAIT) &&
3301             page->mapping->host->i_size > 16 * 1024 * 1024) {
3302                 u64 len;
3303                 while (start <= end) {
3304                         len = end - start + 1;
3305                         write_lock(&map->lock);
3306                         em = lookup_extent_mapping(map, start, len);
3307                         if (IS_ERR_OR_NULL(em)) {
3308                                 write_unlock(&map->lock);
3309                                 break;
3310                         }
3311                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3312                             em->start != start) {
3313                                 write_unlock(&map->lock);
3314                                 free_extent_map(em);
3315                                 break;
3316                         }
3317                         if (!test_range_bit(tree, em->start,
3318                                             extent_map_end(em) - 1,
3319                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
3320                                             0, NULL)) {
3321                                 remove_extent_mapping(map, em);
3322                                 /* once for the rb tree */
3323                                 free_extent_map(em);
3324                         }
3325                         start = extent_map_end(em);
3326                         write_unlock(&map->lock);
3327
3328                         /* once for us */
3329                         free_extent_map(em);
3330                 }
3331         }
3332         return try_release_extent_state(map, tree, page, mask);
3333 }
3334
3335 /*
3336  * helper function for fiemap, which doesn't want to see any holes.
3337  * This maps until we find something past 'last'
3338  */
3339 static struct extent_map *get_extent_skip_holes(struct inode *inode,
3340                                                 u64 offset,
3341                                                 u64 last,
3342                                                 get_extent_t *get_extent)
3343 {
3344         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
3345         struct extent_map *em;
3346         u64 len;
3347
3348         if (offset >= last)
3349                 return NULL;
3350
3351         while(1) {
3352                 len = last - offset;
3353                 if (len == 0)
3354                         break;
3355                 len = (len + sectorsize - 1) & ~(sectorsize - 1);
3356                 em = get_extent(inode, NULL, 0, offset, len, 0);
3357                 if (IS_ERR_OR_NULL(em))
3358                         return em;
3359
3360                 /* if this isn't a hole return it */
3361                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
3362                     em->block_start != EXTENT_MAP_HOLE) {
3363                         return em;
3364                 }
3365
3366                 /* this is a hole, advance to the next extent */
3367                 offset = extent_map_end(em);
3368                 free_extent_map(em);
3369                 if (offset >= last)
3370                         break;
3371         }
3372         return NULL;
3373 }
3374
3375 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
3376                 __u64 start, __u64 len, get_extent_t *get_extent)
3377 {
3378         int ret = 0;
3379         u64 off = start;
3380         u64 max = start + len;
3381         u32 flags = 0;
3382         u32 found_type;
3383         u64 last;
3384         u64 last_for_get_extent = 0;
3385         u64 disko = 0;
3386         u64 isize = i_size_read(inode);
3387         struct btrfs_key found_key;
3388         struct extent_map *em = NULL;
3389         struct extent_state *cached_state = NULL;
3390         struct btrfs_path *path;
3391         struct btrfs_file_extent_item *item;
3392         int end = 0;
3393         u64 em_start = 0;
3394         u64 em_len = 0;
3395         u64 em_end = 0;
3396         unsigned long emflags;
3397
3398         if (len == 0)
3399                 return -EINVAL;
3400
3401         path = btrfs_alloc_path();
3402         if (!path)
3403                 return -ENOMEM;
3404         path->leave_spinning = 1;
3405
3406         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
3407         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
3408
3409         /*
3410          * lookup the last file extent.  We're not using i_size here
3411          * because there might be preallocation past i_size
3412          */
3413         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3414                                        path, btrfs_ino(inode), -1, 0);
3415         if (ret < 0) {
3416                 btrfs_free_path(path);
3417                 return ret;
3418         }
3419         WARN_ON(!ret);
3420         path->slots[0]--;
3421         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3422                               struct btrfs_file_extent_item);
3423         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3424         found_type = btrfs_key_type(&found_key);
3425
3426         /* No extents, but there might be delalloc bits */
3427         if (found_key.objectid != btrfs_ino(inode) ||
3428             found_type != BTRFS_EXTENT_DATA_KEY) {
3429                 /* have to trust i_size as the end */
3430                 last = (u64)-1;
3431                 last_for_get_extent = isize;
3432         } else {
3433                 /*
3434                  * remember the start of the last extent.  There are a
3435                  * bunch of different factors that go into the length of the
3436                  * extent, so its much less complex to remember where it started
3437                  */
3438                 last = found_key.offset;
3439                 last_for_get_extent = last + 1;
3440         }
3441         btrfs_free_path(path);
3442
3443         /*
3444          * we might have some extents allocated but more delalloc past those
3445          * extents.  so, we trust isize unless the start of the last extent is
3446          * beyond isize
3447          */
3448         if (last < isize) {
3449                 last = (u64)-1;
3450                 last_for_get_extent = isize;
3451         }
3452
3453         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3454                          &cached_state, GFP_NOFS);
3455
3456         em = get_extent_skip_holes(inode, start, last_for_get_extent,
3457                                    get_extent);
3458         if (!em)
3459                 goto out;
3460         if (IS_ERR(em)) {
3461                 ret = PTR_ERR(em);
3462                 goto out;
3463         }
3464
3465         while (!end) {
3466                 u64 offset_in_extent;
3467
3468                 /* break if the extent we found is outside the range */
3469                 if (em->start >= max || extent_map_end(em) < off)
3470                         break;
3471
3472                 /*
3473                  * get_extent may return an extent that starts before our
3474                  * requested range.  We have to make sure the ranges
3475                  * we return to fiemap always move forward and don't
3476                  * overlap, so adjust the offsets here
3477                  */
3478                 em_start = max(em->start, off);
3479
3480                 /*
3481                  * record the offset from the start of the extent
3482                  * for adjusting the disk offset below
3483                  */
3484                 offset_in_extent = em_start - em->start;
3485                 em_end = extent_map_end(em);
3486                 em_len = em_end - em_start;
3487                 emflags = em->flags;
3488                 disko = 0;
3489                 flags = 0;
3490
3491                 /*
3492                  * bump off for our next call to get_extent
3493                  */
3494                 off = extent_map_end(em);
3495                 if (off >= max)
3496                         end = 1;
3497
3498                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3499                         end = 1;
3500                         flags |= FIEMAP_EXTENT_LAST;
3501                 } else if (em->block_start == EXTENT_MAP_INLINE) {
3502                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
3503                                   FIEMAP_EXTENT_NOT_ALIGNED);
3504                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3505                         flags |= (FIEMAP_EXTENT_DELALLOC |
3506                                   FIEMAP_EXTENT_UNKNOWN);
3507                 } else {
3508                         disko = em->block_start + offset_in_extent;
3509                 }
3510                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3511                         flags |= FIEMAP_EXTENT_ENCODED;
3512
3513                 free_extent_map(em);
3514                 em = NULL;
3515                 if ((em_start >= last) || em_len == (u64)-1 ||
3516                    (last == (u64)-1 && isize <= em_end)) {
3517                         flags |= FIEMAP_EXTENT_LAST;
3518                         end = 1;
3519                 }
3520
3521                 /* now scan forward to see if this is really the last extent. */
3522                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
3523                                            get_extent);
3524                 if (IS_ERR(em)) {
3525                         ret = PTR_ERR(em);
3526                         goto out;
3527                 }
3528                 if (!em) {
3529                         flags |= FIEMAP_EXTENT_LAST;
3530                         end = 1;
3531                 }
3532                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3533                                               em_len, flags);
3534                 if (ret)
3535                         goto out_free;
3536         }
3537 out_free:
3538         free_extent_map(em);
3539 out:
3540         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3541                              &cached_state, GFP_NOFS);
3542         return ret;
3543 }
3544
3545 inline struct page *extent_buffer_page(struct extent_buffer *eb,
3546                                               unsigned long i)
3547 {
3548         struct page *p;
3549         struct address_space *mapping;
3550
3551         if (i == 0)
3552                 return eb->first_page;
3553         i += eb->start >> PAGE_CACHE_SHIFT;
3554         mapping = eb->first_page->mapping;
3555         if (!mapping)
3556                 return NULL;
3557
3558         /*
3559          * extent_buffer_page is only called after pinning the page
3560          * by increasing the reference count.  So we know the page must
3561          * be in the radix tree.
3562          */
3563         rcu_read_lock();
3564         p = radix_tree_lookup(&mapping->page_tree, i);
3565         rcu_read_unlock();
3566
3567         return p;
3568 }
3569
3570 inline unsigned long num_extent_pages(u64 start, u64 len)
3571 {
3572         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3573                 (start >> PAGE_CACHE_SHIFT);
3574 }
3575
3576 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3577                                                    u64 start,
3578                                                    unsigned long len,
3579                                                    gfp_t mask)
3580 {
3581         struct extent_buffer *eb = NULL;
3582 #if LEAK_DEBUG
3583         unsigned long flags;
3584 #endif
3585
3586         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3587         if (eb == NULL)
3588                 return NULL;
3589         eb->start = start;
3590         eb->len = len;
3591         rwlock_init(&eb->lock);
3592         atomic_set(&eb->write_locks, 0);
3593         atomic_set(&eb->read_locks, 0);
3594         atomic_set(&eb->blocking_readers, 0);
3595         atomic_set(&eb->blocking_writers, 0);
3596         atomic_set(&eb->spinning_readers, 0);
3597         atomic_set(&eb->spinning_writers, 0);
3598         eb->lock_nested = 0;
3599         init_waitqueue_head(&eb->write_lock_wq);
3600         init_waitqueue_head(&eb->read_lock_wq);
3601
3602 #if LEAK_DEBUG
3603         spin_lock_irqsave(&leak_lock, flags);
3604         list_add(&eb->leak_list, &buffers);
3605         spin_unlock_irqrestore(&leak_lock, flags);
3606 #endif
3607         atomic_set(&eb->refs, 1);
3608
3609         return eb;
3610 }
3611
3612 static void __free_extent_buffer(struct extent_buffer *eb)
3613 {
3614 #if LEAK_DEBUG
3615         unsigned long flags;
3616         spin_lock_irqsave(&leak_lock, flags);
3617         list_del(&eb->leak_list);
3618         spin_unlock_irqrestore(&leak_lock, flags);
3619 #endif
3620         kmem_cache_free(extent_buffer_cache, eb);
3621 }
3622
3623 /*
3624  * Helper for releasing extent buffer page.
3625  */
3626 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3627                                                 unsigned long start_idx)
3628 {
3629         unsigned long index;
3630         struct page *page;
3631
3632         if (!eb->first_page)
3633                 return;
3634
3635         index = num_extent_pages(eb->start, eb->len);
3636         if (start_idx >= index)
3637                 return;
3638
3639         do {
3640                 index--;
3641                 page = extent_buffer_page(eb, index);
3642                 if (page)
3643                         page_cache_release(page);
3644         } while (index != start_idx);
3645 }
3646
3647 /*
3648  * Helper for releasing the extent buffer.
3649  */
3650 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3651 {
3652         btrfs_release_extent_buffer_page(eb, 0);
3653         __free_extent_buffer(eb);
3654 }
3655
3656 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3657                                           u64 start, unsigned long len,
3658                                           struct page *page0)
3659 {
3660         unsigned long num_pages = num_extent_pages(start, len);
3661         unsigned long i;
3662         unsigned long index = start >> PAGE_CACHE_SHIFT;
3663         struct extent_buffer *eb;
3664         struct extent_buffer *exists = NULL;
3665         struct page *p;
3666         struct address_space *mapping = tree->mapping;
3667         int uptodate = 1;
3668         int ret;
3669
3670         rcu_read_lock();
3671         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3672         if (eb && atomic_inc_not_zero(&eb->refs)) {
3673                 rcu_read_unlock();
3674                 mark_page_accessed(eb->first_page);
3675                 return eb;
3676         }
3677         rcu_read_unlock();
3678
3679         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
3680         if (!eb)
3681                 return NULL;
3682
3683         if (page0) {
3684                 eb->first_page = page0;
3685                 i = 1;
3686                 index++;
3687                 page_cache_get(page0);
3688                 mark_page_accessed(page0);
3689                 set_page_extent_mapped(page0);
3690                 set_page_extent_head(page0, len);
3691                 uptodate = PageUptodate(page0);
3692         } else {
3693                 i = 0;
3694         }
3695         for (; i < num_pages; i++, index++) {
3696                 p = find_or_create_page(mapping, index, GFP_NOFS);
3697                 if (!p) {
3698                         WARN_ON(1);
3699                         goto free_eb;
3700                 }
3701                 set_page_extent_mapped(p);
3702                 mark_page_accessed(p);
3703                 if (i == 0) {
3704                         eb->first_page = p;
3705                         set_page_extent_head(p, len);
3706                 } else {
3707                         set_page_private(p, EXTENT_PAGE_PRIVATE);
3708                 }
3709                 if (!PageUptodate(p))
3710                         uptodate = 0;
3711
3712                 /*
3713                  * see below about how we avoid a nasty race with release page
3714                  * and why we unlock later
3715                  */
3716                 if (i != 0)
3717                         unlock_page(p);
3718         }
3719         if (uptodate)
3720                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3721
3722         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
3723         if (ret)
3724                 goto free_eb;
3725
3726         spin_lock(&tree->buffer_lock);
3727         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
3728         if (ret == -EEXIST) {
3729                 exists = radix_tree_lookup(&tree->buffer,
3730                                                 start >> PAGE_CACHE_SHIFT);
3731                 /* add one reference for the caller */
3732                 atomic_inc(&exists->refs);
3733                 spin_unlock(&tree->buffer_lock);
3734                 radix_tree_preload_end();
3735                 goto free_eb;
3736         }
3737         /* add one reference for the tree */
3738         atomic_inc(&eb->refs);
3739         spin_unlock(&tree->buffer_lock);
3740         radix_tree_preload_end();
3741
3742         /*
3743          * there is a race where release page may have
3744          * tried to find this extent buffer in the radix
3745          * but failed.  It will tell the VM it is safe to
3746          * reclaim the, and it will clear the page private bit.
3747          * We must make sure to set the page private bit properly
3748          * after the extent buffer is in the radix tree so
3749          * it doesn't get lost
3750          */
3751         set_page_extent_mapped(eb->first_page);
3752         set_page_extent_head(eb->first_page, eb->len);
3753         if (!page0)
3754                 unlock_page(eb->first_page);
3755         return eb;
3756
3757 free_eb:
3758         if (eb->first_page && !page0)
3759                 unlock_page(eb->first_page);
3760
3761         if (!atomic_dec_and_test(&eb->refs))
3762                 return exists;
3763         btrfs_release_extent_buffer(eb);
3764         return exists;
3765 }
3766
3767 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3768                                          u64 start, unsigned long len)
3769 {
3770         struct extent_buffer *eb;
3771
3772         rcu_read_lock();
3773         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3774         if (eb && atomic_inc_not_zero(&eb->refs)) {
3775                 rcu_read_unlock();
3776                 mark_page_accessed(eb->first_page);
3777                 return eb;
3778         }
3779         rcu_read_unlock();
3780
3781         return NULL;
3782 }
3783
3784 void free_extent_buffer(struct extent_buffer *eb)
3785 {
3786         if (!eb)
3787                 return;
3788
3789         if (!atomic_dec_and_test(&eb->refs))
3790                 return;
3791
3792         WARN_ON(1);
3793 }
3794
3795 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3796                               struct extent_buffer *eb)
3797 {
3798         unsigned long i;
3799         unsigned long num_pages;
3800         struct page *page;
3801
3802         num_pages = num_extent_pages(eb->start, eb->len);
3803
3804         for (i = 0; i < num_pages; i++) {
3805                 page = extent_buffer_page(eb, i);
3806                 if (!PageDirty(page))
3807                         continue;
3808
3809                 lock_page(page);
3810                 WARN_ON(!PagePrivate(page));
3811
3812                 set_page_extent_mapped(page);
3813                 if (i == 0)
3814                         set_page_extent_head(page, eb->len);
3815
3816                 clear_page_dirty_for_io(page);
3817                 spin_lock_irq(&page->mapping->tree_lock);
3818                 if (!PageDirty(page)) {
3819                         radix_tree_tag_clear(&page->mapping->page_tree,
3820                                                 page_index(page),
3821                                                 PAGECACHE_TAG_DIRTY);
3822                 }
3823                 spin_unlock_irq(&page->mapping->tree_lock);
3824                 ClearPageError(page);
3825                 unlock_page(page);
3826         }
3827         return 0;
3828 }
3829
3830 int set_extent_buffer_dirty(struct extent_io_tree *tree,
3831                              struct extent_buffer *eb)
3832 {
3833         unsigned long i;
3834         unsigned long num_pages;
3835         int was_dirty = 0;
3836
3837         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3838         num_pages = num_extent_pages(eb->start, eb->len);
3839         for (i = 0; i < num_pages; i++)
3840                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3841         return was_dirty;
3842 }
3843
3844 static int __eb_straddles_pages(u64 start, u64 len)
3845 {
3846         if (len < PAGE_CACHE_SIZE)
3847                 return 1;
3848         if (start & (PAGE_CACHE_SIZE - 1))
3849                 return 1;
3850         if ((start + len) & (PAGE_CACHE_SIZE - 1))
3851                 return 1;
3852         return 0;
3853 }
3854
3855 static int eb_straddles_pages(struct extent_buffer *eb)
3856 {
3857         return __eb_straddles_pages(eb->start, eb->len);
3858 }
3859
3860 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3861                                 struct extent_buffer *eb,
3862                                 struct extent_state **cached_state)
3863 {
3864         unsigned long i;
3865         struct page *page;
3866         unsigned long num_pages;
3867
3868         num_pages = num_extent_pages(eb->start, eb->len);
3869         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3870
3871         if (eb_straddles_pages(eb)) {
3872                 clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3873                                       cached_state, GFP_NOFS);
3874         }
3875         for (i = 0; i < num_pages; i++) {
3876                 page = extent_buffer_page(eb, i);
3877                 if (page)
3878                         ClearPageUptodate(page);
3879         }
3880         return 0;
3881 }
3882
3883 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3884                                 struct extent_buffer *eb)
3885 {
3886         unsigned long i;
3887         struct page *page;
3888         unsigned long num_pages;
3889
3890         num_pages = num_extent_pages(eb->start, eb->len);
3891
3892         if (eb_straddles_pages(eb)) {
3893                 set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3894                                     NULL, GFP_NOFS);
3895         }
3896         for (i = 0; i < num_pages; i++) {
3897                 page = extent_buffer_page(eb, i);
3898                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3899                     ((i == num_pages - 1) &&
3900                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3901                         check_page_uptodate(tree, page);
3902                         continue;
3903                 }
3904                 SetPageUptodate(page);
3905         }
3906         return 0;
3907 }
3908
3909 int extent_range_uptodate(struct extent_io_tree *tree,
3910                           u64 start, u64 end)
3911 {
3912         struct page *page;
3913         int ret;
3914         int pg_uptodate = 1;
3915         int uptodate;
3916         unsigned long index;
3917
3918         if (__eb_straddles_pages(start, end - start + 1)) {
3919                 ret = test_range_bit(tree, start, end,
3920                                      EXTENT_UPTODATE, 1, NULL);
3921                 if (ret)
3922                         return 1;
3923         }
3924         while (start <= end) {
3925                 index = start >> PAGE_CACHE_SHIFT;
3926                 page = find_get_page(tree->mapping, index);
3927                 if (!page)
3928                         return 1;
3929                 uptodate = PageUptodate(page);
3930                 page_cache_release(page);
3931                 if (!uptodate) {
3932                         pg_uptodate = 0;
3933                         break;
3934                 }
3935                 start += PAGE_CACHE_SIZE;
3936         }
3937         return pg_uptodate;
3938 }
3939
3940 int extent_buffer_uptodate(struct extent_io_tree *tree,
3941                            struct extent_buffer *eb,
3942                            struct extent_state *cached_state)
3943 {
3944         int ret = 0;
3945         unsigned long num_pages;
3946         unsigned long i;
3947         struct page *page;
3948         int pg_uptodate = 1;
3949
3950         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3951                 return 1;
3952
3953         if (eb_straddles_pages(eb)) {
3954                 ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3955                                    EXTENT_UPTODATE, 1, cached_state);
3956                 if (ret)
3957                         return ret;
3958         }
3959
3960         num_pages = num_extent_pages(eb->start, eb->len);
3961         for (i = 0; i < num_pages; i++) {
3962                 page = extent_buffer_page(eb, i);
3963                 if (!PageUptodate(page)) {
3964                         pg_uptodate = 0;
3965                         break;
3966                 }
3967         }
3968         return pg_uptodate;
3969 }
3970
3971 int read_extent_buffer_pages(struct extent_io_tree *tree,
3972                              struct extent_buffer *eb, u64 start, int wait,
3973                              get_extent_t *get_extent, int mirror_num)
3974 {
3975         unsigned long i;
3976         unsigned long start_i;
3977         struct page *page;
3978         int err;
3979         int ret = 0;
3980         int locked_pages = 0;
3981         int all_uptodate = 1;
3982         int inc_all_pages = 0;
3983         unsigned long num_pages;
3984         struct bio *bio = NULL;
3985         unsigned long bio_flags = 0;
3986
3987         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3988                 return 0;
3989
3990         if (eb_straddles_pages(eb)) {
3991                 if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3992                                    EXTENT_UPTODATE, 1, NULL)) {
3993                         return 0;
3994                 }
3995         }
3996
3997         if (start) {
3998                 WARN_ON(start < eb->start);
3999                 start_i = (start >> PAGE_CACHE_SHIFT) -
4000                         (eb->start >> PAGE_CACHE_SHIFT);
4001         } else {
4002                 start_i = 0;
4003         }
4004
4005         num_pages = num_extent_pages(eb->start, eb->len);
4006         for (i = start_i; i < num_pages; i++) {
4007                 page = extent_buffer_page(eb, i);
4008                 if (wait == WAIT_NONE) {
4009                         if (!trylock_page(page))
4010                                 goto unlock_exit;
4011                 } else {
4012                         lock_page(page);
4013                 }
4014                 locked_pages++;
4015                 if (!PageUptodate(page))
4016                         all_uptodate = 0;
4017         }
4018         if (all_uptodate) {
4019                 if (start_i == 0)
4020                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4021                 goto unlock_exit;
4022         }
4023
4024         for (i = start_i; i < num_pages; i++) {
4025                 page = extent_buffer_page(eb, i);
4026
4027                 WARN_ON(!PagePrivate(page));
4028
4029                 set_page_extent_mapped(page);
4030                 if (i == 0)
4031                         set_page_extent_head(page, eb->len);
4032
4033                 if (inc_all_pages)
4034                         page_cache_get(page);
4035                 if (!PageUptodate(page)) {
4036                         if (start_i == 0)
4037                                 inc_all_pages = 1;
4038                         ClearPageError(page);
4039                         err = __extent_read_full_page(tree, page,
4040                                                       get_extent, &bio,
4041                                                       mirror_num, &bio_flags);
4042                         if (err)
4043                                 ret = err;
4044                 } else {
4045                         unlock_page(page);
4046                 }
4047         }
4048
4049         if (bio)
4050                 submit_one_bio(READ, bio, mirror_num, bio_flags);
4051
4052         if (ret || wait != WAIT_COMPLETE)
4053                 return ret;
4054
4055         for (i = start_i; i < num_pages; i++) {
4056                 page = extent_buffer_page(eb, i);
4057                 wait_on_page_locked(page);
4058                 if (!PageUptodate(page))
4059                         ret = -EIO;
4060         }
4061
4062         if (!ret)
4063                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4064         return ret;
4065
4066 unlock_exit:
4067         i = start_i;
4068         while (locked_pages > 0) {
4069                 page = extent_buffer_page(eb, i);
4070                 i++;
4071                 unlock_page(page);
4072                 locked_pages--;
4073         }
4074         return ret;
4075 }
4076
4077 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4078                         unsigned long start,
4079                         unsigned long len)
4080 {
4081         size_t cur;
4082         size_t offset;
4083         struct page *page;
4084         char *kaddr;
4085         char *dst = (char *)dstv;
4086         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4087         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4088
4089         WARN_ON(start > eb->len);
4090         WARN_ON(start + len > eb->start + eb->len);
4091
4092         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4093
4094         while (len > 0) {
4095                 page = extent_buffer_page(eb, i);
4096
4097                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4098                 kaddr = page_address(page);
4099                 memcpy(dst, kaddr + offset, cur);
4100
4101                 dst += cur;
4102                 len -= cur;
4103                 offset = 0;
4104                 i++;
4105         }
4106 }
4107
4108 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4109                                unsigned long min_len, char **map,
4110                                unsigned long *map_start,
4111                                unsigned long *map_len)
4112 {
4113         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4114         char *kaddr;
4115         struct page *p;
4116         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4117         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4118         unsigned long end_i = (start_offset + start + min_len - 1) >>
4119                 PAGE_CACHE_SHIFT;
4120
4121         if (i != end_i)
4122                 return -EINVAL;
4123
4124         if (i == 0) {
4125                 offset = start_offset;
4126                 *map_start = 0;
4127         } else {
4128                 offset = 0;
4129                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4130         }
4131
4132         if (start + min_len > eb->len) {
4133                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4134                        "wanted %lu %lu\n", (unsigned long long)eb->start,
4135                        eb->len, start, min_len);
4136                 WARN_ON(1);
4137                 return -EINVAL;
4138         }
4139
4140         p = extent_buffer_page(eb, i);
4141         kaddr = page_address(p);
4142         *map = kaddr + offset;
4143         *map_len = PAGE_CACHE_SIZE - offset;
4144         return 0;
4145 }
4146
4147 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4148                           unsigned long start,
4149                           unsigned long len)
4150 {
4151         size_t cur;
4152         size_t offset;
4153         struct page *page;
4154         char *kaddr;
4155         char *ptr = (char *)ptrv;
4156         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4157         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4158         int ret = 0;
4159
4160         WARN_ON(start > eb->len);
4161         WARN_ON(start + len > eb->start + eb->len);
4162
4163         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4164
4165         while (len > 0) {
4166                 page = extent_buffer_page(eb, i);
4167
4168                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4169
4170                 kaddr = page_address(page);
4171                 ret = memcmp(ptr, kaddr + offset, cur);
4172                 if (ret)
4173                         break;
4174
4175                 ptr += cur;
4176                 len -= cur;
4177                 offset = 0;
4178                 i++;
4179         }
4180         return ret;
4181 }
4182
4183 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4184                          unsigned long start, unsigned long len)
4185 {
4186         size_t cur;
4187         size_t offset;
4188         struct page *page;
4189         char *kaddr;
4190         char *src = (char *)srcv;
4191         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4192         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4193
4194         WARN_ON(start > eb->len);
4195         WARN_ON(start + len > eb->start + eb->len);
4196
4197         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4198
4199         while (len > 0) {
4200                 page = extent_buffer_page(eb, i);
4201                 WARN_ON(!PageUptodate(page));
4202
4203                 cur = min(len, PAGE_CACHE_SIZE - offset);
4204                 kaddr = page_address(page);
4205                 memcpy(kaddr + offset, src, cur);
4206
4207                 src += cur;
4208                 len -= cur;
4209                 offset = 0;
4210                 i++;
4211         }
4212 }
4213
4214 void memset_extent_buffer(struct extent_buffer *eb, char c,
4215                           unsigned long start, unsigned long len)
4216 {
4217         size_t cur;
4218         size_t offset;
4219         struct page *page;
4220         char *kaddr;
4221         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4222         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4223
4224         WARN_ON(start > eb->len);
4225         WARN_ON(start + len > eb->start + eb->len);
4226
4227         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4228
4229         while (len > 0) {
4230                 page = extent_buffer_page(eb, i);
4231                 WARN_ON(!PageUptodate(page));
4232
4233                 cur = min(len, PAGE_CACHE_SIZE - offset);
4234                 kaddr = page_address(page);
4235                 memset(kaddr + offset, c, cur);
4236
4237                 len -= cur;
4238                 offset = 0;
4239                 i++;
4240         }
4241 }
4242
4243 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
4244                         unsigned long dst_offset, unsigned long src_offset,
4245                         unsigned long len)
4246 {
4247         u64 dst_len = dst->len;
4248         size_t cur;
4249         size_t offset;
4250         struct page *page;
4251         char *kaddr;
4252         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4253         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4254
4255         WARN_ON(src->len != dst_len);
4256
4257         offset = (start_offset + dst_offset) &
4258                 ((unsigned long)PAGE_CACHE_SIZE - 1);
4259
4260         while (len > 0) {
4261                 page = extent_buffer_page(dst, i);
4262                 WARN_ON(!PageUptodate(page));
4263
4264                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
4265
4266                 kaddr = page_address(page);
4267                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4268
4269                 src_offset += cur;
4270                 len -= cur;
4271                 offset = 0;
4272                 i++;
4273         }
4274 }
4275
4276 static void move_pages(struct page *dst_page, struct page *src_page,
4277                        unsigned long dst_off, unsigned long src_off,
4278                        unsigned long len)
4279 {
4280         char *dst_kaddr = page_address(dst_page);
4281         if (dst_page == src_page) {
4282                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
4283         } else {
4284                 char *src_kaddr = page_address(src_page);
4285                 char *p = dst_kaddr + dst_off + len;
4286                 char *s = src_kaddr + src_off + len;
4287
4288                 while (len--)
4289                         *--p = *--s;
4290         }
4291 }
4292
4293 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4294 {
4295         unsigned long distance = (src > dst) ? src - dst : dst - src;
4296         return distance < len;
4297 }
4298
4299 static void copy_pages(struct page *dst_page, struct page *src_page,
4300                        unsigned long dst_off, unsigned long src_off,
4301                        unsigned long len)
4302 {
4303         char *dst_kaddr = page_address(dst_page);
4304         char *src_kaddr;
4305
4306         if (dst_page != src_page) {
4307                 src_kaddr = page_address(src_page);
4308         } else {
4309                 src_kaddr = dst_kaddr;
4310                 BUG_ON(areas_overlap(src_off, dst_off, len));
4311         }
4312
4313         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
4314 }
4315
4316 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4317                            unsigned long src_offset, unsigned long len)
4318 {
4319         size_t cur;
4320         size_t dst_off_in_page;
4321         size_t src_off_in_page;
4322         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4323         unsigned long dst_i;
4324         unsigned long src_i;
4325
4326         if (src_offset + len > dst->len) {
4327                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4328                        "len %lu dst len %lu\n", src_offset, len, dst->len);
4329                 BUG_ON(1);
4330         }
4331         if (dst_offset + len > dst->len) {
4332                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4333                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
4334                 BUG_ON(1);
4335         }
4336
4337         while (len > 0) {
4338                 dst_off_in_page = (start_offset + dst_offset) &
4339                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4340                 src_off_in_page = (start_offset + src_offset) &
4341                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4342
4343                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4344                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
4345
4346                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
4347                                                src_off_in_page));
4348                 cur = min_t(unsigned long, cur,
4349                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
4350
4351                 copy_pages(extent_buffer_page(dst, dst_i),
4352                            extent_buffer_page(dst, src_i),
4353                            dst_off_in_page, src_off_in_page, cur);
4354
4355                 src_offset += cur;
4356                 dst_offset += cur;
4357                 len -= cur;
4358         }
4359 }
4360
4361 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4362                            unsigned long src_offset, unsigned long len)
4363 {
4364         size_t cur;
4365         size_t dst_off_in_page;
4366         size_t src_off_in_page;
4367         unsigned long dst_end = dst_offset + len - 1;
4368         unsigned long src_end = src_offset + len - 1;
4369         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4370         unsigned long dst_i;
4371         unsigned long src_i;
4372
4373         if (src_offset + len > dst->len) {
4374                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4375                        "len %lu len %lu\n", src_offset, len, dst->len);
4376                 BUG_ON(1);
4377         }
4378         if (dst_offset + len > dst->len) {
4379                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4380                        "len %lu len %lu\n", dst_offset, len, dst->len);
4381                 BUG_ON(1);
4382         }
4383         if (!areas_overlap(src_offset, dst_offset, len)) {
4384                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4385                 return;
4386         }
4387         while (len > 0) {
4388                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4389                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4390
4391                 dst_off_in_page = (start_offset + dst_end) &
4392                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4393                 src_off_in_page = (start_offset + src_end) &
4394                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4395
4396                 cur = min_t(unsigned long, len, src_off_in_page + 1);
4397                 cur = min(cur, dst_off_in_page + 1);
4398                 move_pages(extent_buffer_page(dst, dst_i),
4399                            extent_buffer_page(dst, src_i),
4400                            dst_off_in_page - cur + 1,
4401                            src_off_in_page - cur + 1, cur);
4402
4403                 dst_end -= cur;
4404                 src_end -= cur;
4405                 len -= cur;
4406         }
4407 }
4408
4409 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4410 {
4411         struct extent_buffer *eb =
4412                         container_of(head, struct extent_buffer, rcu_head);
4413
4414         btrfs_release_extent_buffer(eb);
4415 }
4416
4417 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
4418 {
4419         u64 start = page_offset(page);
4420         struct extent_buffer *eb;
4421         int ret = 1;
4422
4423         spin_lock(&tree->buffer_lock);
4424         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4425         if (!eb) {
4426                 spin_unlock(&tree->buffer_lock);
4427                 return ret;
4428         }
4429
4430         if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
4431                 ret = 0;
4432                 goto out;
4433         }
4434
4435         /*
4436          * set @eb->refs to 0 if it is already 1, and then release the @eb.
4437          * Or go back.
4438          */
4439         if (atomic_cmpxchg(&eb->refs, 1, 0) != 1) {
4440                 ret = 0;
4441                 goto out;
4442         }
4443
4444         radix_tree_delete(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4445 out:
4446         spin_unlock(&tree->buffer_lock);
4447
4448         /* at this point we can safely release the extent buffer */
4449         if (atomic_read(&eb->refs) == 0)
4450                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4451         return ret;
4452 }