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         if (state->end < end && !need_resched())
 517                 next_node = rb_next(&state->rb_node);
 518         else
 519                 next_node = NULL;
 520
 521         /* the state doesn't have the wanted bits, go ahead */
 522         if (!(state->state & bits))
 523                 goto next;
 524
 525         /*
 526          *     | ---- desired range ---- |
 527          *  | state | or
 528          *  | ------------- state -------------- |
 529          *
 530          * We need to split the extent we found, and may flip
 531          * bits on second half.
 532          *
 533          * If the extent we found extends past our range, we
 534          * just split and search again.  It'll get split again
 535          * the next time though.
 536          *
 537          * If the extent we found is inside our range, we clear
 538          * the desired bit on it.
 539          */
 540
 541         if (state->start < start) {
 542                 prealloc = alloc_extent_state_atomic(prealloc);
 543                 BUG_ON(!prealloc);
 544                 err = split_state(tree, state, prealloc, start);
 545                 BUG_ON(err == -EEXIST);
 546                 prealloc = NULL;
 547                 if (err)
 548                         goto out;
 549                 if (state->end <= end) {
 550                         set |= clear_state_bit(tree, state, &bits, wake);
 551                         if (last_end == (u64)-1)
 552                                 goto out;
 553                         start = last_end + 1;
 554                 }
 555                 goto search_again;
 556         }
 557         /*
 558          * | ---- desired range ---- |
 559          *                        | state |
 560          * We need to split the extent, and clear the bit
 561          * on the first half
 562          */
 563         if (state->start <= end && state->end > end) {
 564                 prealloc = alloc_extent_state_atomic(prealloc);
 565                 BUG_ON(!prealloc);
 566                 err = split_state(tree, state, prealloc, end + 1);
 567                 BUG_ON(err == -EEXIST);
 568                 if (wake)
 569                         wake_up(&state->wq);
 570
 571                 set |= clear_state_bit(tree, prealloc, &bits, wake);
 572
 573                 prealloc = NULL;
 574                 goto out;
 575         }
 576
 577         set |= clear_state_bit(tree, state, &bits, wake);
 578 next:
 579         if (last_end == (u64)-1)
 580                 goto out;
 581         start = last_end + 1;
 582         if (start <= end && next_node) {
 583                 state = rb_entry(next_node, struct extent_state,
 584                                  rb_node);
 585                 goto hit_next;
 586         }
 587         goto search_again;
 588
 589 out:
 590         spin_unlock(&tree->lock);
 591         if (prealloc)
 592                 free_extent_state(prealloc);
 593
 594         return set;
 595
 596 search_again:
 597         if (start > end)
 598                 goto out;
 599         spin_unlock(&tree->lock);
 600         if (mask & __GFP_WAIT)
 601                 cond_resched();
 602         goto again;
 603 }
 604
 605 static int wait_on_state(struct extent_io_tree *tree,
 606                          struct extent_state *state)
 607                 __releases(tree->lock)
 608                 __acquires(tree->lock)
 609 {
 610         DEFINE_WAIT(wait);
 611         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 612         spin_unlock(&tree->lock);
 613         schedule();
 614         spin_lock(&tree->lock);
 615         finish_wait(&state->wq, &wait);
 616         return 0;
 617 }
 618
 619 /*
 620  * waits for one or more bits to clear on a range in the state tree.
 621  * The range [start, end] is inclusive.
 622  * The tree lock is taken by this function
 623  */
 624 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 625 {
 626         struct extent_state *state;
 627         struct rb_node *node;
 628
 629         spin_lock(&tree->lock);
 630 again:
 631         while (1) {
 632                 /*
 633                  * this search will find all the extents that end after
 634                  * our range starts
 635                  */
 636                 node = tree_search(tree, start);
 637                 if (!node)
 638                         break;
 639
 640                 state = rb_entry(node, struct extent_state, rb_node);
 641
 642                 if (state->start > end)
 643                         goto out;
 644
 645                 if (state->state & bits) {
 646                         start = state->start;
 647                         atomic_inc(&state->refs);
 648                         wait_on_state(tree, state);
 649                         free_extent_state(state);
 650                         goto again;
 651                 }
 652                 start = state->end + 1;
 653
 654                 if (start > end)
 655                         break;
 656
 657                 cond_resched_lock(&tree->lock);
 658         }
 659 out:
 660         spin_unlock(&tree->lock);
 661         return 0;
 662 }
 663
 664 static void set_state_bits(struct extent_io_tree *tree,
 665                            struct extent_state *state,
 666                            int *bits)
 667 {
 668         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 669
 670         set_state_cb(tree, state, bits);
 671         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 672                 u64 range = state->end - state->start + 1;
 673                 tree->dirty_bytes += range;
 674         }
 675         state->state |= bits_to_set;
 676 }
 677
 678 static void cache_state(struct extent_state *state,
 679                         struct extent_state **cached_ptr)
 680 {
 681         if (cached_ptr && !(*cached_ptr)) {
 682                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 683                         *cached_ptr = state;
 684                         atomic_inc(&state->refs);
 685                 }
 686         }
 687 }
 688
 689 static void uncache_state(struct extent_state **cached_ptr)
 690 {
 691         if (cached_ptr && (*cached_ptr)) {
 692                 struct extent_state *state = *cached_ptr;
 693                 *cached_ptr = NULL;
 694                 free_extent_state(state);
 695         }
 696 }
 697
 698 /*
 699  * set some bits on a range in the tree.  This may require allocations or
 700  * sleeping, so the gfp mask is used to indicate what is allowed.
 701  *
 702  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 703  * part of the range already has the desired bits set.  The start of the
 704  * existing range is returned in failed_start in this case.
 705  *
 706  * [start, end] is inclusive This takes the tree lock.
 707  */
 708
 709 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 710                    int bits, int exclusive_bits, u64 *failed_start,
 711                    struct extent_state **cached_state, gfp_t mask)
 712 {
 713         struct extent_state *state;
 714         struct extent_state *prealloc = NULL;
 715         struct rb_node *node;
 716         int err = 0;
 717         u64 last_start;
 718         u64 last_end;
 719
 720         bits |= EXTENT_FIRST_DELALLOC;
 721 again:
 722         if (!prealloc && (mask & __GFP_WAIT)) {
 723                 prealloc = alloc_extent_state(mask);
 724                 BUG_ON(!prealloc);
 725         }
 726
 727         spin_lock(&tree->lock);
 728         if (cached_state && *cached_state) {
 729                 state = *cached_state;
 730                 if (state->start <= start && state->end > start &&
 731                     state->tree) {
 732                         node = &state->rb_node;
 733                         goto hit_next;
 734                 }
 735         }
 736         /*
 737          * this search will find all the extents that end after
 738          * our range starts.
 739          */
 740         node = tree_search(tree, start);
 741         if (!node) {
 742                 prealloc = alloc_extent_state_atomic(prealloc);
 743                 BUG_ON(!prealloc);
 744                 err = insert_state(tree, prealloc, start, end, &bits);
 745                 prealloc = NULL;
 746                 BUG_ON(err == -EEXIST);
 747                 goto out;
 748         }
 749         state = rb_entry(node, struct extent_state, rb_node);
 750 hit_next:
 751         last_start = state->start;
 752         last_end = state->end;
 753
 754         /*
 755          * | ---- desired range ---- |
 756          * | state |
 757          *
 758          * Just lock what we found and keep going
 759          */
 760         if (state->start == start && state->end <= end) {
 761                 struct rb_node *next_node;
 762                 if (state->state & exclusive_bits) {
 763                         *failed_start = state->start;
 764                         err = -EEXIST;
 765                         goto out;
 766                 }
 767
 768                 set_state_bits(tree, state, &bits);
 769
 770                 cache_state(state, cached_state);
 771                 merge_state(tree, state);
 772                 if (last_end == (u64)-1)
 773                         goto out;
 774
 775                 start = last_end + 1;
 776                 next_node = rb_next(&state->rb_node);
 777                 if (next_node && start < end && prealloc && !need_resched()) {
 778                         state = rb_entry(next_node, struct extent_state,
 779                                          rb_node);
 780                         if (state->start == start)
 781                                 goto hit_next;
 782                 }
 783                 goto search_again;
 784         }
 785
 786         /*
 787          *     | ---- desired range ---- |
 788          * | state |
 789          *   or
 790          * | ------------- state -------------- |
 791          *
 792          * We need to split the extent we found, and may flip bits on
 793          * second half.
 794          *
 795          * If the extent we found extends past our
 796          * range, we just split and search again.  It'll get split
 797          * again the next time though.
 798          *
 799          * If the extent we found is inside our range, we set the
 800          * desired bit on it.
 801          */
 802         if (state->start < start) {
 803                 if (state->state & exclusive_bits) {
 804                         *failed_start = start;
 805                         err = -EEXIST;
 806                         goto out;
 807                 }
 808
 809                 prealloc = alloc_extent_state_atomic(prealloc);
 810                 BUG_ON(!prealloc);
 811                 err = split_state(tree, state, prealloc, start);
 812                 BUG_ON(err == -EEXIST);
 813                 prealloc = NULL;
 814                 if (err)
 815                         goto out;
 816                 if (state->end <= end) {
 817                         set_state_bits(tree, state, &bits);
 818                         cache_state(state, cached_state);
 819                         merge_state(tree, state);
 820                         if (last_end == (u64)-1)
 821                                 goto out;
 822                         start = last_end + 1;
 823                 }
 824                 goto search_again;
 825         }
 826         /*
 827          * | ---- desired range ---- |
 828          *     | state | or               | state |
 829          *
 830          * There's a hole, we need to insert something in it and
 831          * ignore the extent we found.
 832          */
 833         if (state->start > start) {
 834                 u64 this_end;
 835                 if (end < last_start)
 836                         this_end = end;
 837                 else
 838                         this_end = last_start - 1;
 839
 840                 prealloc = alloc_extent_state_atomic(prealloc);
 841                 BUG_ON(!prealloc);
 842
 843                 /*
 844                  * Avoid to free 'prealloc' if it can be merged with
 845                  * the later extent.
 846                  */
 847                 err = insert_state(tree, prealloc, start, this_end,
 848                                    &bits);
 849                 BUG_ON(err == -EEXIST);
 850                 if (err) {
 851                         free_extent_state(prealloc);
 852                         prealloc = NULL;
 853                         goto out;
 854                 }
 855                 cache_state(prealloc, cached_state);
 856                 prealloc = NULL;
 857                 start = this_end + 1;
 858                 goto search_again;
 859         }
 860         /*
 861          * | ---- desired range ---- |
 862          *                        | state |
 863          * We need to split the extent, and set the bit
 864          * on the first half
 865          */
 866         if (state->start <= end && state->end > end) {
 867                 if (state->state & exclusive_bits) {
 868                         *failed_start = start;
 869                         err = -EEXIST;
 870                         goto out;
 871                 }
 872
 873                 prealloc = alloc_extent_state_atomic(prealloc);
 874                 BUG_ON(!prealloc);
 875                 err = split_state(tree, state, prealloc, end + 1);
 876                 BUG_ON(err == -EEXIST);
 877
 878                 set_state_bits(tree, prealloc, &bits);
 879                 cache_state(prealloc, cached_state);
 880                 merge_state(tree, prealloc);
 881                 prealloc = NULL;
 882                 goto out;
 883         }
 884
 885         goto search_again;
 886
 887 out:
 888         spin_unlock(&tree->lock);
 889         if (prealloc)
 890                 free_extent_state(prealloc);
 891
 892         return err;
 893
 894 search_again:
 895         if (start > end)
 896                 goto out;
 897         spin_unlock(&tree->lock);
 898         if (mask & __GFP_WAIT)
 899                 cond_resched();
 900         goto again;
 901 }
 902
 903 /**
 904  * convert_extent - convert all bits in a given range from one bit to another
 905  * @tree:       the io tree to search
 906  * @start:      the start offset in bytes
 907  * @end:        the end offset in bytes (inclusive)
 908  * @bits:       the bits to set in this range
 909  * @clear_bits: the bits to clear in this range
 910  * @mask:       the allocation mask
 911  *
 912  * This will go through and set bits for the given range.  If any states exist
 913  * already in this range they are set with the given bit and cleared of the
 914  * clear_bits.  This is only meant to be used by things that are mergeable, ie
 915  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
 916  * boundary bits like LOCK.
 917  */
 918 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 919                        int bits, int clear_bits, gfp_t mask)
 920 {
 921         struct extent_state *state;
 922         struct extent_state *prealloc = NULL;
 923         struct rb_node *node;
 924         int err = 0;
 925         u64 last_start;
 926         u64 last_end;
 927
 928 again:
 929         if (!prealloc && (mask & __GFP_WAIT)) {
 930                 prealloc = alloc_extent_state(mask);
 931                 if (!prealloc)
 932                         return -ENOMEM;
 933         }
 934
 935         spin_lock(&tree->lock);
 936         /*
 937          * this search will find all the extents that end after
 938          * our range starts.
 939          */
 940         node = tree_search(tree, start);
 941         if (!node) {
 942                 prealloc = alloc_extent_state_atomic(prealloc);
 943                 if (!prealloc) {
 944                         err = -ENOMEM;
 945                         goto out;
 946                 }
 947                 err = insert_state(tree, prealloc, start, end, &bits);
 948                 prealloc = NULL;
 949                 BUG_ON(err == -EEXIST);
 950                 goto out;
 951         }
 952         state = rb_entry(node, struct extent_state, rb_node);
 953 hit_next:
 954         last_start = state->start;
 955         last_end = state->end;
 956
 957         /*
 958          * | ---- desired range ---- |
 959          * | state |
 960          *
 961          * Just lock what we found and keep going
 962          */
 963         if (state->start == start && state->end <= end) {
 964                 struct rb_node *next_node;
 965
 966                 set_state_bits(tree, state, &bits);
 967                 clear_state_bit(tree, state, &clear_bits, 0);
 968                 if (last_end == (u64)-1)
 969                         goto out;
 970
 971                 start = last_end + 1;
 972                 next_node = rb_next(&state->rb_node);
 973                 if (next_node && start < end && prealloc && !need_resched()) {
 974                         state = rb_entry(next_node, struct extent_state,
 975                                          rb_node);
 976                         if (state->start == start)
 977                                 goto hit_next;
 978                 }
 979                 goto search_again;
 980         }
 981
 982         /*
 983          *     | ---- desired range ---- |
 984          * | state |
 985          *   or
 986          * | ------------- state -------------- |
 987          *
 988          * We need to split the extent we found, and may flip bits on
 989          * second half.
 990          *
 991          * If the extent we found extends past our
 992          * range, we just split and search again.  It'll get split
 993          * again the next time though.
 994          *
 995          * If the extent we found is inside our range, we set the
 996          * desired bit on it.
 997          */
 998         if (state->start < start) {
 999                 prealloc = alloc_extent_state_atomic(prealloc);
1000                 if (!prealloc) {
1001                         err = -ENOMEM;
1002                         goto out;
1003                 }
1004                 err = split_state(tree, state, prealloc, start);
1005                 BUG_ON(err == -EEXIST);
1006                 prealloc = NULL;
1007                 if (err)
1008                         goto out;
1009                 if (state->end <= end) {
1010                         set_state_bits(tree, state, &bits);
1011                         clear_state_bit(tree, state, &clear_bits, 0);
1012                         if (last_end == (u64)-1)
1013                                 goto out;
1014                         start = last_end + 1;
1015                 }
1016                 goto search_again;
1017         }
1018         /*
1019          * | ---- desired range ---- |
1020          *     | state | or               | state |
1021          *
1022          * There's a hole, we need to insert something in it and
1023          * ignore the extent we found.
1024          */
1025         if (state->start > start) {
1026                 u64 this_end;
1027                 if (end < last_start)
1028                         this_end = end;
1029                 else
1030                         this_end = last_start - 1;
1031
1032                 prealloc = alloc_extent_state_atomic(prealloc);
1033                 if (!prealloc) {
1034                         err = -ENOMEM;
1035                         goto out;
1036                 }
1037
1038                 /*
1039                  * Avoid to free 'prealloc' if it can be merged with
1040                  * the later extent.
1041                  */
1042                 err = insert_state(tree, prealloc, start, this_end,
1043                                    &bits);
1044                 BUG_ON(err == -EEXIST);
1045                 if (err) {
1046                         free_extent_state(prealloc);
1047                         prealloc = NULL;
1048                         goto out;
1049                 }
1050                 prealloc = NULL;
1051                 start = this_end + 1;
1052                 goto search_again;
1053         }
1054         /*
1055          * | ---- desired range ---- |
1056          *                        | state |
1057          * We need to split the extent, and set the bit
1058          * on the first half
1059          */
1060         if (state->start <= end && state->end > end) {
1061                 prealloc = alloc_extent_state_atomic(prealloc);
1062                 if (!prealloc) {
1063                         err = -ENOMEM;
1064                         goto out;
1065                 }
1066
1067                 err = split_state(tree, state, prealloc, end + 1);
1068                 BUG_ON(err == -EEXIST);
1069
1070                 set_state_bits(tree, prealloc, &bits);
1071                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1072                 prealloc = NULL;
1073                 goto out;
1074         }
1075
1076         goto search_again;
1077
1078 out:
1079         spin_unlock(&tree->lock);
1080         if (prealloc)
1081                 free_extent_state(prealloc);
1082
1083         return err;
1084
1085 search_again:
1086         if (start > end)
1087                 goto out;
1088         spin_unlock(&tree->lock);
1089         if (mask & __GFP_WAIT)
1090                 cond_resched();
1091         goto again;
1092 }
1093
1094 /* wrappers around set/clear extent bit */
1095 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1096                      gfp_t mask)
1097 {
1098         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
1099                               NULL, mask);
1100 }
1101
1102 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1103                     int bits, gfp_t mask)
1104 {
1105         return set_extent_bit(tree, start, end, bits, 0, NULL,
1106                               NULL, mask);
1107 }
1108
1109 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1110                       int bits, gfp_t mask)
1111 {
1112         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1113 }
1114
1115 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1116                         struct extent_state **cached_state, gfp_t mask)
1117 {
1118         return set_extent_bit(tree, start, end,
1119                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1120                               0, NULL, cached_state, mask);
1121 }
1122
1123 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1124                        gfp_t mask)
1125 {
1126         return clear_extent_bit(tree, start, end,
1127                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1128                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1129 }
1130
1131 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1132                      gfp_t mask)
1133 {
1134         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
1135                               NULL, mask);
1136 }
1137
1138 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1139                         struct extent_state **cached_state, gfp_t mask)
1140 {
1141         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
1142                               NULL, cached_state, mask);
1143 }
1144
1145 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
1146                                  u64 end, struct extent_state **cached_state,
1147                                  gfp_t mask)
1148 {
1149         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1150                                 cached_state, mask);
1151 }
1152
1153 /*
1154  * either insert or lock state struct between start and end use mask to tell
1155  * us if waiting is desired.
1156  */
1157 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1158                      int bits, struct extent_state **cached_state, gfp_t mask)
1159 {
1160         int err;
1161         u64 failed_start;
1162         while (1) {
1163                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1164                                      EXTENT_LOCKED, &failed_start,
1165                                      cached_state, mask);
1166                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
1167                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1168                         start = failed_start;
1169                 } else {
1170                         break;
1171                 }
1172                 WARN_ON(start > end);
1173         }
1174         return err;
1175 }
1176
1177 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1178 {
1179         return lock_extent_bits(tree, start, end, 0, NULL, mask);
1180 }
1181
1182 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1183                     gfp_t mask)
1184 {
1185         int err;
1186         u64 failed_start;
1187
1188         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1189                              &failed_start, NULL, mask);
1190         if (err == -EEXIST) {
1191                 if (failed_start > start)
1192                         clear_extent_bit(tree, start, failed_start - 1,
1193                                          EXTENT_LOCKED, 1, 0, NULL, mask);
1194                 return 0;
1195         }
1196         return 1;
1197 }
1198
1199 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1200                          struct extent_state **cached, gfp_t mask)
1201 {
1202         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1203                                 mask);
1204 }
1205
1206 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1207 {
1208         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1209                                 mask);
1210 }
1211
1212 /*
1213  * helper function to set both pages and extents in the tree writeback
1214  */
1215 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1216 {
1217         unsigned long index = start >> PAGE_CACHE_SHIFT;
1218         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1219         struct page *page;
1220
1221         while (index <= end_index) {
1222                 page = find_get_page(tree->mapping, index);
1223                 BUG_ON(!page);
1224                 set_page_writeback(page);
1225                 page_cache_release(page);
1226                 index++;
1227         }
1228         return 0;
1229 }
1230
1231 /* find the first state struct with 'bits' set after 'start', and
1232  * return it.  tree->lock must be held.  NULL will returned if
1233  * nothing was found after 'start'
1234  */
1235 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1236                                                  u64 start, int bits)
1237 {
1238         struct rb_node *node;
1239         struct extent_state *state;
1240
1241         /*
1242          * this search will find all the extents that end after
1243          * our range starts.
1244          */
1245         node = tree_search(tree, start);
1246         if (!node)
1247                 goto out;
1248
1249         while (1) {
1250                 state = rb_entry(node, struct extent_state, rb_node);
1251                 if (state->end >= start && (state->state & bits))
1252                         return state;
1253
1254                 node = rb_next(node);
1255                 if (!node)
1256                         break;
1257         }
1258 out:
1259         return NULL;
1260 }
1261
1262 /*
1263  * find the first offset in the io tree with 'bits' set. zero is
1264  * returned if we find something, and *start_ret and *end_ret are
1265  * set to reflect the state struct that was found.
1266  *
1267  * If nothing was found, 1 is returned, < 0 on error
1268  */
1269 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1270                           u64 *start_ret, u64 *end_ret, int bits)
1271 {
1272         struct extent_state *state;
1273         int ret = 1;
1274
1275         spin_lock(&tree->lock);
1276         state = find_first_extent_bit_state(tree, start, bits);
1277         if (state) {
1278                 *start_ret = state->start;
1279                 *end_ret = state->end;
1280                 ret = 0;
1281         }
1282         spin_unlock(&tree->lock);
1283         return ret;
1284 }
1285
1286 /*
1287  * find a contiguous range of bytes in the file marked as delalloc, not
1288  * more than 'max_bytes'.  start and end are used to return the range,
1289  *
1290  * 1 is returned if we find something, 0 if nothing was in the tree
1291  */
1292 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1293                                         u64 *start, u64 *end, u64 max_bytes,
1294                                         struct extent_state **cached_state)
1295 {
1296         struct rb_node *node;
1297         struct extent_state *state;
1298         u64 cur_start = *start;
1299         u64 found = 0;
1300         u64 total_bytes = 0;
1301
1302         spin_lock(&tree->lock);
1303
1304         /*
1305          * this search will find all the extents that end after
1306          * our range starts.
1307          */
1308         node = tree_search(tree, cur_start);
1309         if (!node) {
1310                 if (!found)
1311                         *end = (u64)-1;
1312                 goto out;
1313         }
1314
1315         while (1) {
1316                 state = rb_entry(node, struct extent_state, rb_node);
1317                 if (found && (state->start != cur_start ||
1318                               (state->state & EXTENT_BOUNDARY))) {
1319                         goto out;
1320                 }
1321                 if (!(state->state & EXTENT_DELALLOC)) {
1322                         if (!found)
1323                                 *end = state->end;
1324                         goto out;
1325                 }
1326                 if (!found) {
1327                         *start = state->start;
1328                         *cached_state = state;
1329                         atomic_inc(&state->refs);
1330                 }
1331                 found++;
1332                 *end = state->end;
1333                 cur_start = state->end + 1;
1334                 node = rb_next(node);
1335                 if (!node)
1336                         break;
1337                 total_bytes += state->end - state->start + 1;
1338                 if (total_bytes >= max_bytes)
1339                         break;
1340         }
1341 out:
1342         spin_unlock(&tree->lock);
1343         return found;
1344 }
1345
1346 static noinline int __unlock_for_delalloc(struct inode *inode,
1347                                           struct page *locked_page,
1348                                           u64 start, u64 end)
1349 {
1350         int ret;
1351         struct page *pages[16];
1352         unsigned long index = start >> PAGE_CACHE_SHIFT;
1353         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1354         unsigned long nr_pages = end_index - index + 1;
1355         int i;
1356
1357         if (index == locked_page->index && end_index == index)
1358                 return 0;
1359
1360         while (nr_pages > 0) {
1361                 ret = find_get_pages_contig(inode->i_mapping, index,
1362                                      min_t(unsigned long, nr_pages,
1363                                      ARRAY_SIZE(pages)), pages);
1364                 for (i = 0; i < ret; i++) {
1365                         if (pages[i] != locked_page)
1366                                 unlock_page(pages[i]);
1367                         page_cache_release(pages[i]);
1368                 }
1369                 nr_pages -= ret;
1370                 index += ret;
1371                 cond_resched();
1372         }
1373         return 0;
1374 }
1375
1376 static noinline int lock_delalloc_pages(struct inode *inode,
1377                                         struct page *locked_page,
1378                                         u64 delalloc_start,
1379                                         u64 delalloc_end)
1380 {
1381         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1382         unsigned long start_index = index;
1383         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1384         unsigned long pages_locked = 0;
1385         struct page *pages[16];
1386         unsigned long nrpages;
1387         int ret;
1388         int i;
1389
1390         /* the caller is responsible for locking the start index */
1391         if (index == locked_page->index && index == end_index)
1392                 return 0;
1393
1394         /* skip the page at the start index */
1395         nrpages = end_index - index + 1;
1396         while (nrpages > 0) {
1397                 ret = find_get_pages_contig(inode->i_mapping, index,
1398                                      min_t(unsigned long,
1399                                      nrpages, ARRAY_SIZE(pages)), pages);
1400                 if (ret == 0) {
1401                         ret = -EAGAIN;
1402                         goto done;
1403                 }
1404                 /* now we have an array of pages, lock them all */
1405                 for (i = 0; i < ret; i++) {
1406                         /*
1407                          * the caller is taking responsibility for
1408                          * locked_page
1409                          */
1410                         if (pages[i] != locked_page) {
1411                                 lock_page(pages[i]);
1412                                 if (!PageDirty(pages[i]) ||
1413                                     pages[i]->mapping != inode->i_mapping) {
1414                                         ret = -EAGAIN;
1415                                         unlock_page(pages[i]);
1416                                         page_cache_release(pages[i]);
1417                                         goto done;
1418                                 }
1419                         }
1420                         page_cache_release(pages[i]);
1421                         pages_locked++;
1422                 }
1423                 nrpages -= ret;
1424                 index += ret;
1425                 cond_resched();
1426         }
1427         ret = 0;
1428 done:
1429         if (ret && pages_locked) {
1430                 __unlock_for_delalloc(inode, locked_page,
1431                               delalloc_start,
1432                               ((u64)(start_index + pages_locked - 1)) <<
1433                               PAGE_CACHE_SHIFT);
1434         }
1435         return ret;
1436 }
1437
1438 /*
1439  * find a contiguous range of bytes in the file marked as delalloc, not
1440  * more than 'max_bytes'.  start and end are used to return the range,
1441  *
1442  * 1 is returned if we find something, 0 if nothing was in the tree
1443  */
1444 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1445                                              struct extent_io_tree *tree,
1446                                              struct page *locked_page,
1447                                              u64 *start, u64 *end,
1448                                              u64 max_bytes)
1449 {
1450         u64 delalloc_start;
1451         u64 delalloc_end;
1452         u64 found;
1453         struct extent_state *cached_state = NULL;
1454         int ret;
1455         int loops = 0;
1456
1457 again:
1458         /* step one, find a bunch of delalloc bytes starting at start */
1459         delalloc_start = *start;
1460         delalloc_end = 0;
1461         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1462                                     max_bytes, &cached_state);
1463         if (!found || delalloc_end <= *start) {
1464                 *start = delalloc_start;
1465                 *end = delalloc_end;
1466                 free_extent_state(cached_state);
1467                 return found;
1468         }
1469
1470         /*
1471          * start comes from the offset of locked_page.  We have to lock
1472          * pages in order, so we can't process delalloc bytes before
1473          * locked_page
1474          */
1475         if (delalloc_start < *start)
1476                 delalloc_start = *start;
1477
1478         /*
1479          * make sure to limit the number of pages we try to lock down
1480          * if we're looping.
1481          */
1482         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1483                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1484
1485         /* step two, lock all the pages after the page that has start */
1486         ret = lock_delalloc_pages(inode, locked_page,
1487                                   delalloc_start, delalloc_end);
1488         if (ret == -EAGAIN) {
1489                 /* some of the pages are gone, lets avoid looping by
1490                  * shortening the size of the delalloc range we're searching
1491                  */
1492                 free_extent_state(cached_state);
1493                 if (!loops) {
1494                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1495                         max_bytes = PAGE_CACHE_SIZE - offset;
1496                         loops = 1;
1497                         goto again;
1498                 } else {
1499                         found = 0;
1500                         goto out_failed;
1501                 }
1502         }
1503         BUG_ON(ret);
1504
1505         /* step three, lock the state bits for the whole range */
1506         lock_extent_bits(tree, delalloc_start, delalloc_end,
1507                          0, &cached_state, GFP_NOFS);
1508
1509         /* then test to make sure it is all still delalloc */
1510         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1511                              EXTENT_DELALLOC, 1, cached_state);
1512         if (!ret) {
1513                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1514                                      &cached_state, GFP_NOFS);
1515                 __unlock_for_delalloc(inode, locked_page,
1516                               delalloc_start, delalloc_end);
1517                 cond_resched();
1518                 goto again;
1519         }
1520         free_extent_state(cached_state);
1521         *start = delalloc_start;
1522         *end = delalloc_end;
1523 out_failed:
1524         return found;
1525 }
1526
1527 int extent_clear_unlock_delalloc(struct inode *inode,
1528                                 struct extent_io_tree *tree,
1529                                 u64 start, u64 end, struct page *locked_page,
1530                                 unsigned long op)
1531 {
1532         int ret;
1533         struct page *pages[16];
1534         unsigned long index = start >> PAGE_CACHE_SHIFT;
1535         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1536         unsigned long nr_pages = end_index - index + 1;
1537         int i;
1538         int clear_bits = 0;
1539
1540         if (op & EXTENT_CLEAR_UNLOCK)
1541                 clear_bits |= EXTENT_LOCKED;
1542         if (op & EXTENT_CLEAR_DIRTY)
1543                 clear_bits |= EXTENT_DIRTY;
1544
1545         if (op & EXTENT_CLEAR_DELALLOC)
1546                 clear_bits |= EXTENT_DELALLOC;
1547
1548         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1549         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1550                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1551                     EXTENT_SET_PRIVATE2)))
1552                 return 0;
1553
1554         while (nr_pages > 0) {
1555                 ret = find_get_pages_contig(inode->i_mapping, index,
1556                                      min_t(unsigned long,
1557                                      nr_pages, ARRAY_SIZE(pages)), pages);
1558                 for (i = 0; i < ret; i++) {
1559
1560                         if (op & EXTENT_SET_PRIVATE2)
1561                                 SetPagePrivate2(pages[i]);
1562
1563                         if (pages[i] == locked_page) {
1564                                 page_cache_release(pages[i]);
1565                                 continue;
1566                         }
1567                         if (op & EXTENT_CLEAR_DIRTY)
1568                                 clear_page_dirty_for_io(pages[i]);
1569                         if (op & EXTENT_SET_WRITEBACK)
1570                                 set_page_writeback(pages[i]);
1571                         if (op & EXTENT_END_WRITEBACK)
1572                                 end_page_writeback(pages[i]);
1573                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1574                                 unlock_page(pages[i]);
1575                         page_cache_release(pages[i]);
1576                 }
1577                 nr_pages -= ret;
1578                 index += ret;
1579                 cond_resched();
1580         }
1581         return 0;
1582 }
1583
1584 /*
1585  * count the number of bytes in the tree that have a given bit(s)
1586  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1587  * cached.  The total number found is returned.
1588  */
1589 u64 count_range_bits(struct extent_io_tree *tree,
1590                      u64 *start, u64 search_end, u64 max_bytes,
1591                      unsigned long bits, int contig)
1592 {
1593         struct rb_node *node;
1594         struct extent_state *state;
1595         u64 cur_start = *start;
1596         u64 total_bytes = 0;
1597         u64 last = 0;
1598         int found = 0;
1599
1600         if (search_end <= cur_start) {
1601                 WARN_ON(1);
1602                 return 0;
1603         }
1604
1605         spin_lock(&tree->lock);
1606         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1607                 total_bytes = tree->dirty_bytes;
1608                 goto out;
1609         }
1610         /*
1611          * this search will find all the extents that end after
1612          * our range starts.
1613          */
1614         node = tree_search(tree, cur_start);
1615         if (!node)
1616                 goto out;
1617
1618         while (1) {
1619                 state = rb_entry(node, struct extent_state, rb_node);
1620                 if (state->start > search_end)
1621                         break;
1622                 if (contig && found && state->start > last + 1)
1623                         break;
1624                 if (state->end >= cur_start && (state->state & bits) == bits) {
1625                         total_bytes += min(search_end, state->end) + 1 -
1626                                        max(cur_start, state->start);
1627                         if (total_bytes >= max_bytes)
1628                                 break;
1629                         if (!found) {
1630                                 *start = max(cur_start, state->start);
1631                                 found = 1;
1632                         }
1633                         last = state->end;
1634                 } else if (contig && found) {
1635                         break;
1636                 }
1637                 node = rb_next(node);
1638                 if (!node)
1639                         break;
1640         }
1641 out:
1642         spin_unlock(&tree->lock);
1643         return total_bytes;
1644 }
1645
1646 /*
1647  * set the private field for a given byte offset in the tree.  If there isn't
1648  * an extent_state there already, this does nothing.
1649  */
1650 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1651 {
1652         struct rb_node *node;
1653         struct extent_state *state;
1654         int ret = 0;
1655
1656         spin_lock(&tree->lock);
1657         /*
1658          * this search will find all the extents that end after
1659          * our range starts.
1660          */
1661         node = tree_search(tree, start);
1662         if (!node) {
1663                 ret = -ENOENT;
1664                 goto out;
1665         }
1666         state = rb_entry(node, struct extent_state, rb_node);
1667         if (state->start != start) {
1668                 ret = -ENOENT;
1669                 goto out;
1670         }
1671         state->private = private;
1672 out:
1673         spin_unlock(&tree->lock);
1674         return ret;
1675 }
1676
1677 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1678 {
1679         struct rb_node *node;
1680         struct extent_state *state;
1681         int ret = 0;
1682
1683         spin_lock(&tree->lock);
1684         /*
1685          * this search will find all the extents that end after
1686          * our range starts.
1687          */
1688         node = tree_search(tree, start);
1689         if (!node) {
1690                 ret = -ENOENT;
1691                 goto out;
1692         }
1693         state = rb_entry(node, struct extent_state, rb_node);
1694         if (state->start != start) {
1695                 ret = -ENOENT;
1696                 goto out;
1697         }
1698         *private = state->private;
1699 out:
1700         spin_unlock(&tree->lock);
1701         return ret;
1702 }
1703
1704 /*
1705  * searches a range in the state tree for a given mask.
1706  * If 'filled' == 1, this returns 1 only if every extent in the tree
1707  * has the bits set.  Otherwise, 1 is returned if any bit in the
1708  * range is found set.
1709  */
1710 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1711                    int bits, int filled, struct extent_state *cached)
1712 {
1713         struct extent_state *state = NULL;
1714         struct rb_node *node;
1715         int bitset = 0;
1716
1717         spin_lock(&tree->lock);
1718         if (cached && cached->tree && cached->start <= start &&
1719             cached->end > start)
1720                 node = &cached->rb_node;
1721         else
1722                 node = tree_search(tree, start);
1723         while (node && start <= end) {
1724                 state = rb_entry(node, struct extent_state, rb_node);
1725
1726                 if (filled && state->start > start) {
1727                         bitset = 0;
1728                         break;
1729                 }
1730
1731                 if (state->start > end)
1732                         break;
1733
1734                 if (state->state & bits) {
1735                         bitset = 1;
1736                         if (!filled)
1737                                 break;
1738                 } else if (filled) {
1739                         bitset = 0;
1740                         break;
1741                 }
1742
1743                 if (state->end == (u64)-1)
1744                         break;
1745
1746                 start = state->end + 1;
1747                 if (start > end)
1748                         break;
1749                 node = rb_next(node);
1750                 if (!node) {
1751                         if (filled)
1752                                 bitset = 0;
1753                         break;
1754                 }
1755         }
1756         spin_unlock(&tree->lock);
1757         return bitset;
1758 }
1759
1760 /*
1761  * helper function to set a given page up to date if all the
1762  * extents in the tree for that page are up to date
1763  */
1764 static int check_page_uptodate(struct extent_io_tree *tree,
1765                                struct page *page)
1766 {
1767         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1768         u64 end = start + PAGE_CACHE_SIZE - 1;
1769         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1770                 SetPageUptodate(page);
1771         return 0;
1772 }
1773
1774 /*
1775  * helper function to unlock a page if all the extents in the tree
1776  * for that page are unlocked
1777  */
1778 static int check_page_locked(struct extent_io_tree *tree,
1779                              struct page *page)
1780 {
1781         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1782         u64 end = start + PAGE_CACHE_SIZE - 1;
1783         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1784                 unlock_page(page);
1785         return 0;
1786 }
1787
1788 /*
1789  * helper function to end page writeback if all the extents
1790  * in the tree for that page are done with writeback
1791  */
1792 static int check_page_writeback(struct extent_io_tree *tree,
1793                              struct page *page)
1794 {
1795         end_page_writeback(page);
1796         return 0;
1797 }
1798
1799 /*
1800  * When IO fails, either with EIO or csum verification fails, we
1801  * try other mirrors that might have a good copy of the data.  This
1802  * io_failure_record is used to record state as we go through all the
1803  * mirrors.  If another mirror has good data, the page is set up to date
1804  * and things continue.  If a good mirror can't be found, the original
1805  * bio end_io callback is called to indicate things have failed.
1806  */
1807 struct io_failure_record {
1808         struct page *page;
1809         u64 start;
1810         u64 len;
1811         u64 logical;
1812         unsigned long bio_flags;
1813         int this_mirror;
1814         int failed_mirror;
1815         int in_validation;
1816 };
1817
1818 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1819                                 int did_repair)
1820 {
1821         int ret;
1822         int err = 0;
1823         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1824
1825         set_state_private(failure_tree, rec->start, 0);
1826         ret = clear_extent_bits(failure_tree, rec->start,
1827                                 rec->start + rec->len - 1,
1828                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1829         if (ret)
1830                 err = ret;
1831
1832         if (did_repair) {
1833                 ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1834                                         rec->start + rec->len - 1,
1835                                         EXTENT_DAMAGED, GFP_NOFS);
1836                 if (ret && !err)
1837                         err = ret;
1838         }
1839
1840         kfree(rec);
1841         return err;
1842 }
1843
1844 static void repair_io_failure_callback(struct bio *bio, int err)
1845 {
1846         complete(bio->bi_private);
1847 }
1848
1849 /*
1850  * this bypasses the standard btrfs submit functions deliberately, as
1851  * the standard behavior is to write all copies in a raid setup. here we only
1852  * want to write the one bad copy. so we do the mapping for ourselves and issue
1853  * submit_bio directly.
1854  * to avoid any synchonization issues, wait for the data after writing, which
1855  * actually prevents the read that triggered the error from finishing.
1856  * currently, there can be no more than two copies of every data bit. thus,
1857  * exactly one rewrite is required.
1858  */
1859 int repair_io_failure(struct btrfs_mapping_tree *map_tree, u64 start,
1860                         u64 length, u64 logical, struct page *page,
1861                         int mirror_num)
1862 {
1863         struct bio *bio;
1864         struct btrfs_device *dev;
1865         DECLARE_COMPLETION_ONSTACK(compl);
1866         u64 map_length = 0;
1867         u64 sector;
1868         struct btrfs_bio *bbio = NULL;
1869         int ret;
1870
1871         BUG_ON(!mirror_num);
1872
1873         bio = bio_alloc(GFP_NOFS, 1);
1874         if (!bio)
1875                 return -EIO;
1876         bio->bi_private = &compl;
1877         bio->bi_end_io = repair_io_failure_callback;
1878         bio->bi_size = 0;
1879         map_length = length;
1880
1881         ret = btrfs_map_block(map_tree, WRITE, logical,
1882                               &map_length, &bbio, mirror_num);
1883         if (ret) {
1884                 bio_put(bio);
1885                 return -EIO;
1886         }
1887         BUG_ON(mirror_num != bbio->mirror_num);
1888         sector = bbio->stripes[mirror_num-1].physical >> 9;
1889         bio->bi_sector = sector;
1890         dev = bbio->stripes[mirror_num-1].dev;
1891         kfree(bbio);
1892         if (!dev || !dev->bdev || !dev->writeable) {
1893                 bio_put(bio);
1894                 return -EIO;
1895         }
1896         bio->bi_bdev = dev->bdev;
1897         bio_add_page(bio, page, length, start-page_offset(page));
1898         btrfsic_submit_bio(WRITE_SYNC, bio);
1899         wait_for_completion(&compl);
1900
1901         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
1902                 /* try to remap that extent elsewhere? */
1903                 bio_put(bio);
1904                 return -EIO;
1905         }
1906
1907         printk(KERN_INFO "btrfs read error corrected: ino %lu off %llu (dev %s "
1908                         "sector %llu)\n", page->mapping->host->i_ino, start,
1909                         dev->name, sector);
1910
1911         bio_put(bio);
1912         return 0;
1913 }
1914
1915 /*
1916  * each time an IO finishes, we do a fast check in the IO failure tree
1917  * to see if we need to process or clean up an io_failure_record
1918  */
1919 static int clean_io_failure(u64 start, struct page *page)
1920 {
1921         u64 private;
1922         u64 private_failure;
1923         struct io_failure_record *failrec;
1924         struct btrfs_mapping_tree *map_tree;
1925         struct extent_state *state;
1926         int num_copies;
1927         int did_repair = 0;
1928         int ret;
1929         struct inode *inode = page->mapping->host;
1930
1931         private = 0;
1932         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
1933                                 (u64)-1, 1, EXTENT_DIRTY, 0);
1934         if (!ret)
1935                 return 0;
1936
1937         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
1938                                 &private_failure);
1939         if (ret)
1940                 return 0;
1941
1942         failrec = (struct io_failure_record *)(unsigned long) private_failure;
1943         BUG_ON(!failrec->this_mirror);
1944
1945         if (failrec->in_validation) {
1946                 /* there was no real error, just free the record */
1947                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
1948                          failrec->start);
1949                 did_repair = 1;
1950                 goto out;
1951         }
1952
1953         spin_lock(&BTRFS_I(inode)->io_tree.lock);
1954         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
1955                                             failrec->start,
1956                                             EXTENT_LOCKED);
1957         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
1958
1959         if (state && state->start == failrec->start) {
1960                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
1961                 num_copies = btrfs_num_copies(map_tree, failrec->logical,
1962                                                 failrec->len);
1963                 if (num_copies > 1)  {
1964                         ret = repair_io_failure(map_tree, start, failrec->len,
1965                                                 failrec->logical, page,
1966                                                 failrec->failed_mirror);
1967                         did_repair = !ret;
1968                 }
1969         }
1970
1971 out:
1972         if (!ret)
1973                 ret = free_io_failure(inode, failrec, did_repair);
1974
1975         return ret;
1976 }
1977
1978 /*
1979  * this is a generic handler for readpage errors (default
1980  * readpage_io_failed_hook). if other copies exist, read those and write back
1981  * good data to the failed position. does not investigate in remapping the
1982  * failed extent elsewhere, hoping the device will be smart enough to do this as
1983  * needed
1984  */
1985
1986 static int bio_readpage_error(struct bio *failed_bio, struct page *page,
1987                                 u64 start, u64 end, int failed_mirror,
1988                                 struct extent_state *state)
1989 {
1990         struct io_failure_record *failrec = NULL;
1991         u64 private;
1992         struct extent_map *em;
1993         struct inode *inode = page->mapping->host;
1994         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1995         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1996         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
1997         struct bio *bio;
1998         int num_copies;
1999         int ret;
2000         int read_mode;
2001         u64 logical;
2002
2003         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2004
2005         ret = get_state_private(failure_tree, start, &private);
2006         if (ret) {
2007                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2008                 if (!failrec)
2009                         return -ENOMEM;
2010                 failrec->start = start;
2011                 failrec->len = end - start + 1;
2012                 failrec->this_mirror = 0;
2013                 failrec->bio_flags = 0;
2014                 failrec->in_validation = 0;
2015
2016                 read_lock(&em_tree->lock);
2017                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2018                 if (!em) {
2019                         read_unlock(&em_tree->lock);
2020                         kfree(failrec);
2021                         return -EIO;
2022                 }
2023
2024                 if (em->start > start || em->start + em->len < start) {
2025                         free_extent_map(em);
2026                         em = NULL;
2027                 }
2028                 read_unlock(&em_tree->lock);
2029
2030                 if (!em || IS_ERR(em)) {
2031                         kfree(failrec);
2032                         return -EIO;
2033                 }
2034                 logical = start - em->start;
2035                 logical = em->block_start + logical;
2036                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2037                         logical = em->block_start;
2038                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2039                         extent_set_compress_type(&failrec->bio_flags,
2040                                                  em->compress_type);
2041                 }
2042                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2043                          "len=%llu\n", logical, start, failrec->len);
2044                 failrec->logical = logical;
2045                 free_extent_map(em);
2046
2047                 /* set the bits in the private failure tree */
2048                 ret = set_extent_bits(failure_tree, start, end,
2049                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2050                 if (ret >= 0)
2051                         ret = set_state_private(failure_tree, start,
2052                                                 (u64)(unsigned long)failrec);
2053                 /* set the bits in the inode's tree */
2054                 if (ret >= 0)
2055                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2056                                                 GFP_NOFS);
2057                 if (ret < 0) {
2058                         kfree(failrec);
2059                         return ret;
2060                 }
2061         } else {
2062                 failrec = (struct io_failure_record *)(unsigned long)private;
2063                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2064                          "start=%llu, len=%llu, validation=%d\n",
2065                          failrec->logical, failrec->start, failrec->len,
2066                          failrec->in_validation);
2067                 /*
2068                  * when data can be on disk more than twice, add to failrec here
2069                  * (e.g. with a list for failed_mirror) to make
2070                  * clean_io_failure() clean all those errors at once.
2071                  */
2072         }
2073         num_copies = btrfs_num_copies(
2074                               &BTRFS_I(inode)->root->fs_info->mapping_tree,
2075                               failrec->logical, failrec->len);
2076         if (num_copies == 1) {
2077                 /*
2078                  * we only have a single copy of the data, so don't bother with
2079                  * all the retry and error correction code that follows. no
2080                  * matter what the error is, it is very likely to persist.
2081                  */
2082                 pr_debug("bio_readpage_error: cannot repair, num_copies == 1. "
2083                          "state=%p, num_copies=%d, next_mirror %d, "
2084                          "failed_mirror %d\n", state, num_copies,
2085                          failrec->this_mirror, failed_mirror);
2086                 free_io_failure(inode, failrec, 0);
2087                 return -EIO;
2088         }
2089
2090         if (!state) {
2091                 spin_lock(&tree->lock);
2092                 state = find_first_extent_bit_state(tree, failrec->start,
2093                                                     EXTENT_LOCKED);
2094                 if (state && state->start != failrec->start)
2095                         state = NULL;
2096                 spin_unlock(&tree->lock);
2097         }
2098
2099         /*
2100          * there are two premises:
2101          *      a) deliver good data to the caller
2102          *      b) correct the bad sectors on disk
2103          */
2104         if (failed_bio->bi_vcnt > 1) {
2105                 /*
2106                  * to fulfill b), we need to know the exact failing sectors, as
2107                  * we don't want to rewrite any more than the failed ones. thus,
2108                  * we need separate read requests for the failed bio
2109                  *
2110                  * if the following BUG_ON triggers, our validation request got
2111                  * merged. we need separate requests for our algorithm to work.
2112                  */
2113                 BUG_ON(failrec->in_validation);
2114                 failrec->in_validation = 1;
2115                 failrec->this_mirror = failed_mirror;
2116                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2117         } else {
2118                 /*
2119                  * we're ready to fulfill a) and b) alongside. get a good copy
2120                  * of the failed sector and if we succeed, we have setup
2121                  * everything for repair_io_failure to do the rest for us.
2122                  */
2123                 if (failrec->in_validation) {
2124                         BUG_ON(failrec->this_mirror != failed_mirror);
2125                         failrec->in_validation = 0;
2126                         failrec->this_mirror = 0;
2127                 }
2128                 failrec->failed_mirror = failed_mirror;
2129                 failrec->this_mirror++;
2130                 if (failrec->this_mirror == failed_mirror)
2131                         failrec->this_mirror++;
2132                 read_mode = READ_SYNC;
2133         }
2134
2135         if (!state || failrec->this_mirror > num_copies) {
2136                 pr_debug("bio_readpage_error: (fail) state=%p, num_copies=%d, "
2137                          "next_mirror %d, failed_mirror %d\n", state,
2138                          num_copies, failrec->this_mirror, failed_mirror);
2139                 free_io_failure(inode, failrec, 0);
2140                 return -EIO;
2141         }
2142
2143         bio = bio_alloc(GFP_NOFS, 1);
2144         bio->bi_private = state;
2145         bio->bi_end_io = failed_bio->bi_end_io;
2146         bio->bi_sector = failrec->logical >> 9;
2147         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2148         bio->bi_size = 0;
2149
2150         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2151
2152         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2153                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2154                  failrec->this_mirror, num_copies, failrec->in_validation);
2155
2156         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2157                                          failrec->this_mirror,
2158                                          failrec->bio_flags, 0);
2159         return ret;
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 attach_extent_buffer_page(struct extent_buffer *eb, struct page *page)
2477 {
2478         if (!PagePrivate(page)) {
2479                 SetPagePrivate(page);
2480                 page_cache_get(page);
2481                 set_page_private(page, (unsigned long)eb);
2482         } else {
2483                 WARN_ON(page->private != (unsigned long)eb);
2484         }
2485 }
2486
2487 void set_page_extent_mapped(struct page *page)
2488 {
2489         if (!PagePrivate(page)) {
2490                 SetPagePrivate(page);
2491                 page_cache_get(page);
2492                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2493         }
2494 }
2495
2496 /*
2497  * basic readpage implementation.  Locked extent state structs are inserted
2498  * into the tree that are removed when the IO is done (by the end_io
2499  * handlers)
2500  */
2501 static int __extent_read_full_page(struct extent_io_tree *tree,
2502                                    struct page *page,
2503                                    get_extent_t *get_extent,
2504                                    struct bio **bio, int mirror_num,
2505                                    unsigned long *bio_flags)
2506 {
2507         struct inode *inode = page->mapping->host;
2508         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2509         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2510         u64 end;
2511         u64 cur = start;
2512         u64 extent_offset;
2513         u64 last_byte = i_size_read(inode);
2514         u64 block_start;
2515         u64 cur_end;
2516         sector_t sector;
2517         struct extent_map *em;
2518         struct block_device *bdev;
2519         struct btrfs_ordered_extent *ordered;
2520         int ret;
2521         int nr = 0;
2522         size_t pg_offset = 0;
2523         size_t iosize;
2524         size_t disk_io_size;
2525         size_t blocksize = inode->i_sb->s_blocksize;
2526         unsigned long this_bio_flag = 0;
2527
2528         set_page_extent_mapped(page);
2529
2530         if (!PageUptodate(page)) {
2531                 if (cleancache_get_page(page) == 0) {
2532                         BUG_ON(blocksize != PAGE_SIZE);
2533                         goto out;
2534                 }
2535         }
2536
2537         end = page_end;
2538         while (1) {
2539                 lock_extent(tree, start, end, GFP_NOFS);
2540                 ordered = btrfs_lookup_ordered_extent(inode, start);
2541                 if (!ordered)
2542                         break;
2543                 unlock_extent(tree, start, end, GFP_NOFS);
2544                 btrfs_start_ordered_extent(inode, ordered, 1);
2545                 btrfs_put_ordered_extent(ordered);
2546         }
2547
2548         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2549                 char *userpage;
2550                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2551
2552                 if (zero_offset) {
2553                         iosize = PAGE_CACHE_SIZE - zero_offset;
2554                         userpage = kmap_atomic(page, KM_USER0);
2555                         memset(userpage + zero_offset, 0, iosize);
2556                         flush_dcache_page(page);
2557                         kunmap_atomic(userpage, KM_USER0);
2558                 }
2559         }
2560         while (cur <= end) {
2561                 if (cur >= last_byte) {
2562                         char *userpage;
2563                         struct extent_state *cached = NULL;
2564
2565                         iosize = PAGE_CACHE_SIZE - pg_offset;
2566                         userpage = kmap_atomic(page, KM_USER0);
2567                         memset(userpage + pg_offset, 0, iosize);
2568                         flush_dcache_page(page);
2569                         kunmap_atomic(userpage, KM_USER0);
2570                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2571                                             &cached, GFP_NOFS);
2572                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2573                                              &cached, GFP_NOFS);
2574                         break;
2575                 }
2576                 em = get_extent(inode, page, pg_offset, cur,
2577                                 end - cur + 1, 0);
2578                 if (IS_ERR_OR_NULL(em)) {
2579                         SetPageError(page);
2580                         unlock_extent(tree, cur, end, GFP_NOFS);
2581                         break;
2582                 }
2583                 extent_offset = cur - em->start;
2584                 BUG_ON(extent_map_end(em) <= cur);
2585                 BUG_ON(end < cur);
2586
2587                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2588                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2589                         extent_set_compress_type(&this_bio_flag,
2590                                                  em->compress_type);
2591                 }
2592
2593                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2594                 cur_end = min(extent_map_end(em) - 1, end);
2595                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2596                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2597                         disk_io_size = em->block_len;
2598                         sector = em->block_start >> 9;
2599                 } else {
2600                         sector = (em->block_start + extent_offset) >> 9;
2601                         disk_io_size = iosize;
2602                 }
2603                 bdev = em->bdev;
2604                 block_start = em->block_start;
2605                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2606                         block_start = EXTENT_MAP_HOLE;
2607                 free_extent_map(em);
2608                 em = NULL;
2609
2610                 /* we've found a hole, just zero and go on */
2611                 if (block_start == EXTENT_MAP_HOLE) {
2612                         char *userpage;
2613                         struct extent_state *cached = NULL;
2614
2615                         userpage = kmap_atomic(page, KM_USER0);
2616                         memset(userpage + pg_offset, 0, iosize);
2617                         flush_dcache_page(page);
2618                         kunmap_atomic(userpage, KM_USER0);
2619
2620                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2621                                             &cached, GFP_NOFS);
2622                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2623                                              &cached, GFP_NOFS);
2624                         cur = cur + iosize;
2625                         pg_offset += iosize;
2626                         continue;
2627                 }
2628                 /* the get_extent function already copied into the page */
2629                 if (test_range_bit(tree, cur, cur_end,
2630                                    EXTENT_UPTODATE, 1, NULL)) {
2631                         check_page_uptodate(tree, page);
2632                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2633                         cur = cur + iosize;
2634                         pg_offset += iosize;
2635                         continue;
2636                 }
2637                 /* we have an inline extent but it didn't get marked up
2638                  * to date.  Error out
2639                  */
2640                 if (block_start == EXTENT_MAP_INLINE) {
2641                         SetPageError(page);
2642                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2643                         cur = cur + iosize;
2644                         pg_offset += iosize;
2645                         continue;
2646                 }
2647
2648                 ret = 0;
2649                 if (tree->ops && tree->ops->readpage_io_hook) {
2650                         ret = tree->ops->readpage_io_hook(page, cur,
2651                                                           cur + iosize - 1);
2652                 }
2653                 if (!ret) {
2654                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2655                         pnr -= page->index;
2656                         ret = submit_extent_page(READ, tree, page,
2657                                          sector, disk_io_size, pg_offset,
2658                                          bdev, bio, pnr,
2659                                          end_bio_extent_readpage, mirror_num,
2660                                          *bio_flags,
2661                                          this_bio_flag);
2662                         nr++;
2663                         *bio_flags = this_bio_flag;
2664                 }
2665                 if (ret)
2666                         SetPageError(page);
2667                 cur = cur + iosize;
2668                 pg_offset += iosize;
2669         }
2670 out:
2671         if (!nr) {
2672                 if (!PageError(page))
2673                         SetPageUptodate(page);
2674                 unlock_page(page);
2675         }
2676         return 0;
2677 }
2678
2679 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2680                             get_extent_t *get_extent, int mirror_num)
2681 {
2682         struct bio *bio = NULL;
2683         unsigned long bio_flags = 0;
2684         int ret;
2685
2686         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
2687                                       &bio_flags);
2688         if (bio)
2689                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
2690         return ret;
2691 }
2692
2693 static noinline void update_nr_written(struct page *page,
2694                                       struct writeback_control *wbc,
2695                                       unsigned long nr_written)
2696 {
2697         wbc->nr_to_write -= nr_written;
2698         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2699             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2700                 page->mapping->writeback_index = page->index + nr_written;
2701 }
2702
2703 /*
2704  * the writepage semantics are similar to regular writepage.  extent
2705  * records are inserted to lock ranges in the tree, and as dirty areas
2706  * are found, they are marked writeback.  Then the lock bits are removed
2707  * and the end_io handler clears the writeback ranges
2708  */
2709 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2710                               void *data)
2711 {
2712         struct inode *inode = page->mapping->host;
2713         struct extent_page_data *epd = data;
2714         struct extent_io_tree *tree = epd->tree;
2715         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2716         u64 delalloc_start;
2717         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2718         u64 end;
2719         u64 cur = start;
2720         u64 extent_offset;
2721         u64 last_byte = i_size_read(inode);
2722         u64 block_start;
2723         u64 iosize;
2724         sector_t sector;
2725         struct extent_state *cached_state = NULL;
2726         struct extent_map *em;
2727         struct block_device *bdev;
2728         int ret;
2729         int nr = 0;
2730         size_t pg_offset = 0;
2731         size_t blocksize;
2732         loff_t i_size = i_size_read(inode);
2733         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2734         u64 nr_delalloc;
2735         u64 delalloc_end;
2736         int page_started;
2737         int compressed;
2738         int write_flags;
2739         unsigned long nr_written = 0;
2740         bool fill_delalloc = true;
2741
2742         if (wbc->sync_mode == WB_SYNC_ALL)
2743                 write_flags = WRITE_SYNC;
2744         else
2745                 write_flags = WRITE;
2746
2747         trace___extent_writepage(page, inode, wbc);
2748
2749         WARN_ON(!PageLocked(page));
2750
2751         ClearPageError(page);
2752
2753         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2754         if (page->index > end_index ||
2755            (page->index == end_index && !pg_offset)) {
2756                 page->mapping->a_ops->invalidatepage(page, 0);
2757                 unlock_page(page);
2758                 return 0;
2759         }
2760
2761         if (page->index == end_index) {
2762                 char *userpage;
2763
2764                 userpage = kmap_atomic(page, KM_USER0);
2765                 memset(userpage + pg_offset, 0,
2766                        PAGE_CACHE_SIZE - pg_offset);
2767                 kunmap_atomic(userpage, KM_USER0);
2768                 flush_dcache_page(page);
2769         }
2770         pg_offset = 0;
2771
2772         set_page_extent_mapped(page);
2773
2774         if (!tree->ops || !tree->ops->fill_delalloc)
2775                 fill_delalloc = false;
2776
2777         delalloc_start = start;
2778         delalloc_end = 0;
2779         page_started = 0;
2780         if (!epd->extent_locked && fill_delalloc) {
2781                 u64 delalloc_to_write = 0;
2782                 /*
2783                  * make sure the wbc mapping index is at least updated
2784                  * to this page.
2785                  */
2786                 update_nr_written(page, wbc, 0);
2787
2788                 while (delalloc_end < page_end) {
2789                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2790                                                        page,
2791                                                        &delalloc_start,
2792                                                        &delalloc_end,
2793                                                        128 * 1024 * 1024);
2794                         if (nr_delalloc == 0) {
2795                                 delalloc_start = delalloc_end + 1;
2796                                 continue;
2797                         }
2798                         ret = tree->ops->fill_delalloc(inode, page,
2799                                                        delalloc_start,
2800                                                        delalloc_end,
2801                                                        &page_started,
2802                                                        &nr_written);
2803                         BUG_ON(ret);
2804                         /*
2805                          * delalloc_end is already one less than the total
2806                          * length, so we don't subtract one from
2807                          * PAGE_CACHE_SIZE
2808                          */
2809                         delalloc_to_write += (delalloc_end - delalloc_start +
2810                                               PAGE_CACHE_SIZE) >>
2811                                               PAGE_CACHE_SHIFT;
2812                         delalloc_start = delalloc_end + 1;
2813                 }
2814                 if (wbc->nr_to_write < delalloc_to_write) {
2815                         int thresh = 8192;
2816
2817                         if (delalloc_to_write < thresh * 2)
2818                                 thresh = delalloc_to_write;
2819                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2820                                                  thresh);
2821                 }
2822
2823                 /* did the fill delalloc function already unlock and start
2824                  * the IO?
2825                  */
2826                 if (page_started) {
2827                         ret = 0;
2828                         /*
2829                          * we've unlocked the page, so we can't update
2830                          * the mapping's writeback index, just update
2831                          * nr_to_write.
2832                          */
2833                         wbc->nr_to_write -= nr_written;
2834                         goto done_unlocked;
2835                 }
2836         }
2837         if (tree->ops && tree->ops->writepage_start_hook) {
2838                 ret = tree->ops->writepage_start_hook(page, start,
2839                                                       page_end);
2840                 if (ret) {
2841                         /* Fixup worker will requeue */
2842                         if (ret == -EBUSY)
2843                                 wbc->pages_skipped++;
2844                         else
2845                                 redirty_page_for_writepage(wbc, page);
2846                         update_nr_written(page, wbc, nr_written);
2847                         unlock_page(page);
2848                         ret = 0;
2849                         goto done_unlocked;
2850                 }
2851         }
2852
2853         /*
2854          * we don't want to touch the inode after unlocking the page,
2855          * so we update the mapping writeback index now
2856          */
2857         update_nr_written(page, wbc, nr_written + 1);
2858
2859         end = page_end;
2860         if (last_byte <= start) {
2861                 if (tree->ops && tree->ops->writepage_end_io_hook)
2862                         tree->ops->writepage_end_io_hook(page, start,
2863                                                          page_end, NULL, 1);
2864                 goto done;
2865         }
2866
2867         blocksize = inode->i_sb->s_blocksize;
2868
2869         while (cur <= end) {
2870                 if (cur >= last_byte) {
2871                         if (tree->ops && tree->ops->writepage_end_io_hook)
2872                                 tree->ops->writepage_end_io_hook(page, cur,
2873                                                          page_end, NULL, 1);
2874                         break;
2875                 }
2876                 em = epd->get_extent(inode, page, pg_offset, cur,
2877                                      end - cur + 1, 1);
2878                 if (IS_ERR_OR_NULL(em)) {
2879                         SetPageError(page);
2880                         break;
2881                 }
2882
2883                 extent_offset = cur - em->start;
2884                 BUG_ON(extent_map_end(em) <= cur);
2885                 BUG_ON(end < cur);
2886                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2887                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2888                 sector = (em->block_start + extent_offset) >> 9;
2889                 bdev = em->bdev;
2890                 block_start = em->block_start;
2891                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2892                 free_extent_map(em);
2893                 em = NULL;
2894
2895                 /*
2896                  * compressed and inline extents are written through other
2897                  * paths in the FS
2898                  */
2899                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2900                     block_start == EXTENT_MAP_INLINE) {
2901                         /*
2902                          * end_io notification does not happen here for
2903                          * compressed extents
2904                          */
2905                         if (!compressed && tree->ops &&
2906                             tree->ops->writepage_end_io_hook)
2907                                 tree->ops->writepage_end_io_hook(page, cur,
2908                                                          cur + iosize - 1,
2909                                                          NULL, 1);
2910                         else if (compressed) {
2911                                 /* we don't want to end_page_writeback on
2912                                  * a compressed extent.  this happens
2913                                  * elsewhere
2914                                  */
2915                                 nr++;
2916                         }
2917
2918                         cur += iosize;
2919                         pg_offset += iosize;
2920                         continue;
2921                 }
2922                 /* leave this out until we have a page_mkwrite call */
2923                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2924                                    EXTENT_DIRTY, 0, NULL)) {
2925                         cur = cur + iosize;
2926                         pg_offset += iosize;
2927                         continue;
2928                 }
2929
2930                 if (tree->ops && tree->ops->writepage_io_hook) {
2931                         ret = tree->ops->writepage_io_hook(page, cur,
2932                                                 cur + iosize - 1);
2933                 } else {
2934                         ret = 0;
2935                 }
2936                 if (ret) {
2937                         SetPageError(page);
2938                 } else {
2939                         unsigned long max_nr = end_index + 1;
2940
2941                         set_range_writeback(tree, cur, cur + iosize - 1);
2942                         if (!PageWriteback(page)) {
2943                                 printk(KERN_ERR "btrfs warning page %lu not "
2944                                        "writeback, cur %llu end %llu\n",
2945                                        page->index, (unsigned long long)cur,
2946                                        (unsigned long long)end);
2947                         }
2948
2949                         ret = submit_extent_page(write_flags, tree, page,
2950                                                  sector, iosize, pg_offset,
2951                                                  bdev, &epd->bio, max_nr,
2952                                                  end_bio_extent_writepage,
2953                                                  0, 0, 0);
2954                         if (ret)
2955                                 SetPageError(page);
2956                 }
2957                 cur = cur + iosize;
2958                 pg_offset += iosize;
2959                 nr++;
2960         }
2961 done:
2962         if (nr == 0) {
2963                 /* make sure the mapping tag for page dirty gets cleared */
2964                 set_page_writeback(page);
2965                 end_page_writeback(page);
2966         }
2967         unlock_page(page);
2968
2969 done_unlocked:
2970
2971         /* drop our reference on any cached states */
2972         free_extent_state(cached_state);
2973         return 0;
2974 }
2975
2976 /**
2977  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2978  * @mapping: address space structure to write
2979  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2980  * @writepage: function called for each page
2981  * @data: data passed to writepage function
2982  *
2983  * If a page is already under I/O, write_cache_pages() skips it, even
2984  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2985  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2986  * and msync() need to guarantee that all the data which was dirty at the time
2987  * the call was made get new I/O started against them.  If wbc->sync_mode is
2988  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2989  * existing IO to complete.
2990  */
2991 static int extent_write_cache_pages(struct extent_io_tree *tree,
2992                              struct address_space *mapping,
2993                              struct writeback_control *wbc,
2994                              writepage_t writepage, void *data,
2995                              void (*flush_fn)(void *))
2996 {
2997         int ret = 0;
2998         int done = 0;
2999         int nr_to_write_done = 0;
3000         struct pagevec pvec;
3001         int nr_pages;
3002         pgoff_t index;
3003         pgoff_t end;            /* Inclusive */
3004         int scanned = 0;
3005         int tag;
3006
3007         pagevec_init(&pvec, 0);
3008         if (wbc->range_cyclic) {
3009                 index = mapping->writeback_index; /* Start from prev offset */
3010                 end = -1;
3011         } else {
3012                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3013                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3014                 scanned = 1;
3015         }
3016         if (wbc->sync_mode == WB_SYNC_ALL)
3017                 tag = PAGECACHE_TAG_TOWRITE;
3018         else
3019                 tag = PAGECACHE_TAG_DIRTY;
3020 retry:
3021         if (wbc->sync_mode == WB_SYNC_ALL)
3022                 tag_pages_for_writeback(mapping, index, end);
3023         while (!done && !nr_to_write_done && (index <= end) &&
3024                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3025                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3026                 unsigned i;
3027
3028                 scanned = 1;
3029                 for (i = 0; i < nr_pages; i++) {
3030                         struct page *page = pvec.pages[i];
3031
3032                         /*
3033                          * At this point we hold neither mapping->tree_lock nor
3034                          * lock on the page itself: the page may be truncated or
3035                          * invalidated (changing page->mapping to NULL), or even
3036                          * swizzled back from swapper_space to tmpfs file
3037                          * mapping
3038                          */
3039                         if (tree->ops &&
3040                             tree->ops->write_cache_pages_lock_hook) {
3041                                 tree->ops->write_cache_pages_lock_hook(page,
3042                                                                data, flush_fn);
3043                         } else {
3044                                 if (!trylock_page(page)) {
3045                                         flush_fn(data);
3046                                         lock_page(page);
3047                                 }
3048                         }
3049
3050                         if (unlikely(page->mapping != mapping)) {
3051                                 unlock_page(page);
3052                                 continue;
3053                         }
3054
3055                         if (!wbc->range_cyclic && page->index > end) {
3056                                 done = 1;
3057                                 unlock_page(page);
3058                                 continue;
3059                         }
3060
3061                         if (wbc->sync_mode != WB_SYNC_NONE) {
3062                                 if (PageWriteback(page))
3063                                         flush_fn(data);
3064                                 wait_on_page_writeback(page);
3065                         }
3066
3067                         if (PageWriteback(page) ||
3068                             !clear_page_dirty_for_io(page)) {
3069                                 unlock_page(page);
3070                                 continue;
3071                         }
3072
3073                         ret = (*writepage)(page, wbc, data);
3074
3075                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3076                                 unlock_page(page);
3077                                 ret = 0;
3078                         }
3079                         if (ret)
3080                                 done = 1;
3081
3082                         /*
3083                          * the filesystem may choose to bump up nr_to_write.
3084                          * We have to make sure to honor the new nr_to_write
3085                          * at any time
3086                          */
3087                         nr_to_write_done = wbc->nr_to_write <= 0;
3088                 }
3089                 pagevec_release(&pvec);
3090                 cond_resched();
3091         }
3092         if (!scanned && !done) {
3093                 /*
3094                  * We hit the last page and there is more work to be done: wrap
3095                  * back to the start of the file
3096                  */
3097                 scanned = 1;
3098                 index = 0;
3099                 goto retry;
3100         }
3101         return ret;
3102 }
3103
3104 static void flush_epd_write_bio(struct extent_page_data *epd)
3105 {
3106         if (epd->bio) {
3107                 if (epd->sync_io)
3108                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
3109                 else
3110                         submit_one_bio(WRITE, epd->bio, 0, 0);
3111                 epd->bio = NULL;
3112         }
3113 }
3114
3115 static noinline void flush_write_bio(void *data)
3116 {
3117         struct extent_page_data *epd = data;
3118         flush_epd_write_bio(epd);
3119 }
3120
3121 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3122                           get_extent_t *get_extent,
3123                           struct writeback_control *wbc)
3124 {
3125         int ret;
3126         struct extent_page_data epd = {
3127                 .bio = NULL,
3128                 .tree = tree,
3129                 .get_extent = get_extent,
3130                 .extent_locked = 0,
3131                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3132         };
3133
3134         ret = __extent_writepage(page, wbc, &epd);
3135
3136         flush_epd_write_bio(&epd);
3137         return ret;
3138 }
3139
3140 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3141                               u64 start, u64 end, get_extent_t *get_extent,
3142                               int mode)
3143 {
3144         int ret = 0;
3145         struct address_space *mapping = inode->i_mapping;
3146         struct page *page;
3147         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3148                 PAGE_CACHE_SHIFT;
3149
3150         struct extent_page_data epd = {
3151                 .bio = NULL,
3152                 .tree = tree,
3153                 .get_extent = get_extent,
3154                 .extent_locked = 1,
3155                 .sync_io = mode == WB_SYNC_ALL,
3156         };
3157         struct writeback_control wbc_writepages = {
3158                 .sync_mode      = mode,
3159                 .nr_to_write    = nr_pages * 2,
3160                 .range_start    = start,
3161                 .range_end      = end + 1,
3162         };
3163
3164         while (start <= end) {
3165                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3166                 if (clear_page_dirty_for_io(page))
3167                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3168                 else {
3169                         if (tree->ops && tree->ops->writepage_end_io_hook)
3170                                 tree->ops->writepage_end_io_hook(page, start,
3171                                                  start + PAGE_CACHE_SIZE - 1,
3172                                                  NULL, 1);
3173                         unlock_page(page);
3174                 }
3175                 page_cache_release(page);
3176                 start += PAGE_CACHE_SIZE;
3177         }
3178
3179         flush_epd_write_bio(&epd);
3180         return ret;
3181 }
3182
3183 int extent_writepages(struct extent_io_tree *tree,
3184                       struct address_space *mapping,
3185                       get_extent_t *get_extent,
3186                       struct writeback_control *wbc)
3187 {
3188         int ret = 0;
3189         struct extent_page_data epd = {
3190                 .bio = NULL,
3191                 .tree = tree,
3192                 .get_extent = get_extent,
3193                 .extent_locked = 0,
3194                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3195         };
3196
3197         ret = extent_write_cache_pages(tree, mapping, wbc,
3198                                        __extent_writepage, &epd,
3199                                        flush_write_bio);
3200         flush_epd_write_bio(&epd);
3201         return ret;
3202 }
3203
3204 int extent_readpages(struct extent_io_tree *tree,
3205                      struct address_space *mapping,
3206                      struct list_head *pages, unsigned nr_pages,
3207                      get_extent_t get_extent)
3208 {
3209         struct bio *bio = NULL;
3210         unsigned page_idx;
3211         unsigned long bio_flags = 0;
3212
3213         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3214                 struct page *page = list_entry(pages->prev, struct page, lru);
3215
3216                 prefetchw(&page->flags);
3217                 list_del(&page->lru);
3218                 if (!add_to_page_cache_lru(page, mapping,
3219                                         page->index, GFP_NOFS)) {
3220                         __extent_read_full_page(tree, page, get_extent,
3221                                                 &bio, 0, &bio_flags);
3222                 }
3223                 page_cache_release(page);
3224         }
3225         BUG_ON(!list_empty(pages));
3226         if (bio)
3227                 submit_one_bio(READ, bio, 0, bio_flags);
3228         return 0;
3229 }
3230
3231 /*
3232  * basic invalidatepage code, this waits on any locked or writeback
3233  * ranges corresponding to the page, and then deletes any extent state
3234  * records from the tree
3235  */
3236 int extent_invalidatepage(struct extent_io_tree *tree,
3237                           struct page *page, unsigned long offset)
3238 {
3239         struct extent_state *cached_state = NULL;
3240         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
3241         u64 end = start + PAGE_CACHE_SIZE - 1;
3242         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3243
3244         start += (offset + blocksize - 1) & ~(blocksize - 1);
3245         if (start > end)
3246                 return 0;
3247
3248         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
3249         wait_on_page_writeback(page);
3250         clear_extent_bit(tree, start, end,
3251                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3252                          EXTENT_DO_ACCOUNTING,
3253                          1, 1, &cached_state, GFP_NOFS);
3254         return 0;
3255 }
3256
3257 /*
3258  * a helper for releasepage, this tests for areas of the page that
3259  * are locked or under IO and drops the related state bits if it is safe
3260  * to drop the page.
3261  */
3262 int try_release_extent_state(struct extent_map_tree *map,
3263                              struct extent_io_tree *tree, struct page *page,
3264                              gfp_t mask)
3265 {
3266         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3267         u64 end = start + PAGE_CACHE_SIZE - 1;
3268         int ret = 1;
3269
3270         if (test_range_bit(tree, start, end,
3271                            EXTENT_IOBITS, 0, NULL))
3272                 ret = 0;
3273         else {
3274                 if ((mask & GFP_NOFS) == GFP_NOFS)
3275                         mask = GFP_NOFS;
3276                 /*
3277                  * at this point we can safely clear everything except the
3278                  * locked bit and the nodatasum bit
3279                  */
3280                 ret = clear_extent_bit(tree, start, end,
3281                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3282                                  0, 0, NULL, mask);
3283
3284                 /* if clear_extent_bit failed for enomem reasons,
3285                  * we can't allow the release to continue.
3286                  */
3287                 if (ret < 0)
3288                         ret = 0;
3289                 else
3290                         ret = 1;
3291         }
3292         return ret;
3293 }
3294
3295 /*
3296  * a helper for releasepage.  As long as there are no locked extents
3297  * in the range corresponding to the page, both state records and extent
3298  * map records are removed
3299  */
3300 int try_release_extent_mapping(struct extent_map_tree *map,
3301                                struct extent_io_tree *tree, struct page *page,
3302                                gfp_t mask)
3303 {
3304         struct extent_map *em;
3305         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3306         u64 end = start + PAGE_CACHE_SIZE - 1;
3307
3308         if ((mask & __GFP_WAIT) &&
3309             page->mapping->host->i_size > 16 * 1024 * 1024) {
3310                 u64 len;
3311                 while (start <= end) {
3312                         len = end - start + 1;
3313                         write_lock(&map->lock);
3314                         em = lookup_extent_mapping(map, start, len);
3315                         if (!em) {
3316                                 write_unlock(&map->lock);
3317                                 break;
3318                         }
3319                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3320                             em->start != start) {
3321                                 write_unlock(&map->lock);
3322                                 free_extent_map(em);
3323                                 break;
3324                         }
3325                         if (!test_range_bit(tree, em->start,
3326                                             extent_map_end(em) - 1,
3327                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
3328                                             0, NULL)) {
3329                                 remove_extent_mapping(map, em);
3330                                 /* once for the rb tree */
3331                                 free_extent_map(em);
3332                         }
3333                         start = extent_map_end(em);
3334                         write_unlock(&map->lock);
3335
3336                         /* once for us */
3337                         free_extent_map(em);
3338                 }
3339         }
3340         return try_release_extent_state(map, tree, page, mask);
3341 }
3342
3343 /*
3344  * helper function for fiemap, which doesn't want to see any holes.
3345  * This maps until we find something past 'last'
3346  */
3347 static struct extent_map *get_extent_skip_holes(struct inode *inode,
3348                                                 u64 offset,
3349                                                 u64 last,
3350                                                 get_extent_t *get_extent)
3351 {
3352         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
3353         struct extent_map *em;
3354         u64 len;
3355
3356         if (offset >= last)
3357                 return NULL;
3358
3359         while(1) {
3360                 len = last - offset;
3361                 if (len == 0)
3362                         break;
3363                 len = (len + sectorsize - 1) & ~(sectorsize - 1);
3364                 em = get_extent(inode, NULL, 0, offset, len, 0);
3365                 if (IS_ERR_OR_NULL(em))
3366                         return em;
3367
3368                 /* if this isn't a hole return it */
3369                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
3370                     em->block_start != EXTENT_MAP_HOLE) {
3371                         return em;
3372                 }
3373
3374                 /* this is a hole, advance to the next extent */
3375                 offset = extent_map_end(em);
3376                 free_extent_map(em);
3377                 if (offset >= last)
3378                         break;
3379         }
3380         return NULL;
3381 }
3382
3383 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
3384                 __u64 start, __u64 len, get_extent_t *get_extent)
3385 {
3386         int ret = 0;
3387         u64 off = start;
3388         u64 max = start + len;
3389         u32 flags = 0;
3390         u32 found_type;
3391         u64 last;
3392         u64 last_for_get_extent = 0;
3393         u64 disko = 0;
3394         u64 isize = i_size_read(inode);
3395         struct btrfs_key found_key;
3396         struct extent_map *em = NULL;
3397         struct extent_state *cached_state = NULL;
3398         struct btrfs_path *path;
3399         struct btrfs_file_extent_item *item;
3400         int end = 0;
3401         u64 em_start = 0;
3402         u64 em_len = 0;
3403         u64 em_end = 0;
3404         unsigned long emflags;
3405
3406         if (len == 0)
3407                 return -EINVAL;
3408
3409         path = btrfs_alloc_path();
3410         if (!path)
3411                 return -ENOMEM;
3412         path->leave_spinning = 1;
3413
3414         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
3415         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
3416
3417         /*
3418          * lookup the last file extent.  We're not using i_size here
3419          * because there might be preallocation past i_size
3420          */
3421         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3422                                        path, btrfs_ino(inode), -1, 0);
3423         if (ret < 0) {
3424                 btrfs_free_path(path);
3425                 return ret;
3426         }
3427         WARN_ON(!ret);
3428         path->slots[0]--;
3429         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3430                               struct btrfs_file_extent_item);
3431         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3432         found_type = btrfs_key_type(&found_key);
3433
3434         /* No extents, but there might be delalloc bits */
3435         if (found_key.objectid != btrfs_ino(inode) ||
3436             found_type != BTRFS_EXTENT_DATA_KEY) {
3437                 /* have to trust i_size as the end */
3438                 last = (u64)-1;
3439                 last_for_get_extent = isize;
3440         } else {
3441                 /*
3442                  * remember the start of the last extent.  There are a
3443                  * bunch of different factors that go into the length of the
3444                  * extent, so its much less complex to remember where it started
3445                  */
3446                 last = found_key.offset;
3447                 last_for_get_extent = last + 1;
3448         }
3449         btrfs_free_path(path);
3450
3451         /*
3452          * we might have some extents allocated but more delalloc past those
3453          * extents.  so, we trust isize unless the start of the last extent is
3454          * beyond isize
3455          */
3456         if (last < isize) {
3457                 last = (u64)-1;
3458                 last_for_get_extent = isize;
3459         }
3460
3461         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3462                          &cached_state, GFP_NOFS);
3463
3464         em = get_extent_skip_holes(inode, start, last_for_get_extent,
3465                                    get_extent);
3466         if (!em)
3467                 goto out;
3468         if (IS_ERR(em)) {
3469                 ret = PTR_ERR(em);
3470                 goto out;
3471         }
3472
3473         while (!end) {
3474                 u64 offset_in_extent;
3475
3476                 /* break if the extent we found is outside the range */
3477                 if (em->start >= max || extent_map_end(em) < off)
3478                         break;
3479
3480                 /*
3481                  * get_extent may return an extent that starts before our
3482                  * requested range.  We have to make sure the ranges
3483                  * we return to fiemap always move forward and don't
3484                  * overlap, so adjust the offsets here
3485                  */
3486                 em_start = max(em->start, off);
3487
3488                 /*
3489                  * record the offset from the start of the extent
3490                  * for adjusting the disk offset below
3491                  */
3492                 offset_in_extent = em_start - em->start;
3493                 em_end = extent_map_end(em);
3494                 em_len = em_end - em_start;
3495                 emflags = em->flags;
3496                 disko = 0;
3497                 flags = 0;
3498
3499                 /*
3500                  * bump off for our next call to get_extent
3501                  */
3502                 off = extent_map_end(em);
3503                 if (off >= max)
3504                         end = 1;
3505
3506                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3507                         end = 1;
3508                         flags |= FIEMAP_EXTENT_LAST;
3509                 } else if (em->block_start == EXTENT_MAP_INLINE) {
3510                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
3511                                   FIEMAP_EXTENT_NOT_ALIGNED);
3512                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3513                         flags |= (FIEMAP_EXTENT_DELALLOC |
3514                                   FIEMAP_EXTENT_UNKNOWN);
3515                 } else {
3516                         disko = em->block_start + offset_in_extent;
3517                 }
3518                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3519                         flags |= FIEMAP_EXTENT_ENCODED;
3520
3521                 free_extent_map(em);
3522                 em = NULL;
3523                 if ((em_start >= last) || em_len == (u64)-1 ||
3524                    (last == (u64)-1 && isize <= em_end)) {
3525                         flags |= FIEMAP_EXTENT_LAST;
3526                         end = 1;
3527                 }
3528
3529                 /* now scan forward to see if this is really the last extent. */
3530                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
3531                                            get_extent);
3532                 if (IS_ERR(em)) {
3533                         ret = PTR_ERR(em);
3534                         goto out;
3535                 }
3536                 if (!em) {
3537                         flags |= FIEMAP_EXTENT_LAST;
3538                         end = 1;
3539                 }
3540                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3541                                               em_len, flags);
3542                 if (ret)
3543                         goto out_free;
3544         }
3545 out_free:
3546         free_extent_map(em);
3547 out:
3548         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3549                              &cached_state, GFP_NOFS);
3550         return ret;
3551 }
3552
3553 inline struct page *extent_buffer_page(struct extent_buffer *eb,
3554                                               unsigned long i)
3555 {
3556         return eb->pages[i];
3557 }
3558
3559 inline unsigned long num_extent_pages(u64 start, u64 len)
3560 {
3561         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3562                 (start >> PAGE_CACHE_SHIFT);
3563 }
3564
3565 static void __free_extent_buffer(struct extent_buffer *eb)
3566 {
3567 #if LEAK_DEBUG
3568         unsigned long flags;
3569         spin_lock_irqsave(&leak_lock, flags);
3570         list_del(&eb->leak_list);
3571         spin_unlock_irqrestore(&leak_lock, flags);
3572 #endif
3573         if (eb->pages && eb->pages != eb->inline_pages)
3574                 kfree(eb->pages);
3575         kmem_cache_free(extent_buffer_cache, eb);
3576 }
3577
3578 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3579                                                    u64 start,
3580                                                    unsigned long len,
3581                                                    gfp_t mask)
3582 {
3583         struct extent_buffer *eb = NULL;
3584 #if LEAK_DEBUG
3585         unsigned long flags;
3586 #endif
3587
3588         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3589         if (eb == NULL)
3590                 return NULL;
3591         eb->start = start;
3592         eb->len = len;
3593         eb->tree = tree;
3594         rwlock_init(&eb->lock);
3595         atomic_set(&eb->write_locks, 0);
3596         atomic_set(&eb->read_locks, 0);
3597         atomic_set(&eb->blocking_readers, 0);
3598         atomic_set(&eb->blocking_writers, 0);
3599         atomic_set(&eb->spinning_readers, 0);
3600         atomic_set(&eb->spinning_writers, 0);
3601         eb->lock_nested = 0;
3602         init_waitqueue_head(&eb->write_lock_wq);
3603         init_waitqueue_head(&eb->read_lock_wq);
3604
3605 #if LEAK_DEBUG
3606         spin_lock_irqsave(&leak_lock, flags);
3607         list_add(&eb->leak_list, &buffers);
3608         spin_unlock_irqrestore(&leak_lock, flags);
3609 #endif
3610         spin_lock_init(&eb->refs_lock);
3611         atomic_set(&eb->refs, 1);
3612         atomic_set(&eb->pages_reading, 0);
3613
3614         if (len > MAX_INLINE_EXTENT_BUFFER_SIZE) {
3615                 struct page **pages;
3616                 int num_pages = (len + PAGE_CACHE_SIZE - 1) >>
3617                         PAGE_CACHE_SHIFT;
3618                 pages = kzalloc(num_pages, mask);
3619                 if (!pages) {
3620                         __free_extent_buffer(eb);
3621                         return NULL;
3622                 }
3623                 eb->pages = pages;
3624         } else {
3625                 eb->pages = eb->inline_pages;
3626         }
3627
3628         return eb;
3629 }
3630
3631 /*
3632  * Helper for releasing extent buffer page.
3633  */
3634 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3635                                                 unsigned long start_idx)
3636 {
3637         unsigned long index;
3638         struct page *page;
3639
3640         index = num_extent_pages(eb->start, eb->len);
3641         if (start_idx >= index)
3642                 return;
3643
3644         do {
3645                 index--;
3646                 page = extent_buffer_page(eb, index);
3647                 if (page) {
3648                         spin_lock(&page->mapping->private_lock);
3649                         /*
3650                          * We do this since we'll remove the pages after we've
3651                          * removed the eb from the radix tree, so we could race
3652                          * and have this page now attached to the new eb.  So
3653                          * only clear page_private if it's still connected to
3654                          * this eb.
3655                          */
3656                         if (PagePrivate(page) &&
3657                             page->private == (unsigned long)eb) {
3658                                 BUG_ON(PageDirty(page));
3659                                 BUG_ON(PageWriteback(page));
3660                                 /*
3661                                  * We need to make sure we haven't be attached
3662                                  * to a new eb.
3663                                  */
3664                                 ClearPagePrivate(page);
3665                                 set_page_private(page, 0);
3666                                 /* One for the page private */
3667                                 page_cache_release(page);
3668                         }
3669                         spin_unlock(&page->mapping->private_lock);
3670
3671                         /* One for when we alloced the page */
3672                         page_cache_release(page);
3673                 }
3674         } while (index != start_idx);
3675 }
3676
3677 /*
3678  * Helper for releasing the extent buffer.
3679  */
3680 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3681 {
3682         btrfs_release_extent_buffer_page(eb, 0);
3683         __free_extent_buffer(eb);
3684 }
3685
3686 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
3687 {
3688         unsigned long num_pages, i;
3689
3690         num_pages = num_extent_pages(eb->start, eb->len);
3691         for (i = 0; i < num_pages; i++) {
3692                 struct page *p = extent_buffer_page(eb, i);
3693                 mark_page_accessed(p);
3694         }
3695 }
3696
3697 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3698                                           u64 start, unsigned long len)
3699 {
3700         unsigned long num_pages = num_extent_pages(start, len);
3701         unsigned long i;
3702         unsigned long index = start >> PAGE_CACHE_SHIFT;
3703         struct extent_buffer *eb;
3704         struct extent_buffer *exists = NULL;
3705         struct page *p;
3706         struct address_space *mapping = tree->mapping;
3707         int uptodate = 1;
3708         int ret;
3709
3710         rcu_read_lock();
3711         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3712         if (eb && atomic_inc_not_zero(&eb->refs)) {
3713                 rcu_read_unlock();
3714                 mark_extent_buffer_accessed(eb);
3715                 return eb;
3716         }
3717         rcu_read_unlock();
3718
3719         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
3720         if (!eb)
3721                 return NULL;
3722
3723         for (i = 0; i < num_pages; i++, index++) {
3724                 p = find_or_create_page(mapping, index, GFP_NOFS);
3725                 if (!p) {
3726                         WARN_ON(1);
3727                         goto free_eb;
3728                 }
3729
3730                 spin_lock(&mapping->private_lock);
3731                 if (PagePrivate(p)) {
3732                         /*
3733                          * We could have already allocated an eb for this page
3734                          * and attached one so lets see if we can get a ref on
3735                          * the existing eb, and if we can we know it's good and
3736                          * we can just return that one, else we know we can just
3737                          * overwrite page->private.
3738                          */
3739                         exists = (struct extent_buffer *)p->private;
3740                         if (atomic_inc_not_zero(&exists->refs)) {
3741                                 spin_unlock(&mapping->private_lock);
3742                                 unlock_page(p);
3743                                 mark_extent_buffer_accessed(exists);
3744                                 goto free_eb;
3745                         }
3746
3747                         /*
3748                          * Do this so attach doesn't complain and we need to
3749                          * drop the ref the old guy had.
3750                          */
3751                         ClearPagePrivate(p);
3752                         page_cache_release(p);
3753                 }
3754                 attach_extent_buffer_page(eb, p);
3755                 spin_unlock(&mapping->private_lock);
3756                 mark_page_accessed(p);
3757                 eb->pages[i] = p;
3758                 if (!PageUptodate(p))
3759                         uptodate = 0;
3760
3761                 /*
3762                  * see below about how we avoid a nasty race with release page
3763                  * and why we unlock later
3764                  */
3765         }
3766         if (uptodate)
3767                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3768 again:
3769         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
3770         if (ret)
3771                 goto free_eb;
3772
3773         spin_lock(&tree->buffer_lock);
3774         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
3775         if (ret == -EEXIST) {
3776                 exists = radix_tree_lookup(&tree->buffer,
3777                                                 start >> PAGE_CACHE_SHIFT);
3778                 if (!atomic_inc_not_zero(&exists->refs)) {
3779                         spin_unlock(&tree->buffer_lock);
3780                         radix_tree_preload_end();
3781                         exists = NULL;
3782                         goto again;
3783                 }
3784                 spin_unlock(&tree->buffer_lock);
3785                 radix_tree_preload_end();
3786                 mark_extent_buffer_accessed(exists);
3787                 goto free_eb;
3788         }
3789         /* add one reference for the tree */
3790         spin_lock(&eb->refs_lock);
3791         atomic_inc(&eb->refs);
3792         set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags);
3793         spin_unlock(&eb->refs_lock);
3794         spin_unlock(&tree->buffer_lock);
3795         radix_tree_preload_end();
3796
3797         /*
3798          * there is a race where release page may have
3799          * tried to find this extent buffer in the radix
3800          * but failed.  It will tell the VM it is safe to
3801          * reclaim the, and it will clear the page private bit.
3802          * We must make sure to set the page private bit properly
3803          * after the extent buffer is in the radix tree so
3804          * it doesn't get lost
3805          */
3806         SetPageChecked(eb->pages[0]);
3807         for (i = 1; i < num_pages; i++) {
3808                 p = extent_buffer_page(eb, i);
3809                 ClearPageChecked(p);
3810                 unlock_page(p);
3811         }
3812         unlock_page(eb->pages[0]);
3813         return eb;
3814
3815 free_eb:
3816         for (i = 0; i < num_pages; i++) {
3817                 if (eb->pages[i])
3818                         unlock_page(eb->pages[i]);
3819         }
3820
3821         if (!atomic_dec_and_test(&eb->refs))
3822                 return exists;
3823         btrfs_release_extent_buffer(eb);
3824         return exists;
3825 }
3826
3827 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3828                                          u64 start, unsigned long len)
3829 {
3830         struct extent_buffer *eb;
3831
3832         rcu_read_lock();
3833         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3834         if (eb && atomic_inc_not_zero(&eb->refs)) {
3835                 rcu_read_unlock();
3836                 mark_extent_buffer_accessed(eb);
3837                 return eb;
3838         }
3839         rcu_read_unlock();
3840
3841         return NULL;
3842 }
3843
3844 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3845 {
3846         struct extent_buffer *eb =
3847                         container_of(head, struct extent_buffer, rcu_head);
3848
3849         __free_extent_buffer(eb);
3850 }
3851
3852 static int extent_buffer_under_io(struct extent_buffer *eb,
3853                                   struct page *locked_page)
3854 {
3855         unsigned long num_pages, i;
3856
3857         num_pages = num_extent_pages(eb->start, eb->len);
3858         for (i = 0; i < num_pages; i++) {
3859                 struct page *page = eb->pages[i];
3860                 int need_unlock = 0;
3861
3862                 if (!page)
3863                         continue;
3864
3865                 if (page != locked_page) {
3866                         if (!trylock_page(page))
3867                                 return 1;
3868                         need_unlock = 1;
3869                 }
3870
3871                 if (PageDirty(page) || PageWriteback(page)) {
3872                         if (need_unlock)
3873                                 unlock_page(page);
3874                         return 1;
3875                 }
3876                 if (need_unlock)
3877                         unlock_page(page);
3878         }
3879
3880         return 0;
3881 }
3882
3883 /* Expects to have eb->eb_lock already held */
3884 static void release_extent_buffer(struct extent_buffer *eb, gfp_t mask)
3885 {
3886         WARN_ON(atomic_read(&eb->refs) == 0);
3887         if (atomic_dec_and_test(&eb->refs)) {
3888                 struct extent_io_tree *tree = eb->tree;
3889                 int ret;
3890
3891                 spin_unlock(&eb->refs_lock);
3892
3893                 might_sleep_if(mask & __GFP_WAIT);
3894                 ret = clear_extent_bit(tree, eb->start,
3895                                        eb->start + eb->len - 1, -1, 0, 0,
3896                                        NULL, mask);
3897                 if (ret < 0) {
3898                         unsigned long num_pages, i;
3899
3900                         num_pages = num_extent_pages(eb->start, eb->len);
3901                         /*
3902                          * We failed to clear the state bits which likely means
3903                          * ENOMEM, so just re-up the eb ref and continue, we
3904                          * will get freed later on via releasepage or something
3905                          * else and will be ok.
3906                          */
3907                         spin_lock(&eb->tree->mapping->private_lock);
3908                         spin_lock(&eb->refs_lock);
3909                         set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags);
3910                         atomic_inc(&eb->refs);
3911
3912                         /*
3913                          * We may have started to reclaim the pages for a newly
3914                          * allocated eb, make sure we own all of them again.
3915                          */
3916                         for (i = 0; i < num_pages; i++) {
3917                                 struct page *page = eb->pages[i];
3918
3919                                 if (!page) {
3920                                         WARN_ON(1);
3921                                         continue;
3922                                 }
3923
3924                                 BUG_ON(!PagePrivate(page));
3925                                 if (page->private != (unsigned long)eb) {
3926                                         ClearPagePrivate(page);
3927                                         page_cache_release(page);
3928                                         attach_extent_buffer_page(eb, page);
3929                                 }
3930                         }
3931                         spin_unlock(&eb->refs_lock);
3932                         spin_unlock(&eb->tree->mapping->private_lock);
3933                         return;
3934                 }
3935
3936                 spin_lock(&tree->buffer_lock);
3937                 radix_tree_delete(&tree->buffer,
3938                                   eb->start >> PAGE_CACHE_SHIFT);
3939                 spin_unlock(&tree->buffer_lock);
3940
3941                 /* Should be safe to release our pages at this point */
3942                 btrfs_release_extent_buffer_page(eb, 0);
3943
3944                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3945                 return;
3946         }
3947         spin_unlock(&eb->refs_lock);
3948 }
3949
3950 void free_extent_buffer(struct extent_buffer *eb)
3951 {
3952         if (!eb)
3953                 return;
3954
3955         spin_lock(&eb->refs_lock);
3956         if (atomic_read(&eb->refs) == 2 &&
3957             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
3958             !extent_buffer_under_io(eb, NULL) &&
3959             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3960                 atomic_dec(&eb->refs);
3961
3962         /*
3963          * I know this is terrible, but it's temporary until we stop tracking
3964          * the uptodate bits and such for the extent buffers.
3965          */
3966         release_extent_buffer(eb, GFP_ATOMIC);
3967 }
3968
3969 void free_extent_buffer_stale(struct extent_buffer *eb)
3970 {
3971         if (!eb)
3972                 return;
3973
3974         spin_lock(&eb->refs_lock);
3975         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
3976
3977         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb, NULL) &&
3978             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3979                 atomic_dec(&eb->refs);
3980         release_extent_buffer(eb, GFP_NOFS);
3981 }
3982
3983 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3984                               struct extent_buffer *eb)
3985 {
3986         unsigned long i;
3987         unsigned long num_pages;
3988         struct page *page;
3989
3990         num_pages = num_extent_pages(eb->start, eb->len);
3991
3992         for (i = 0; i < num_pages; i++) {
3993                 page = extent_buffer_page(eb, i);
3994                 if (!PageDirty(page))
3995                         continue;
3996
3997                 lock_page(page);
3998                 WARN_ON(!PagePrivate(page));
3999
4000                 clear_page_dirty_for_io(page);
4001                 spin_lock_irq(&page->mapping->tree_lock);
4002                 if (!PageDirty(page)) {
4003                         radix_tree_tag_clear(&page->mapping->page_tree,
4004                                                 page_index(page),
4005                                                 PAGECACHE_TAG_DIRTY);
4006                 }
4007                 spin_unlock_irq(&page->mapping->tree_lock);
4008                 ClearPageError(page);
4009                 unlock_page(page);
4010         }
4011         return 0;
4012 }
4013
4014 int set_extent_buffer_dirty(struct extent_io_tree *tree,
4015                              struct extent_buffer *eb)
4016 {
4017         unsigned long i;
4018         unsigned long num_pages;
4019         int was_dirty = 0;
4020
4021         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4022         num_pages = num_extent_pages(eb->start, eb->len);
4023         WARN_ON(atomic_read(&eb->refs) == 0);
4024         for (i = 0; i < num_pages; i++)
4025                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
4026         return was_dirty;
4027 }
4028
4029 static int __eb_straddles_pages(u64 start, u64 len)
4030 {
4031         if (len < PAGE_CACHE_SIZE)
4032                 return 1;
4033         if (start & (PAGE_CACHE_SIZE - 1))
4034                 return 1;
4035         if ((start + len) & (PAGE_CACHE_SIZE - 1))
4036                 return 1;
4037         return 0;
4038 }
4039
4040 static int eb_straddles_pages(struct extent_buffer *eb)
4041 {
4042         return __eb_straddles_pages(eb->start, eb->len);
4043 }
4044
4045 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
4046                                 struct extent_buffer *eb,
4047                                 struct extent_state **cached_state)
4048 {
4049         unsigned long i;
4050         struct page *page;
4051         unsigned long num_pages;
4052
4053         num_pages = num_extent_pages(eb->start, eb->len);
4054         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4055
4056         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
4057                               cached_state, GFP_NOFS);
4058
4059         for (i = 0; i < num_pages; i++) {
4060                 page = extent_buffer_page(eb, i);
4061                 if (page)
4062                         ClearPageUptodate(page);
4063         }
4064         return 0;
4065 }
4066
4067 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
4068                                 struct extent_buffer *eb)
4069 {
4070         unsigned long i;
4071         struct page *page;
4072         unsigned long num_pages;
4073
4074         num_pages = num_extent_pages(eb->start, eb->len);
4075
4076         if (eb_straddles_pages(eb)) {
4077                 set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
4078                                     NULL, GFP_NOFS);
4079         }
4080         for (i = 0; i < num_pages; i++) {
4081                 page = extent_buffer_page(eb, i);
4082                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
4083                     ((i == num_pages - 1) &&
4084                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
4085                         check_page_uptodate(tree, page);
4086                         continue;
4087                 }
4088                 SetPageUptodate(page);
4089         }
4090         return 0;
4091 }
4092
4093 int extent_range_uptodate(struct extent_io_tree *tree,
4094                           u64 start, u64 end)
4095 {
4096         struct page *page;
4097         int ret;
4098         int pg_uptodate = 1;
4099         int uptodate;
4100         unsigned long index;
4101
4102         if (__eb_straddles_pages(start, end - start + 1)) {
4103                 ret = test_range_bit(tree, start, end,
4104                                      EXTENT_UPTODATE, 1, NULL);
4105                 if (ret)
4106                         return 1;
4107         }
4108         while (start <= end) {
4109                 index = start >> PAGE_CACHE_SHIFT;
4110                 page = find_get_page(tree->mapping, index);
4111                 if (!page)
4112                         return 1;
4113                 uptodate = PageUptodate(page);
4114                 page_cache_release(page);
4115                 if (!uptodate) {
4116                         pg_uptodate = 0;
4117                         break;
4118                 }
4119                 start += PAGE_CACHE_SIZE;
4120         }
4121         return pg_uptodate;
4122 }
4123
4124 int extent_buffer_uptodate(struct extent_io_tree *tree,
4125                            struct extent_buffer *eb,
4126                            struct extent_state *cached_state)
4127 {
4128         int ret = 0;
4129         unsigned long num_pages;
4130         unsigned long i;
4131         struct page *page;
4132         int pg_uptodate = 1;
4133
4134         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4135                 return 1;
4136
4137         if (eb_straddles_pages(eb)) {
4138                 ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
4139                                    EXTENT_UPTODATE, 1, cached_state);
4140                 if (ret)
4141                         return ret;
4142         }
4143
4144         num_pages = num_extent_pages(eb->start, eb->len);
4145         for (i = 0; i < num_pages; i++) {
4146                 page = extent_buffer_page(eb, i);
4147                 if (!PageUptodate(page)) {
4148                         pg_uptodate = 0;
4149                         break;
4150                 }
4151         }
4152         return pg_uptodate;
4153 }
4154
4155 int read_extent_buffer_pages(struct extent_io_tree *tree,
4156                              struct extent_buffer *eb, u64 start, int wait,
4157                              get_extent_t *get_extent, int mirror_num)
4158 {
4159         unsigned long i;
4160         unsigned long start_i;
4161         struct page *page;
4162         int err;
4163         int ret = 0;
4164         int locked_pages = 0;
4165         int all_uptodate = 1;
4166         unsigned long num_pages;
4167         unsigned long num_reads = 0;
4168         struct bio *bio = NULL;
4169         unsigned long bio_flags = 0;
4170
4171         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4172                 return 0;
4173
4174         if (eb_straddles_pages(eb)) {
4175                 if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
4176                                    EXTENT_UPTODATE, 1, NULL)) {
4177                         return 0;
4178                 }
4179         }
4180
4181         if (start) {
4182                 WARN_ON(start < eb->start);
4183                 start_i = (start >> PAGE_CACHE_SHIFT) -
4184                         (eb->start >> PAGE_CACHE_SHIFT);
4185         } else {
4186                 start_i = 0;
4187         }
4188
4189         num_pages = num_extent_pages(eb->start, eb->len);
4190         for (i = start_i; i < num_pages; i++) {
4191                 page = extent_buffer_page(eb, i);
4192                 if (wait == WAIT_NONE) {
4193                         if (!trylock_page(page))
4194                                 goto unlock_exit;
4195                 } else {
4196                         lock_page(page);
4197                 }
4198                 locked_pages++;
4199                 if (!PageUptodate(page)) {
4200                         num_reads++;
4201                         all_uptodate = 0;
4202                 }
4203         }
4204         if (all_uptodate) {
4205                 if (start_i == 0)
4206                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4207                 goto unlock_exit;
4208         }
4209
4210         atomic_set(&eb->pages_reading, num_reads);
4211         for (i = start_i; i < num_pages; i++) {
4212                 page = extent_buffer_page(eb, i);
4213                 if (!PageUptodate(page)) {
4214                         ClearPageError(page);
4215                         err = __extent_read_full_page(tree, page,
4216                                                       get_extent, &bio,
4217                                                       mirror_num, &bio_flags);
4218                         if (err)
4219                                 ret = err;
4220                 } else {
4221                         unlock_page(page);
4222                 }
4223         }
4224
4225         if (bio)
4226                 submit_one_bio(READ, bio, mirror_num, bio_flags);
4227
4228         if (ret || wait != WAIT_COMPLETE)
4229                 return ret;
4230
4231         for (i = start_i; i < num_pages; i++) {
4232                 page = extent_buffer_page(eb, i);
4233                 wait_on_page_locked(page);
4234                 if (!PageUptodate(page))
4235                         ret = -EIO;
4236         }
4237
4238         if (!ret)
4239                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4240         return ret;
4241
4242 unlock_exit:
4243         i = start_i;
4244         while (locked_pages > 0) {
4245                 page = extent_buffer_page(eb, i);
4246                 i++;
4247                 unlock_page(page);
4248                 locked_pages--;
4249         }
4250         return ret;
4251 }
4252
4253 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4254                         unsigned long start,
4255                         unsigned long len)
4256 {
4257         size_t cur;
4258         size_t offset;
4259         struct page *page;
4260         char *kaddr;
4261         char *dst = (char *)dstv;
4262         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4263         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4264
4265         WARN_ON(start > eb->len);
4266         WARN_ON(start + len > eb->start + eb->len);
4267
4268         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4269
4270         while (len > 0) {
4271                 page = extent_buffer_page(eb, i);
4272
4273                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4274                 kaddr = page_address(page);
4275                 memcpy(dst, kaddr + offset, cur);
4276
4277                 dst += cur;
4278                 len -= cur;
4279                 offset = 0;
4280                 i++;
4281         }
4282 }
4283
4284 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4285                                unsigned long min_len, char **map,
4286                                unsigned long *map_start,
4287                                unsigned long *map_len)
4288 {
4289         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4290         char *kaddr;
4291         struct page *p;
4292         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4293         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4294         unsigned long end_i = (start_offset + start + min_len - 1) >>
4295                 PAGE_CACHE_SHIFT;
4296
4297         if (i != end_i)
4298                 return -EINVAL;
4299
4300         if (i == 0) {
4301                 offset = start_offset;
4302                 *map_start = 0;
4303         } else {
4304                 offset = 0;
4305                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4306         }
4307
4308         if (start + min_len > eb->len) {
4309                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4310                        "wanted %lu %lu\n", (unsigned long long)eb->start,
4311                        eb->len, start, min_len);
4312                 WARN_ON(1);
4313                 return -EINVAL;
4314         }
4315
4316         p = extent_buffer_page(eb, i);
4317         kaddr = page_address(p);
4318         *map = kaddr + offset;
4319         *map_len = PAGE_CACHE_SIZE - offset;
4320         return 0;
4321 }
4322
4323 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4324                           unsigned long start,
4325                           unsigned long len)
4326 {
4327         size_t cur;
4328         size_t offset;
4329         struct page *page;
4330         char *kaddr;
4331         char *ptr = (char *)ptrv;
4332         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4333         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4334         int ret = 0;
4335
4336         WARN_ON(start > eb->len);
4337         WARN_ON(start + len > eb->start + eb->len);
4338
4339         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4340
4341         while (len > 0) {
4342                 page = extent_buffer_page(eb, i);
4343
4344                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4345
4346                 kaddr = page_address(page);
4347                 ret = memcmp(ptr, kaddr + offset, cur);
4348                 if (ret)
4349                         break;
4350
4351                 ptr += cur;
4352                 len -= cur;
4353                 offset = 0;
4354                 i++;
4355         }
4356         return ret;
4357 }
4358
4359 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4360                          unsigned long start, unsigned long len)
4361 {
4362         size_t cur;
4363         size_t offset;
4364         struct page *page;
4365         char *kaddr;
4366         char *src = (char *)srcv;
4367         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4368         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4369
4370         WARN_ON(start > eb->len);
4371         WARN_ON(start + len > eb->start + eb->len);
4372
4373         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4374
4375         while (len > 0) {
4376                 page = extent_buffer_page(eb, i);
4377                 WARN_ON(!PageUptodate(page));
4378
4379                 cur = min(len, PAGE_CACHE_SIZE - offset);
4380                 kaddr = page_address(page);
4381                 memcpy(kaddr + offset, src, cur);
4382
4383                 src += cur;
4384                 len -= cur;
4385                 offset = 0;
4386                 i++;
4387         }
4388 }
4389
4390 void memset_extent_buffer(struct extent_buffer *eb, char c,
4391                           unsigned long start, unsigned long len)
4392 {
4393         size_t cur;
4394         size_t offset;
4395         struct page *page;
4396         char *kaddr;
4397         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4398         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4399
4400         WARN_ON(start > eb->len);
4401         WARN_ON(start + len > eb->start + eb->len);
4402
4403         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4404
4405         while (len > 0) {
4406                 page = extent_buffer_page(eb, i);
4407                 WARN_ON(!PageUptodate(page));
4408
4409                 cur = min(len, PAGE_CACHE_SIZE - offset);
4410                 kaddr = page_address(page);
4411                 memset(kaddr + offset, c, cur);
4412
4413                 len -= cur;
4414                 offset = 0;
4415                 i++;
4416         }
4417 }
4418
4419 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
4420                         unsigned long dst_offset, unsigned long src_offset,
4421                         unsigned long len)
4422 {
4423         u64 dst_len = dst->len;
4424         size_t cur;
4425         size_t offset;
4426         struct page *page;
4427         char *kaddr;
4428         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4429         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4430
4431         WARN_ON(src->len != dst_len);
4432
4433         offset = (start_offset + dst_offset) &
4434                 ((unsigned long)PAGE_CACHE_SIZE - 1);
4435
4436         while (len > 0) {
4437                 page = extent_buffer_page(dst, i);
4438                 WARN_ON(!PageUptodate(page));
4439
4440                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
4441
4442                 kaddr = page_address(page);
4443                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4444
4445                 src_offset += cur;
4446                 len -= cur;
4447                 offset = 0;
4448                 i++;
4449         }
4450 }
4451
4452 static void move_pages(struct page *dst_page, struct page *src_page,
4453                        unsigned long dst_off, unsigned long src_off,
4454                        unsigned long len)
4455 {
4456         char *dst_kaddr = page_address(dst_page);
4457         if (dst_page == src_page) {
4458                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
4459         } else {
4460                 char *src_kaddr = page_address(src_page);
4461                 char *p = dst_kaddr + dst_off + len;
4462                 char *s = src_kaddr + src_off + len;
4463
4464                 while (len--)
4465                         *--p = *--s;
4466         }
4467 }
4468
4469 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4470 {
4471         unsigned long distance = (src > dst) ? src - dst : dst - src;
4472         return distance < len;
4473 }
4474
4475 static void copy_pages(struct page *dst_page, struct page *src_page,
4476                        unsigned long dst_off, unsigned long src_off,
4477                        unsigned long len)
4478 {
4479         char *dst_kaddr = page_address(dst_page);
4480         char *src_kaddr;
4481         int must_memmove = 0;
4482
4483         if (dst_page != src_page) {
4484                 src_kaddr = page_address(src_page);
4485         } else {
4486                 src_kaddr = dst_kaddr;
4487                 if (areas_overlap(src_off, dst_off, len))
4488                         must_memmove = 1;
4489         }
4490
4491         if (must_memmove)
4492                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
4493         else
4494                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
4495 }
4496
4497 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4498                            unsigned long src_offset, unsigned long len)
4499 {
4500         size_t cur;
4501         size_t dst_off_in_page;
4502         size_t src_off_in_page;
4503         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4504         unsigned long dst_i;
4505         unsigned long src_i;
4506
4507         if (src_offset + len > dst->len) {
4508                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4509                        "len %lu dst len %lu\n", src_offset, len, dst->len);
4510                 BUG_ON(1);
4511         }
4512         if (dst_offset + len > dst->len) {
4513                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4514                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
4515                 BUG_ON(1);
4516         }
4517
4518         while (len > 0) {
4519                 dst_off_in_page = (start_offset + dst_offset) &
4520                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4521                 src_off_in_page = (start_offset + src_offset) &
4522                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4523
4524                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4525                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
4526
4527                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
4528                                                src_off_in_page));
4529                 cur = min_t(unsigned long, cur,
4530                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
4531
4532                 copy_pages(extent_buffer_page(dst, dst_i),
4533                            extent_buffer_page(dst, src_i),
4534                            dst_off_in_page, src_off_in_page, cur);
4535
4536                 src_offset += cur;
4537                 dst_offset += cur;
4538                 len -= cur;
4539         }
4540 }
4541
4542 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4543                            unsigned long src_offset, unsigned long len)
4544 {
4545         size_t cur;
4546         size_t dst_off_in_page;
4547         size_t src_off_in_page;
4548         unsigned long dst_end = dst_offset + len - 1;
4549         unsigned long src_end = src_offset + len - 1;
4550         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4551         unsigned long dst_i;
4552         unsigned long src_i;
4553
4554         if (src_offset + len > dst->len) {
4555                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4556                        "len %lu len %lu\n", src_offset, len, dst->len);
4557                 BUG_ON(1);
4558         }
4559         if (dst_offset + len > dst->len) {
4560                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4561                        "len %lu len %lu\n", dst_offset, len, dst->len);
4562                 BUG_ON(1);
4563         }
4564         if (dst_offset < src_offset) {
4565                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4566                 return;
4567         }
4568         while (len > 0) {
4569                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4570                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4571
4572                 dst_off_in_page = (start_offset + dst_end) &
4573                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4574                 src_off_in_page = (start_offset + src_end) &
4575                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4576
4577                 cur = min_t(unsigned long, len, src_off_in_page + 1);
4578                 cur = min(cur, dst_off_in_page + 1);
4579                 move_pages(extent_buffer_page(dst, dst_i),
4580                            extent_buffer_page(dst, src_i),
4581                            dst_off_in_page - cur + 1,
4582                            src_off_in_page - cur + 1, cur);
4583
4584                 dst_end -= cur;
4585                 src_end -= cur;
4586                 len -= cur;
4587         }
4588 }
4589
4590 int try_release_extent_buffer(struct page *page, gfp_t mask)
4591 {
4592         struct extent_buffer *eb;
4593
4594         /*
4595          * We need to make sure noboody is attaching this page to an eb right
4596          * now.
4597          */
4598         spin_lock(&page->mapping->private_lock);
4599         if (!PagePrivate(page)) {
4600                 spin_unlock(&page->mapping->private_lock);
4601                 return 1;
4602         }
4603
4604         eb = (struct extent_buffer *)page->private;
4605         BUG_ON(!eb);
4606
4607         /*
4608          * This is a little awful but should be ok, we need to make sure that
4609          * the eb doesn't disappear out from under us while we're looking at
4610          * this page.
4611          */
4612         spin_lock(&eb->refs_lock);
4613         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb, page)) {
4614                 spin_unlock(&eb->refs_lock);
4615                 spin_unlock(&page->mapping->private_lock);
4616                 return 0;
4617         }
4618         spin_unlock(&page->mapping->private_lock);
4619
4620         if ((mask & GFP_NOFS) == GFP_NOFS)
4621                 mask = GFP_NOFS;
4622
4623         /*
4624          * If tree ref isn't set then we know the ref on this eb is a real ref,
4625          * so just return, this page will likely be freed soon anyway.
4626          */
4627         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4628                 spin_unlock(&eb->refs_lock);
4629                 return 0;
4630         }
4631         release_extent_buffer(eb, mask);
4632
4633         return 1;
4634 }