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/gfp.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/page-flags.h>
   8 #include <linux/module.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/swap.h>
  12 #include <linux/writeback.h>
  13 #include <linux/pagevec.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "compat.h"
  17 #include "ctree.h"
  18 #include "btrfs_inode.h"
  19
  20 static struct kmem_cache *extent_state_cache;
  21 static struct kmem_cache *extent_buffer_cache;
  22
  23 static LIST_HEAD(buffers);
  24 static LIST_HEAD(states);
  25
  26 #define LEAK_DEBUG 0
  27 #if LEAK_DEBUG
  28 static DEFINE_SPINLOCK(leak_lock);
  29 #endif
  30
  31 #define BUFFER_LRU_MAX 64
  32
  33 struct tree_entry {
  34         u64 start;
  35         u64 end;
  36         struct rb_node rb_node;
  37 };
  38
  39 struct extent_page_data {
  40         struct bio *bio;
  41         struct extent_io_tree *tree;
  42         get_extent_t *get_extent;
  43
  44         /* tells writepage not to lock the state bits for this range
  45          * it still does the unlocking
  46          */
  47         unsigned int extent_locked:1;
  48
  49         /* tells the submit_bio code to use a WRITE_SYNC */
  50         unsigned int sync_io:1;
  51 };
  52
  53 int __init extent_io_init(void)
  54 {
  55         extent_state_cache = kmem_cache_create("extent_state",
  56                         sizeof(struct extent_state), 0,
  57                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  58         if (!extent_state_cache)
  59                 return -ENOMEM;
  60
  61         extent_buffer_cache = kmem_cache_create("extent_buffers",
  62                         sizeof(struct extent_buffer), 0,
  63                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  64         if (!extent_buffer_cache)
  65                 goto free_state_cache;
  66         return 0;
  67
  68 free_state_cache:
  69         kmem_cache_destroy(extent_state_cache);
  70         return -ENOMEM;
  71 }
  72
  73 void extent_io_exit(void)
  74 {
  75         struct extent_state *state;
  76         struct extent_buffer *eb;
  77
  78         while (!list_empty(&states)) {
  79                 state = list_entry(states.next, struct extent_state, leak_list);
  80                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  81                        "state %lu in tree %p refs %d\n",
  82                        (unsigned long long)state->start,
  83                        (unsigned long long)state->end,
  84                        state->state, state->tree, atomic_read(&state->refs));
  85                 list_del(&state->leak_list);
  86                 kmem_cache_free(extent_state_cache, state);
  87
  88         }
  89
  90         while (!list_empty(&buffers)) {
  91                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  92                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  93                        "refs %d\n", (unsigned long long)eb->start,
  94                        eb->len, atomic_read(&eb->refs));
  95                 list_del(&eb->leak_list);
  96                 kmem_cache_free(extent_buffer_cache, eb);
  97         }
  98         if (extent_state_cache)
  99                 kmem_cache_destroy(extent_state_cache);
 100         if (extent_buffer_cache)
 101                 kmem_cache_destroy(extent_buffer_cache);
 102 }
 103
 104 void extent_io_tree_init(struct extent_io_tree *tree,
 105                           struct address_space *mapping, gfp_t mask)
 106 {
 107         tree->state.rb_node = NULL;
 108         tree->buffer.rb_node = NULL;
 109         tree->ops = NULL;
 110         tree->dirty_bytes = 0;
 111         spin_lock_init(&tree->lock);
 112         spin_lock_init(&tree->buffer_lock);
 113         tree->mapping = mapping;
 114 }
 115
 116 static struct extent_state *alloc_extent_state(gfp_t mask)
 117 {
 118         struct extent_state *state;
 119 #if LEAK_DEBUG
 120         unsigned long flags;
 121 #endif
 122
 123         state = kmem_cache_alloc(extent_state_cache, mask);
 124         if (!state)
 125                 return state;
 126         state->state = 0;
 127         state->private = 0;
 128         state->tree = NULL;
 129 #if LEAK_DEBUG
 130         spin_lock_irqsave(&leak_lock, flags);
 131         list_add(&state->leak_list, &states);
 132         spin_unlock_irqrestore(&leak_lock, flags);
 133 #endif
 134         atomic_set(&state->refs, 1);
 135         init_waitqueue_head(&state->wq);
 136         return state;
 137 }
 138
 139 static void free_extent_state(struct extent_state *state)
 140 {
 141         if (!state)
 142                 return;
 143         if (atomic_dec_and_test(&state->refs)) {
 144 #if LEAK_DEBUG
 145                 unsigned long flags;
 146 #endif
 147                 WARN_ON(state->tree);
 148 #if LEAK_DEBUG
 149                 spin_lock_irqsave(&leak_lock, flags);
 150                 list_del(&state->leak_list);
 151                 spin_unlock_irqrestore(&leak_lock, flags);
 152 #endif
 153                 kmem_cache_free(extent_state_cache, state);
 154         }
 155 }
 156
 157 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 158                                    struct rb_node *node)
 159 {
 160         struct rb_node **p = &root->rb_node;
 161         struct rb_node *parent = NULL;
 162         struct tree_entry *entry;
 163
 164         while (*p) {
 165                 parent = *p;
 166                 entry = rb_entry(parent, struct tree_entry, rb_node);
 167
 168                 if (offset < entry->start)
 169                         p = &(*p)->rb_left;
 170                 else if (offset > entry->end)
 171                         p = &(*p)->rb_right;
 172                 else
 173                         return parent;
 174         }
 175
 176         entry = rb_entry(node, struct tree_entry, rb_node);
 177         rb_link_node(node, parent, p);
 178         rb_insert_color(node, root);
 179         return NULL;
 180 }
 181
 182 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 183                                      struct rb_node **prev_ret,
 184                                      struct rb_node **next_ret)
 185 {
 186         struct rb_root *root = &tree->state;
 187         struct rb_node *n = root->rb_node;
 188         struct rb_node *prev = NULL;
 189         struct rb_node *orig_prev = NULL;
 190         struct tree_entry *entry;
 191         struct tree_entry *prev_entry = NULL;
 192
 193         while (n) {
 194                 entry = rb_entry(n, struct tree_entry, rb_node);
 195                 prev = n;
 196                 prev_entry = entry;
 197
 198                 if (offset < entry->start)
 199                         n = n->rb_left;
 200                 else if (offset > entry->end)
 201                         n = n->rb_right;
 202                 else
 203                         return n;
 204         }
 205
 206         if (prev_ret) {
 207                 orig_prev = prev;
 208                 while (prev && offset > prev_entry->end) {
 209                         prev = rb_next(prev);
 210                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 211                 }
 212                 *prev_ret = prev;
 213                 prev = orig_prev;
 214         }
 215
 216         if (next_ret) {
 217                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 218                 while (prev && offset < prev_entry->start) {
 219                         prev = rb_prev(prev);
 220                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 221                 }
 222                 *next_ret = prev;
 223         }
 224         return NULL;
 225 }
 226
 227 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 228                                           u64 offset)
 229 {
 230         struct rb_node *prev = NULL;
 231         struct rb_node *ret;
 232
 233         ret = __etree_search(tree, offset, &prev, NULL);
 234         if (!ret)
 235                 return prev;
 236         return ret;
 237 }
 238
 239 static struct extent_buffer *buffer_tree_insert(struct extent_io_tree *tree,
 240                                           u64 offset, struct rb_node *node)
 241 {
 242         struct rb_root *root = &tree->buffer;
 243         struct rb_node **p = &root->rb_node;
 244         struct rb_node *parent = NULL;
 245         struct extent_buffer *eb;
 246
 247         while (*p) {
 248                 parent = *p;
 249                 eb = rb_entry(parent, struct extent_buffer, rb_node);
 250
 251                 if (offset < eb->start)
 252                         p = &(*p)->rb_left;
 253                 else if (offset > eb->start)
 254                         p = &(*p)->rb_right;
 255                 else
 256                         return eb;
 257         }
 258
 259         rb_link_node(node, parent, p);
 260         rb_insert_color(node, root);
 261         return NULL;
 262 }
 263
 264 static struct extent_buffer *buffer_search(struct extent_io_tree *tree,
 265                                            u64 offset)
 266 {
 267         struct rb_root *root = &tree->buffer;
 268         struct rb_node *n = root->rb_node;
 269         struct extent_buffer *eb;
 270
 271         while (n) {
 272                 eb = rb_entry(n, struct extent_buffer, rb_node);
 273                 if (offset < eb->start)
 274                         n = n->rb_left;
 275                 else if (offset > eb->start)
 276                         n = n->rb_right;
 277                 else
 278                         return eb;
 279         }
 280         return NULL;
 281 }
 282
 283 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 284                      struct extent_state *other)
 285 {
 286         if (tree->ops && tree->ops->merge_extent_hook)
 287                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 288                                              other);
 289 }
 290
 291 /*
 292  * utility function to look for merge candidates inside a given range.
 293  * Any extents with matching state are merged together into a single
 294  * extent in the tree.  Extents with EXTENT_IO in their state field
 295  * are not merged because the end_io handlers need to be able to do
 296  * operations on them without sleeping (or doing allocations/splits).
 297  *
 298  * This should be called with the tree lock held.
 299  */
 300 static int merge_state(struct extent_io_tree *tree,
 301                        struct extent_state *state)
 302 {
 303         struct extent_state *other;
 304         struct rb_node *other_node;
 305
 306         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 307                 return 0;
 308
 309         other_node = rb_prev(&state->rb_node);
 310         if (other_node) {
 311                 other = rb_entry(other_node, struct extent_state, rb_node);
 312                 if (other->end == state->start - 1 &&
 313                     other->state == state->state) {
 314                         merge_cb(tree, state, other);
 315                         state->start = other->start;
 316                         other->tree = NULL;
 317                         rb_erase(&other->rb_node, &tree->state);
 318                         free_extent_state(other);
 319                 }
 320         }
 321         other_node = rb_next(&state->rb_node);
 322         if (other_node) {
 323                 other = rb_entry(other_node, struct extent_state, rb_node);
 324                 if (other->start == state->end + 1 &&
 325                     other->state == state->state) {
 326                         merge_cb(tree, state, other);
 327                         other->start = state->start;
 328                         state->tree = NULL;
 329                         rb_erase(&state->rb_node, &tree->state);
 330                         free_extent_state(state);
 331                         state = NULL;
 332                 }
 333         }
 334
 335         return 0;
 336 }
 337
 338 static int set_state_cb(struct extent_io_tree *tree,
 339                          struct extent_state *state,
 340                          unsigned long bits)
 341 {
 342         if (tree->ops && tree->ops->set_bit_hook) {
 343                 return tree->ops->set_bit_hook(tree->mapping->host,
 344                                                state->start, state->end,
 345                                                state->state, bits);
 346         }
 347
 348         return 0;
 349 }
 350
 351 static void clear_state_cb(struct extent_io_tree *tree,
 352                            struct extent_state *state,
 353                            unsigned long bits)
 354 {
 355         if (tree->ops && tree->ops->clear_bit_hook)
 356                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 357 }
 358
 359 /*
 360  * insert an extent_state struct into the tree.  'bits' are set on the
 361  * struct before it is inserted.
 362  *
 363  * This may return -EEXIST if the extent is already there, in which case the
 364  * state struct is freed.
 365  *
 366  * The tree lock is not taken internally.  This is a utility function and
 367  * probably isn't what you want to call (see set/clear_extent_bit).
 368  */
 369 static int insert_state(struct extent_io_tree *tree,
 370                         struct extent_state *state, u64 start, u64 end,
 371                         int bits)
 372 {
 373         struct rb_node *node;
 374         int ret;
 375
 376         if (end < start) {
 377                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
 378                        (unsigned long long)end,
 379                        (unsigned long long)start);
 380                 WARN_ON(1);
 381         }
 382         state->start = start;
 383         state->end = end;
 384         ret = set_state_cb(tree, state, bits);
 385         if (ret)
 386                 return ret;
 387
 388         if (bits & EXTENT_DIRTY)
 389                 tree->dirty_bytes += end - start + 1;
 390         state->state |= bits;
 391         node = tree_insert(&tree->state, end, &state->rb_node);
 392         if (node) {
 393                 struct extent_state *found;
 394                 found = rb_entry(node, struct extent_state, rb_node);
 395                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 396                        "%llu %llu\n", (unsigned long long)found->start,
 397                        (unsigned long long)found->end,
 398                        (unsigned long long)start, (unsigned long long)end);
 399                 free_extent_state(state);
 400                 return -EEXIST;
 401         }
 402         state->tree = tree;
 403         merge_state(tree, state);
 404         return 0;
 405 }
 406
 407 static int split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 408                      u64 split)
 409 {
 410         if (tree->ops && tree->ops->split_extent_hook)
 411                 return tree->ops->split_extent_hook(tree->mapping->host,
 412                                                     orig, split);
 413         return 0;
 414 }
 415
 416 /*
 417  * split a given extent state struct in two, inserting the preallocated
 418  * struct 'prealloc' as the newly created second half.  'split' indicates an
 419  * offset inside 'orig' where it should be split.
 420  *
 421  * Before calling,
 422  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 423  * are two extent state structs in the tree:
 424  * prealloc: [orig->start, split - 1]
 425  * orig: [ split, orig->end ]
 426  *
 427  * The tree locks are not taken by this function. They need to be held
 428  * by the caller.
 429  */
 430 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 431                        struct extent_state *prealloc, u64 split)
 432 {
 433         struct rb_node *node;
 434
 435         split_cb(tree, orig, split);
 436
 437         prealloc->start = orig->start;
 438         prealloc->end = split - 1;
 439         prealloc->state = orig->state;
 440         orig->start = split;
 441
 442         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 443         if (node) {
 444                 free_extent_state(prealloc);
 445                 return -EEXIST;
 446         }
 447         prealloc->tree = tree;
 448         return 0;
 449 }
 450
 451 /*
 452  * utility function to clear some bits in an extent state struct.
 453  * it will optionally wake up any one waiting on this state (wake == 1), or
 454  * forcibly remove the state from the tree (delete == 1).
 455  *
 456  * If no bits are set on the state struct after clearing things, the
 457  * struct is freed and removed from the tree
 458  */
 459 static int clear_state_bit(struct extent_io_tree *tree,
 460                             struct extent_state *state, int bits, int wake,
 461                             int delete)
 462 {
 463         int bits_to_clear = bits & ~EXTENT_DO_ACCOUNTING;
 464         int ret = state->state & bits_to_clear;
 465
 466         if ((bits & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 467                 u64 range = state->end - state->start + 1;
 468                 WARN_ON(range > tree->dirty_bytes);
 469                 tree->dirty_bytes -= range;
 470         }
 471         clear_state_cb(tree, state, bits);
 472         state->state &= ~bits_to_clear;
 473         if (wake)
 474                 wake_up(&state->wq);
 475         if (delete || state->state == 0) {
 476                 if (state->tree) {
 477                         clear_state_cb(tree, state, state->state);
 478                         rb_erase(&state->rb_node, &tree->state);
 479                         state->tree = NULL;
 480                         free_extent_state(state);
 481                 } else {
 482                         WARN_ON(1);
 483                 }
 484         } else {
 485                 merge_state(tree, state);
 486         }
 487         return ret;
 488 }
 489
 490 /*
 491  * clear some bits on a range in the tree.  This may require splitting
 492  * or inserting elements in the tree, so the gfp mask is used to
 493  * indicate which allocations or sleeping are allowed.
 494  *
 495  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 496  * the given range from the tree regardless of state (ie for truncate).
 497  *
 498  * the range [start, end] is inclusive.
 499  *
 500  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 501  * bits were already set, or zero if none of the bits were already set.
 502  */
 503 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 504                      int bits, int wake, int delete,
 505                      struct extent_state **cached_state,
 506                      gfp_t mask)
 507 {
 508         struct extent_state *state;
 509         struct extent_state *cached;
 510         struct extent_state *prealloc = NULL;
 511         struct rb_node *next_node;
 512         struct rb_node *node;
 513         u64 last_end;
 514         int err;
 515         int set = 0;
 516
 517 again:
 518         if (!prealloc && (mask & __GFP_WAIT)) {
 519                 prealloc = alloc_extent_state(mask);
 520                 if (!prealloc)
 521                         return -ENOMEM;
 522         }
 523
 524         spin_lock(&tree->lock);
 525         if (cached_state) {
 526                 cached = *cached_state;
 527                 *cached_state = NULL;
 528                 cached_state = NULL;
 529                 if (cached && cached->tree && cached->start == start) {
 530                         atomic_dec(&cached->refs);
 531                         state = cached;
 532                         goto hit_next;
 533                 }
 534                 free_extent_state(cached);
 535         }
 536         /*
 537          * this search will find the extents that end after
 538          * our range starts
 539          */
 540         node = tree_search(tree, start);
 541         if (!node)
 542                 goto out;
 543         state = rb_entry(node, struct extent_state, rb_node);
 544 hit_next:
 545         if (state->start > end)
 546                 goto out;
 547         WARN_ON(state->end < start);
 548         last_end = state->end;
 549
 550         /*
 551          *     | ---- desired range ---- |
 552          *  | state | or
 553          *  | ------------- state -------------- |
 554          *
 555          * We need to split the extent we found, and may flip
 556          * bits on second half.
 557          *
 558          * If the extent we found extends past our range, we
 559          * just split and search again.  It'll get split again
 560          * the next time though.
 561          *
 562          * If the extent we found is inside our range, we clear
 563          * the desired bit on it.
 564          */
 565
 566         if (state->start < start) {
 567                 if (!prealloc)
 568                         prealloc = alloc_extent_state(GFP_ATOMIC);
 569                 err = split_state(tree, state, prealloc, start);
 570                 BUG_ON(err == -EEXIST);
 571                 prealloc = NULL;
 572                 if (err)
 573                         goto out;
 574                 if (state->end <= end) {
 575                         set |= clear_state_bit(tree, state, bits, wake,
 576                                                delete);
 577                         if (last_end == (u64)-1)
 578                                 goto out;
 579                         start = last_end + 1;
 580                 }
 581                 goto search_again;
 582         }
 583         /*
 584          * | ---- desired range ---- |
 585          *                        | state |
 586          * We need to split the extent, and clear the bit
 587          * on the first half
 588          */
 589         if (state->start <= end && state->end > end) {
 590                 if (!prealloc)
 591                         prealloc = alloc_extent_state(GFP_ATOMIC);
 592                 err = split_state(tree, state, prealloc, end + 1);
 593                 BUG_ON(err == -EEXIST);
 594                 if (wake)
 595                         wake_up(&state->wq);
 596
 597                 set |= clear_state_bit(tree, prealloc, bits, wake, delete);
 598
 599                 prealloc = NULL;
 600                 goto out;
 601         }
 602
 603         if (state->end < end && prealloc && !need_resched())
 604                 next_node = rb_next(&state->rb_node);
 605         else
 606                 next_node = NULL;
 607
 608         set |= clear_state_bit(tree, state, bits, wake, delete);
 609         if (last_end == (u64)-1)
 610                 goto out;
 611         start = last_end + 1;
 612         if (start <= end && next_node) {
 613                 state = rb_entry(next_node, struct extent_state,
 614                                  rb_node);
 615                 if (state->start == start)
 616                         goto hit_next;
 617         }
 618         goto search_again;
 619
 620 out:
 621         spin_unlock(&tree->lock);
 622         if (prealloc)
 623                 free_extent_state(prealloc);
 624
 625         return set;
 626
 627 search_again:
 628         if (start > end)
 629                 goto out;
 630         spin_unlock(&tree->lock);
 631         if (mask & __GFP_WAIT)
 632                 cond_resched();
 633         goto again;
 634 }
 635
 636 static int wait_on_state(struct extent_io_tree *tree,
 637                          struct extent_state *state)
 638                 __releases(tree->lock)
 639                 __acquires(tree->lock)
 640 {
 641         DEFINE_WAIT(wait);
 642         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 643         spin_unlock(&tree->lock);
 644         schedule();
 645         spin_lock(&tree->lock);
 646         finish_wait(&state->wq, &wait);
 647         return 0;
 648 }
 649
 650 /*
 651  * waits for one or more bits to clear on a range in the state tree.
 652  * The range [start, end] is inclusive.
 653  * The tree lock is taken by this function
 654  */
 655 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 656 {
 657         struct extent_state *state;
 658         struct rb_node *node;
 659
 660         spin_lock(&tree->lock);
 661 again:
 662         while (1) {
 663                 /*
 664                  * this search will find all the extents that end after
 665                  * our range starts
 666                  */
 667                 node = tree_search(tree, start);
 668                 if (!node)
 669                         break;
 670
 671                 state = rb_entry(node, struct extent_state, rb_node);
 672
 673                 if (state->start > end)
 674                         goto out;
 675
 676                 if (state->state & bits) {
 677                         start = state->start;
 678                         atomic_inc(&state->refs);
 679                         wait_on_state(tree, state);
 680                         free_extent_state(state);
 681                         goto again;
 682                 }
 683                 start = state->end + 1;
 684
 685                 if (start > end)
 686                         break;
 687
 688                 if (need_resched()) {
 689                         spin_unlock(&tree->lock);
 690                         cond_resched();
 691                         spin_lock(&tree->lock);
 692                 }
 693         }
 694 out:
 695         spin_unlock(&tree->lock);
 696         return 0;
 697 }
 698
 699 static int set_state_bits(struct extent_io_tree *tree,
 700                            struct extent_state *state,
 701                            int bits)
 702 {
 703         int ret;
 704
 705         ret = set_state_cb(tree, state, bits);
 706         if (ret)
 707                 return ret;
 708
 709         if ((bits & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 710                 u64 range = state->end - state->start + 1;
 711                 tree->dirty_bytes += range;
 712         }
 713         state->state |= bits;
 714
 715         return 0;
 716 }
 717
 718 static void cache_state(struct extent_state *state,
 719                         struct extent_state **cached_ptr)
 720 {
 721         if (cached_ptr && !(*cached_ptr)) {
 722                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 723                         *cached_ptr = state;
 724                         atomic_inc(&state->refs);
 725                 }
 726         }
 727 }
 728
 729 /*
 730  * set some bits on a range in the tree.  This may require allocations or
 731  * sleeping, so the gfp mask is used to indicate what is allowed.
 732  *
 733  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 734  * part of the range already has the desired bits set.  The start of the
 735  * existing range is returned in failed_start in this case.
 736  *
 737  * [start, end] is inclusive This takes the tree lock.
 738  */
 739
 740 static int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 741                           int bits, int exclusive_bits, u64 *failed_start,
 742                           struct extent_state **cached_state,
 743                           gfp_t mask)
 744 {
 745         struct extent_state *state;
 746         struct extent_state *prealloc = NULL;
 747         struct rb_node *node;
 748         int err = 0;
 749         u64 last_start;
 750         u64 last_end;
 751
 752 again:
 753         if (!prealloc && (mask & __GFP_WAIT)) {
 754                 prealloc = alloc_extent_state(mask);
 755                 if (!prealloc)
 756                         return -ENOMEM;
 757         }
 758
 759         spin_lock(&tree->lock);
 760         if (cached_state && *cached_state) {
 761                 state = *cached_state;
 762                 if (state->start == start && state->tree) {
 763                         node = &state->rb_node;
 764                         goto hit_next;
 765                 }
 766         }
 767         /*
 768          * this search will find all the extents that end after
 769          * our range starts.
 770          */
 771         node = tree_search(tree, start);
 772         if (!node) {
 773                 err = insert_state(tree, prealloc, start, end, bits);
 774                 prealloc = NULL;
 775                 BUG_ON(err == -EEXIST);
 776                 goto out;
 777         }
 778         state = rb_entry(node, struct extent_state, rb_node);
 779 hit_next:
 780         last_start = state->start;
 781         last_end = state->end;
 782
 783         /*
 784          * | ---- desired range ---- |
 785          * | state |
 786          *
 787          * Just lock what we found and keep going
 788          */
 789         if (state->start == start && state->end <= end) {
 790                 struct rb_node *next_node;
 791                 if (state->state & exclusive_bits) {
 792                         *failed_start = state->start;
 793                         err = -EEXIST;
 794                         goto out;
 795                 }
 796
 797                 err = set_state_bits(tree, state, bits);
 798                 if (err)
 799                         goto out;
 800
 801                 cache_state(state, cached_state);
 802                 merge_state(tree, state);
 803                 if (last_end == (u64)-1)
 804                         goto out;
 805
 806                 start = last_end + 1;
 807                 if (start < end && prealloc && !need_resched()) {
 808                         next_node = rb_next(node);
 809                         if (next_node) {
 810                                 state = rb_entry(next_node, struct extent_state,
 811                                                  rb_node);
 812                                 if (state->start == start)
 813                                         goto hit_next;
 814                         }
 815                 }
 816                 goto search_again;
 817         }
 818
 819         /*
 820          *     | ---- desired range ---- |
 821          * | state |
 822          *   or
 823          * | ------------- state -------------- |
 824          *
 825          * We need to split the extent we found, and may flip bits on
 826          * second half.
 827          *
 828          * If the extent we found extends past our
 829          * range, we just split and search again.  It'll get split
 830          * again the next time though.
 831          *
 832          * If the extent we found is inside our range, we set the
 833          * desired bit on it.
 834          */
 835         if (state->start < start) {
 836                 if (state->state & exclusive_bits) {
 837                         *failed_start = start;
 838                         err = -EEXIST;
 839                         goto out;
 840                 }
 841                 err = split_state(tree, state, prealloc, start);
 842                 BUG_ON(err == -EEXIST);
 843                 prealloc = NULL;
 844                 if (err)
 845                         goto out;
 846                 if (state->end <= end) {
 847                         err = set_state_bits(tree, state, bits);
 848                         if (err)
 849                                 goto out;
 850                         cache_state(state, cached_state);
 851                         merge_state(tree, state);
 852                         if (last_end == (u64)-1)
 853                                 goto out;
 854                         start = last_end + 1;
 855                 }
 856                 goto search_again;
 857         }
 858         /*
 859          * | ---- desired range ---- |
 860          *     | state | or               | state |
 861          *
 862          * There's a hole, we need to insert something in it and
 863          * ignore the extent we found.
 864          */
 865         if (state->start > start) {
 866                 u64 this_end;
 867                 if (end < last_start)
 868                         this_end = end;
 869                 else
 870                         this_end = last_start - 1;
 871                 err = insert_state(tree, prealloc, start, this_end,
 872                                    bits);
 873                 BUG_ON(err == -EEXIST);
 874                 if (err) {
 875                         prealloc = NULL;
 876                         goto out;
 877                 }
 878                 cache_state(prealloc, cached_state);
 879                 prealloc = NULL;
 880                 start = this_end + 1;
 881                 goto search_again;
 882         }
 883         /*
 884          * | ---- desired range ---- |
 885          *                        | state |
 886          * We need to split the extent, and set the bit
 887          * on the first half
 888          */
 889         if (state->start <= end && state->end > end) {
 890                 if (state->state & exclusive_bits) {
 891                         *failed_start = start;
 892                         err = -EEXIST;
 893                         goto out;
 894                 }
 895                 err = split_state(tree, state, prealloc, end + 1);
 896                 BUG_ON(err == -EEXIST);
 897
 898                 err = set_state_bits(tree, prealloc, bits);
 899                 if (err) {
 900                         prealloc = NULL;
 901                         goto out;
 902                 }
 903                 cache_state(prealloc, cached_state);
 904                 merge_state(tree, prealloc);
 905                 prealloc = NULL;
 906                 goto out;
 907         }
 908
 909         goto search_again;
 910
 911 out:
 912         spin_unlock(&tree->lock);
 913         if (prealloc)
 914                 free_extent_state(prealloc);
 915
 916         return err;
 917
 918 search_again:
 919         if (start > end)
 920                 goto out;
 921         spin_unlock(&tree->lock);
 922         if (mask & __GFP_WAIT)
 923                 cond_resched();
 924         goto again;
 925 }
 926
 927 /* wrappers around set/clear extent bit */
 928 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 929                      gfp_t mask)
 930 {
 931         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 932                               NULL, mask);
 933 }
 934
 935 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 936                     int bits, gfp_t mask)
 937 {
 938         return set_extent_bit(tree, start, end, bits, 0, NULL,
 939                               NULL, mask);
 940 }
 941
 942 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 943                       int bits, gfp_t mask)
 944 {
 945         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
 946 }
 947
 948 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
 949                      gfp_t mask)
 950 {
 951         return set_extent_bit(tree, start, end,
 952                               EXTENT_DELALLOC | EXTENT_DIRTY | EXTENT_UPTODATE,
 953                               0, NULL, NULL, mask);
 954 }
 955
 956 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 957                        gfp_t mask)
 958 {
 959         return clear_extent_bit(tree, start, end,
 960                                 EXTENT_DIRTY | EXTENT_DELALLOC |
 961                                 EXTENT_DO_ACCOUNTING, 0, 0,
 962                                 NULL, mask);
 963 }
 964
 965 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 966                      gfp_t mask)
 967 {
 968         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 969                               NULL, mask);
 970 }
 971
 972 static int clear_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 973                        gfp_t mask)
 974 {
 975         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0,
 976                                 NULL, mask);
 977 }
 978
 979 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 980                         gfp_t mask)
 981 {
 982         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, NULL,
 983                               NULL, mask);
 984 }
 985
 986 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
 987                                  u64 end, gfp_t mask)
 988 {
 989         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
 990                                 NULL, mask);
 991 }
 992
 993 int wait_on_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 994 {
 995         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 996 }
 997
 998 /*
 999  * either insert or lock state struct between start and end use mask to tell
1000  * us if waiting is desired.
1001  */
1002 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1003                      int bits, struct extent_state **cached_state, gfp_t mask)
1004 {
1005         int err;
1006         u64 failed_start;
1007         while (1) {
1008                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1009                                      EXTENT_LOCKED, &failed_start,
1010                                      cached_state, mask);
1011                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
1012                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1013                         start = failed_start;
1014                 } else {
1015                         break;
1016                 }
1017                 WARN_ON(start > end);
1018         }
1019         return err;
1020 }
1021
1022 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1023 {
1024         return lock_extent_bits(tree, start, end, 0, NULL, mask);
1025 }
1026
1027 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1028                     gfp_t mask)
1029 {
1030         int err;
1031         u64 failed_start;
1032
1033         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1034                              &failed_start, NULL, mask);
1035         if (err == -EEXIST) {
1036                 if (failed_start > start)
1037                         clear_extent_bit(tree, start, failed_start - 1,
1038                                          EXTENT_LOCKED, 1, 0, NULL, mask);
1039                 return 0;
1040         }
1041         return 1;
1042 }
1043
1044 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1045                          struct extent_state **cached, gfp_t mask)
1046 {
1047         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1048                                 mask);
1049 }
1050
1051 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1052                   gfp_t mask)
1053 {
1054         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1055                                 mask);
1056 }
1057
1058 /*
1059  * helper function to set pages and extents in the tree dirty
1060  */
1061 int set_range_dirty(struct extent_io_tree *tree, u64 start, u64 end)
1062 {
1063         unsigned long index = start >> PAGE_CACHE_SHIFT;
1064         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1065         struct page *page;
1066
1067         while (index <= end_index) {
1068                 page = find_get_page(tree->mapping, index);
1069                 BUG_ON(!page);
1070                 __set_page_dirty_nobuffers(page);
1071                 page_cache_release(page);
1072                 index++;
1073         }
1074         return 0;
1075 }
1076
1077 /*
1078  * helper function to set both pages and extents in the tree writeback
1079  */
1080 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1081 {
1082         unsigned long index = start >> PAGE_CACHE_SHIFT;
1083         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1084         struct page *page;
1085
1086         while (index <= end_index) {
1087                 page = find_get_page(tree->mapping, index);
1088                 BUG_ON(!page);
1089                 set_page_writeback(page);
1090                 page_cache_release(page);
1091                 index++;
1092         }
1093         return 0;
1094 }
1095
1096 /*
1097  * find the first offset in the io tree with 'bits' set. zero is
1098  * returned if we find something, and *start_ret and *end_ret are
1099  * set to reflect the state struct that was found.
1100  *
1101  * If nothing was found, 1 is returned, < 0 on error
1102  */
1103 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1104                           u64 *start_ret, u64 *end_ret, int bits)
1105 {
1106         struct rb_node *node;
1107         struct extent_state *state;
1108         int ret = 1;
1109
1110         spin_lock(&tree->lock);
1111         /*
1112          * this search will find all the extents that end after
1113          * our range starts.
1114          */
1115         node = tree_search(tree, start);
1116         if (!node)
1117                 goto out;
1118
1119         while (1) {
1120                 state = rb_entry(node, struct extent_state, rb_node);
1121                 if (state->end >= start && (state->state & bits)) {
1122                         *start_ret = state->start;
1123                         *end_ret = state->end;
1124                         ret = 0;
1125                         break;
1126                 }
1127                 node = rb_next(node);
1128                 if (!node)
1129                         break;
1130         }
1131 out:
1132         spin_unlock(&tree->lock);
1133         return ret;
1134 }
1135
1136 /* find the first state struct with 'bits' set after 'start', and
1137  * return it.  tree->lock must be held.  NULL will returned if
1138  * nothing was found after 'start'
1139  */
1140 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1141                                                  u64 start, int bits)
1142 {
1143         struct rb_node *node;
1144         struct extent_state *state;
1145
1146         /*
1147          * this search will find all the extents that end after
1148          * our range starts.
1149          */
1150         node = tree_search(tree, start);
1151         if (!node)
1152                 goto out;
1153
1154         while (1) {
1155                 state = rb_entry(node, struct extent_state, rb_node);
1156                 if (state->end >= start && (state->state & bits))
1157                         return state;
1158
1159                 node = rb_next(node);
1160                 if (!node)
1161                         break;
1162         }
1163 out:
1164         return NULL;
1165 }
1166
1167 /*
1168  * find a contiguous range of bytes in the file marked as delalloc, not
1169  * more than 'max_bytes'.  start and end are used to return the range,
1170  *
1171  * 1 is returned if we find something, 0 if nothing was in the tree
1172  */
1173 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1174                                         u64 *start, u64 *end, u64 max_bytes)
1175 {
1176         struct rb_node *node;
1177         struct extent_state *state;
1178         u64 cur_start = *start;
1179         u64 found = 0;
1180         u64 total_bytes = 0;
1181
1182         spin_lock(&tree->lock);
1183
1184         /*
1185          * this search will find all the extents that end after
1186          * our range starts.
1187          */
1188         node = tree_search(tree, cur_start);
1189         if (!node) {
1190                 if (!found)
1191                         *end = (u64)-1;
1192                 goto out;
1193         }
1194
1195         while (1) {
1196                 state = rb_entry(node, struct extent_state, rb_node);
1197                 if (found && (state->start != cur_start ||
1198                               (state->state & EXTENT_BOUNDARY))) {
1199                         goto out;
1200                 }
1201                 if (!(state->state & EXTENT_DELALLOC)) {
1202                         if (!found)
1203                                 *end = state->end;
1204                         goto out;
1205                 }
1206                 if (!found)
1207                         *start = state->start;
1208                 found++;
1209                 *end = state->end;
1210                 cur_start = state->end + 1;
1211                 node = rb_next(node);
1212                 if (!node)
1213                         break;
1214                 total_bytes += state->end - state->start + 1;
1215                 if (total_bytes >= max_bytes)
1216                         break;
1217         }
1218 out:
1219         spin_unlock(&tree->lock);
1220         return found;
1221 }
1222
1223 static noinline int __unlock_for_delalloc(struct inode *inode,
1224                                           struct page *locked_page,
1225                                           u64 start, u64 end)
1226 {
1227         int ret;
1228         struct page *pages[16];
1229         unsigned long index = start >> PAGE_CACHE_SHIFT;
1230         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1231         unsigned long nr_pages = end_index - index + 1;
1232         int i;
1233
1234         if (index == locked_page->index && end_index == index)
1235                 return 0;
1236
1237         while (nr_pages > 0) {
1238                 ret = find_get_pages_contig(inode->i_mapping, index,
1239                                      min_t(unsigned long, nr_pages,
1240                                      ARRAY_SIZE(pages)), pages);
1241                 for (i = 0; i < ret; i++) {
1242                         if (pages[i] != locked_page)
1243                                 unlock_page(pages[i]);
1244                         page_cache_release(pages[i]);
1245                 }
1246                 nr_pages -= ret;
1247                 index += ret;
1248                 cond_resched();
1249         }
1250         return 0;
1251 }
1252
1253 static noinline int lock_delalloc_pages(struct inode *inode,
1254                                         struct page *locked_page,
1255                                         u64 delalloc_start,
1256                                         u64 delalloc_end)
1257 {
1258         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1259         unsigned long start_index = index;
1260         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1261         unsigned long pages_locked = 0;
1262         struct page *pages[16];
1263         unsigned long nrpages;
1264         int ret;
1265         int i;
1266
1267         /* the caller is responsible for locking the start index */
1268         if (index == locked_page->index && index == end_index)
1269                 return 0;
1270
1271         /* skip the page at the start index */
1272         nrpages = end_index - index + 1;
1273         while (nrpages > 0) {
1274                 ret = find_get_pages_contig(inode->i_mapping, index,
1275                                      min_t(unsigned long,
1276                                      nrpages, ARRAY_SIZE(pages)), pages);
1277                 if (ret == 0) {
1278                         ret = -EAGAIN;
1279                         goto done;
1280                 }
1281                 /* now we have an array of pages, lock them all */
1282                 for (i = 0; i < ret; i++) {
1283                         /*
1284                          * the caller is taking responsibility for
1285                          * locked_page
1286                          */
1287                         if (pages[i] != locked_page) {
1288                                 lock_page(pages[i]);
1289                                 if (!PageDirty(pages[i]) ||
1290                                     pages[i]->mapping != inode->i_mapping) {
1291                                         ret = -EAGAIN;
1292                                         unlock_page(pages[i]);
1293                                         page_cache_release(pages[i]);
1294                                         goto done;
1295                                 }
1296                         }
1297                         page_cache_release(pages[i]);
1298                         pages_locked++;
1299                 }
1300                 nrpages -= ret;
1301                 index += ret;
1302                 cond_resched();
1303         }
1304         ret = 0;
1305 done:
1306         if (ret && pages_locked) {
1307                 __unlock_for_delalloc(inode, locked_page,
1308                               delalloc_start,
1309                               ((u64)(start_index + pages_locked - 1)) <<
1310                               PAGE_CACHE_SHIFT);
1311         }
1312         return ret;
1313 }
1314
1315 /*
1316  * find a contiguous range of bytes in the file marked as delalloc, not
1317  * more than 'max_bytes'.  start and end are used to return the range,
1318  *
1319  * 1 is returned if we find something, 0 if nothing was in the tree
1320  */
1321 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1322                                              struct extent_io_tree *tree,
1323                                              struct page *locked_page,
1324                                              u64 *start, u64 *end,
1325                                              u64 max_bytes)
1326 {
1327         u64 delalloc_start;
1328         u64 delalloc_end;
1329         u64 found;
1330         struct extent_state *cached_state = NULL;
1331         int ret;
1332         int loops = 0;
1333
1334 again:
1335         /* step one, find a bunch of delalloc bytes starting at start */
1336         delalloc_start = *start;
1337         delalloc_end = 0;
1338         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1339                                     max_bytes);
1340         if (!found || delalloc_end <= *start) {
1341                 *start = delalloc_start;
1342                 *end = delalloc_end;
1343                 return found;
1344         }
1345
1346         /*
1347          * start comes from the offset of locked_page.  We have to lock
1348          * pages in order, so we can't process delalloc bytes before
1349          * locked_page
1350          */
1351         if (delalloc_start < *start)
1352                 delalloc_start = *start;
1353
1354         /*
1355          * make sure to limit the number of pages we try to lock down
1356          * if we're looping.
1357          */
1358         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1359                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1360
1361         /* step two, lock all the pages after the page that has start */
1362         ret = lock_delalloc_pages(inode, locked_page,
1363                                   delalloc_start, delalloc_end);
1364         if (ret == -EAGAIN) {
1365                 /* some of the pages are gone, lets avoid looping by
1366                  * shortening the size of the delalloc range we're searching
1367                  */
1368                 free_extent_state(cached_state);
1369                 if (!loops) {
1370                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1371                         max_bytes = PAGE_CACHE_SIZE - offset;
1372                         loops = 1;
1373                         goto again;
1374                 } else {
1375                         found = 0;
1376                         goto out_failed;
1377                 }
1378         }
1379         BUG_ON(ret);
1380
1381         /* step three, lock the state bits for the whole range */
1382         lock_extent_bits(tree, delalloc_start, delalloc_end,
1383                          0, &cached_state, GFP_NOFS);
1384
1385         /* then test to make sure it is all still delalloc */
1386         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1387                              EXTENT_DELALLOC, 1, cached_state);
1388         if (!ret) {
1389                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1390                                      &cached_state, GFP_NOFS);
1391                 __unlock_for_delalloc(inode, locked_page,
1392                               delalloc_start, delalloc_end);
1393                 cond_resched();
1394                 goto again;
1395         }
1396         free_extent_state(cached_state);
1397         *start = delalloc_start;
1398         *end = delalloc_end;
1399 out_failed:
1400         return found;
1401 }
1402
1403 int extent_clear_unlock_delalloc(struct inode *inode,
1404                                 struct extent_io_tree *tree,
1405                                 u64 start, u64 end, struct page *locked_page,
1406                                 unsigned long op)
1407 {
1408         int ret;
1409         struct page *pages[16];
1410         unsigned long index = start >> PAGE_CACHE_SHIFT;
1411         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1412         unsigned long nr_pages = end_index - index + 1;
1413         int i;
1414         int clear_bits = 0;
1415
1416         if (op & EXTENT_CLEAR_UNLOCK)
1417                 clear_bits |= EXTENT_LOCKED;
1418         if (op & EXTENT_CLEAR_DIRTY)
1419                 clear_bits |= EXTENT_DIRTY;
1420
1421         if (op & EXTENT_CLEAR_DELALLOC)
1422                 clear_bits |= EXTENT_DELALLOC;
1423
1424         if (op & EXTENT_CLEAR_ACCOUNTING)
1425                 clear_bits |= EXTENT_DO_ACCOUNTING;
1426
1427         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1428         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1429                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1430                     EXTENT_SET_PRIVATE2)))
1431                 return 0;
1432
1433         while (nr_pages > 0) {
1434                 ret = find_get_pages_contig(inode->i_mapping, index,
1435                                      min_t(unsigned long,
1436                                      nr_pages, ARRAY_SIZE(pages)), pages);
1437                 for (i = 0; i < ret; i++) {
1438
1439                         if (op & EXTENT_SET_PRIVATE2)
1440                                 SetPagePrivate2(pages[i]);
1441
1442                         if (pages[i] == locked_page) {
1443                                 page_cache_release(pages[i]);
1444                                 continue;
1445                         }
1446                         if (op & EXTENT_CLEAR_DIRTY)
1447                                 clear_page_dirty_for_io(pages[i]);
1448                         if (op & EXTENT_SET_WRITEBACK)
1449                                 set_page_writeback(pages[i]);
1450                         if (op & EXTENT_END_WRITEBACK)
1451                                 end_page_writeback(pages[i]);
1452                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1453                                 unlock_page(pages[i]);
1454                         page_cache_release(pages[i]);
1455                 }
1456                 nr_pages -= ret;
1457                 index += ret;
1458                 cond_resched();
1459         }
1460         return 0;
1461 }
1462
1463 /*
1464  * count the number of bytes in the tree that have a given bit(s)
1465  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1466  * cached.  The total number found is returned.
1467  */
1468 u64 count_range_bits(struct extent_io_tree *tree,
1469                      u64 *start, u64 search_end, u64 max_bytes,
1470                      unsigned long bits)
1471 {
1472         struct rb_node *node;
1473         struct extent_state *state;
1474         u64 cur_start = *start;
1475         u64 total_bytes = 0;
1476         int found = 0;
1477
1478         if (search_end <= cur_start) {
1479                 WARN_ON(1);
1480                 return 0;
1481         }
1482
1483         spin_lock(&tree->lock);
1484         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1485                 total_bytes = tree->dirty_bytes;
1486                 goto out;
1487         }
1488         /*
1489          * this search will find all the extents that end after
1490          * our range starts.
1491          */
1492         node = tree_search(tree, cur_start);
1493         if (!node)
1494                 goto out;
1495
1496         while (1) {
1497                 state = rb_entry(node, struct extent_state, rb_node);
1498                 if (state->start > search_end)
1499                         break;
1500                 if (state->end >= cur_start && (state->state & bits)) {
1501                         total_bytes += min(search_end, state->end) + 1 -
1502                                        max(cur_start, state->start);
1503                         if (total_bytes >= max_bytes)
1504                                 break;
1505                         if (!found) {
1506                                 *start = state->start;
1507                                 found = 1;
1508                         }
1509                 }
1510                 node = rb_next(node);
1511                 if (!node)
1512                         break;
1513         }
1514 out:
1515         spin_unlock(&tree->lock);
1516         return total_bytes;
1517 }
1518
1519 /*
1520  * set the private field for a given byte offset in the tree.  If there isn't
1521  * an extent_state there already, this does nothing.
1522  */
1523 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1524 {
1525         struct rb_node *node;
1526         struct extent_state *state;
1527         int ret = 0;
1528
1529         spin_lock(&tree->lock);
1530         /*
1531          * this search will find all the extents that end after
1532          * our range starts.
1533          */
1534         node = tree_search(tree, start);
1535         if (!node) {
1536                 ret = -ENOENT;
1537                 goto out;
1538         }
1539         state = rb_entry(node, struct extent_state, rb_node);
1540         if (state->start != start) {
1541                 ret = -ENOENT;
1542                 goto out;
1543         }
1544         state->private = private;
1545 out:
1546         spin_unlock(&tree->lock);
1547         return ret;
1548 }
1549
1550 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1551 {
1552         struct rb_node *node;
1553         struct extent_state *state;
1554         int ret = 0;
1555
1556         spin_lock(&tree->lock);
1557         /*
1558          * this search will find all the extents that end after
1559          * our range starts.
1560          */
1561         node = tree_search(tree, start);
1562         if (!node) {
1563                 ret = -ENOENT;
1564                 goto out;
1565         }
1566         state = rb_entry(node, struct extent_state, rb_node);
1567         if (state->start != start) {
1568                 ret = -ENOENT;
1569                 goto out;
1570         }
1571         *private = state->private;
1572 out:
1573         spin_unlock(&tree->lock);
1574         return ret;
1575 }
1576
1577 /*
1578  * searches a range in the state tree for a given mask.
1579  * If 'filled' == 1, this returns 1 only if every extent in the tree
1580  * has the bits set.  Otherwise, 1 is returned if any bit in the
1581  * range is found set.
1582  */
1583 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1584                    int bits, int filled, struct extent_state *cached)
1585 {
1586         struct extent_state *state = NULL;
1587         struct rb_node *node;
1588         int bitset = 0;
1589
1590         spin_lock(&tree->lock);
1591         if (cached && cached->tree && cached->start == start)
1592                 node = &cached->rb_node;
1593         else
1594                 node = tree_search(tree, start);
1595         while (node && start <= end) {
1596                 state = rb_entry(node, struct extent_state, rb_node);
1597
1598                 if (filled && state->start > start) {
1599                         bitset = 0;
1600                         break;
1601                 }
1602
1603                 if (state->start > end)
1604                         break;
1605
1606                 if (state->state & bits) {
1607                         bitset = 1;
1608                         if (!filled)
1609                                 break;
1610                 } else if (filled) {
1611                         bitset = 0;
1612                         break;
1613                 }
1614
1615                 if (state->end == (u64)-1)
1616                         break;
1617
1618                 start = state->end + 1;
1619                 if (start > end)
1620                         break;
1621                 node = rb_next(node);
1622                 if (!node) {
1623                         if (filled)
1624                                 bitset = 0;
1625                         break;
1626                 }
1627         }
1628         spin_unlock(&tree->lock);
1629         return bitset;
1630 }
1631
1632 /*
1633  * helper function to set a given page up to date if all the
1634  * extents in the tree for that page are up to date
1635  */
1636 static int check_page_uptodate(struct extent_io_tree *tree,
1637                                struct page *page)
1638 {
1639         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1640         u64 end = start + PAGE_CACHE_SIZE - 1;
1641         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1642                 SetPageUptodate(page);
1643         return 0;
1644 }
1645
1646 /*
1647  * helper function to unlock a page if all the extents in the tree
1648  * for that page are unlocked
1649  */
1650 static int check_page_locked(struct extent_io_tree *tree,
1651                              struct page *page)
1652 {
1653         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1654         u64 end = start + PAGE_CACHE_SIZE - 1;
1655         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1656                 unlock_page(page);
1657         return 0;
1658 }
1659
1660 /*
1661  * helper function to end page writeback if all the extents
1662  * in the tree for that page are done with writeback
1663  */
1664 static int check_page_writeback(struct extent_io_tree *tree,
1665                              struct page *page)
1666 {
1667         end_page_writeback(page);
1668         return 0;
1669 }
1670
1671 /* lots and lots of room for performance fixes in the end_bio funcs */
1672
1673 /*
1674  * after a writepage IO is done, we need to:
1675  * clear the uptodate bits on error
1676  * clear the writeback bits in the extent tree for this IO
1677  * end_page_writeback if the page has no more pending IO
1678  *
1679  * Scheduling is not allowed, so the extent state tree is expected
1680  * to have one and only one object corresponding to this IO.
1681  */
1682 static void end_bio_extent_writepage(struct bio *bio, int err)
1683 {
1684         int uptodate = err == 0;
1685         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1686         struct extent_io_tree *tree;
1687         u64 start;
1688         u64 end;
1689         int whole_page;
1690         int ret;
1691
1692         do {
1693                 struct page *page = bvec->bv_page;
1694                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1695
1696                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1697                          bvec->bv_offset;
1698                 end = start + bvec->bv_len - 1;
1699
1700                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1701                         whole_page = 1;
1702                 else
1703                         whole_page = 0;
1704
1705                 if (--bvec >= bio->bi_io_vec)
1706                         prefetchw(&bvec->bv_page->flags);
1707                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1708                         ret = tree->ops->writepage_end_io_hook(page, start,
1709                                                        end, NULL, uptodate);
1710                         if (ret)
1711                                 uptodate = 0;
1712                 }
1713
1714                 if (!uptodate && tree->ops &&
1715                     tree->ops->writepage_io_failed_hook) {
1716                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1717                                                          start, end, NULL);
1718                         if (ret == 0) {
1719                                 uptodate = (err == 0);
1720                                 continue;
1721                         }
1722                 }
1723
1724                 if (!uptodate) {
1725                         clear_extent_uptodate(tree, start, end, GFP_NOFS);
1726                         ClearPageUptodate(page);
1727                         SetPageError(page);
1728                 }
1729
1730                 if (whole_page)
1731                         end_page_writeback(page);
1732                 else
1733                         check_page_writeback(tree, page);
1734         } while (bvec >= bio->bi_io_vec);
1735
1736         bio_put(bio);
1737 }
1738
1739 /*
1740  * after a readpage IO is done, we need to:
1741  * clear the uptodate bits on error
1742  * set the uptodate bits if things worked
1743  * set the page up to date if all extents in the tree are uptodate
1744  * clear the lock bit in the extent tree
1745  * unlock the page if there are no other extents locked for it
1746  *
1747  * Scheduling is not allowed, so the extent state tree is expected
1748  * to have one and only one object corresponding to this IO.
1749  */
1750 static void end_bio_extent_readpage(struct bio *bio, int err)
1751 {
1752         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1753         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1754         struct extent_io_tree *tree;
1755         u64 start;
1756         u64 end;
1757         int whole_page;
1758         int ret;
1759
1760         if (err)
1761                 uptodate = 0;
1762
1763         do {
1764                 struct page *page = bvec->bv_page;
1765                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1766
1767                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1768                         bvec->bv_offset;
1769                 end = start + bvec->bv_len - 1;
1770
1771                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1772                         whole_page = 1;
1773                 else
1774                         whole_page = 0;
1775
1776                 if (--bvec >= bio->bi_io_vec)
1777                         prefetchw(&bvec->bv_page->flags);
1778
1779                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1780                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1781                                                               NULL);
1782                         if (ret)
1783                                 uptodate = 0;
1784                 }
1785                 if (!uptodate && tree->ops &&
1786                     tree->ops->readpage_io_failed_hook) {
1787                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1788                                                          start, end, NULL);
1789                         if (ret == 0) {
1790                                 uptodate =
1791                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1792                                 if (err)
1793                                         uptodate = 0;
1794                                 continue;
1795                         }
1796                 }
1797
1798                 if (uptodate) {
1799                         set_extent_uptodate(tree, start, end,
1800                                             GFP_ATOMIC);
1801                 }
1802                 unlock_extent(tree, start, end, GFP_ATOMIC);
1803
1804                 if (whole_page) {
1805                         if (uptodate) {
1806                                 SetPageUptodate(page);
1807                         } else {
1808                                 ClearPageUptodate(page);
1809                                 SetPageError(page);
1810                         }
1811                         unlock_page(page);
1812                 } else {
1813                         if (uptodate) {
1814                                 check_page_uptodate(tree, page);
1815                         } else {
1816                                 ClearPageUptodate(page);
1817                                 SetPageError(page);
1818                         }
1819                         check_page_locked(tree, page);
1820                 }
1821         } while (bvec >= bio->bi_io_vec);
1822
1823         bio_put(bio);
1824 }
1825
1826 /*
1827  * IO done from prepare_write is pretty simple, we just unlock
1828  * the structs in the extent tree when done, and set the uptodate bits
1829  * as appropriate.
1830  */
1831 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1832 {
1833         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1834         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1835         struct extent_io_tree *tree;
1836         u64 start;
1837         u64 end;
1838
1839         do {
1840                 struct page *page = bvec->bv_page;
1841                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1842
1843                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1844                         bvec->bv_offset;
1845                 end = start + bvec->bv_len - 1;
1846
1847                 if (--bvec >= bio->bi_io_vec)
1848                         prefetchw(&bvec->bv_page->flags);
1849
1850                 if (uptodate) {
1851                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1852                 } else {
1853                         ClearPageUptodate(page);
1854                         SetPageError(page);
1855                 }
1856
1857                 unlock_extent(tree, start, end, GFP_ATOMIC);
1858
1859         } while (bvec >= bio->bi_io_vec);
1860
1861         bio_put(bio);
1862 }
1863
1864 static struct bio *
1865 extent_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1866                  gfp_t gfp_flags)
1867 {
1868         struct bio *bio;
1869
1870         bio = bio_alloc(gfp_flags, nr_vecs);
1871
1872         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1873                 while (!bio && (nr_vecs /= 2))
1874                         bio = bio_alloc(gfp_flags, nr_vecs);
1875         }
1876
1877         if (bio) {
1878                 bio->bi_size = 0;
1879                 bio->bi_bdev = bdev;
1880                 bio->bi_sector = first_sector;
1881         }
1882         return bio;
1883 }
1884
1885 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1886                           unsigned long bio_flags)
1887 {
1888         int ret = 0;
1889         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1890         struct page *page = bvec->bv_page;
1891         struct extent_io_tree *tree = bio->bi_private;
1892         u64 start;
1893         u64 end;
1894
1895         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1896         end = start + bvec->bv_len - 1;
1897
1898         bio->bi_private = NULL;
1899
1900         bio_get(bio);
1901
1902         if (tree->ops && tree->ops->submit_bio_hook)
1903                 tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1904                                            mirror_num, bio_flags);
1905         else
1906                 submit_bio(rw, bio);
1907         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1908                 ret = -EOPNOTSUPP;
1909         bio_put(bio);
1910         return ret;
1911 }
1912
1913 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1914                               struct page *page, sector_t sector,
1915                               size_t size, unsigned long offset,
1916                               struct block_device *bdev,
1917                               struct bio **bio_ret,
1918                               unsigned long max_pages,
1919                               bio_end_io_t end_io_func,
1920                               int mirror_num,
1921                               unsigned long prev_bio_flags,
1922                               unsigned long bio_flags)
1923 {
1924         int ret = 0;
1925         struct bio *bio;
1926         int nr;
1927         int contig = 0;
1928         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1929         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1930         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1931
1932         if (bio_ret && *bio_ret) {
1933                 bio = *bio_ret;
1934                 if (old_compressed)
1935                         contig = bio->bi_sector == sector;
1936                 else
1937                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
1938                                 sector;
1939
1940                 if (prev_bio_flags != bio_flags || !contig ||
1941                     (tree->ops && tree->ops->merge_bio_hook &&
1942                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
1943                                                bio_flags)) ||
1944                     bio_add_page(bio, page, page_size, offset) < page_size) {
1945                         ret = submit_one_bio(rw, bio, mirror_num,
1946                                              prev_bio_flags);
1947                         bio = NULL;
1948                 } else {
1949                         return 0;
1950                 }
1951         }
1952         if (this_compressed)
1953                 nr = BIO_MAX_PAGES;
1954         else
1955                 nr = bio_get_nr_vecs(bdev);
1956
1957         bio = extent_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1958
1959         bio_add_page(bio, page, page_size, offset);
1960         bio->bi_end_io = end_io_func;
1961         bio->bi_private = tree;
1962
1963         if (bio_ret)
1964                 *bio_ret = bio;
1965         else
1966                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1967
1968         return ret;
1969 }
1970
1971 void set_page_extent_mapped(struct page *page)
1972 {
1973         if (!PagePrivate(page)) {
1974                 SetPagePrivate(page);
1975                 page_cache_get(page);
1976                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1977         }
1978 }
1979
1980 static void set_page_extent_head(struct page *page, unsigned long len)
1981 {
1982         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1983 }
1984
1985 /*
1986  * basic readpage implementation.  Locked extent state structs are inserted
1987  * into the tree that are removed when the IO is done (by the end_io
1988  * handlers)
1989  */
1990 static int __extent_read_full_page(struct extent_io_tree *tree,
1991                                    struct page *page,
1992                                    get_extent_t *get_extent,
1993                                    struct bio **bio, int mirror_num,
1994                                    unsigned long *bio_flags)
1995 {
1996         struct inode *inode = page->mapping->host;
1997         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1998         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1999         u64 end;
2000         u64 cur = start;
2001         u64 extent_offset;
2002         u64 last_byte = i_size_read(inode);
2003         u64 block_start;
2004         u64 cur_end;
2005         sector_t sector;
2006         struct extent_map *em;
2007         struct block_device *bdev;
2008         int ret;
2009         int nr = 0;
2010         size_t page_offset = 0;
2011         size_t iosize;
2012         size_t disk_io_size;
2013         size_t blocksize = inode->i_sb->s_blocksize;
2014         unsigned long this_bio_flag = 0;
2015
2016         set_page_extent_mapped(page);
2017
2018         end = page_end;
2019         lock_extent(tree, start, end, GFP_NOFS);
2020
2021         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2022                 char *userpage;
2023                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2024
2025                 if (zero_offset) {
2026                         iosize = PAGE_CACHE_SIZE - zero_offset;
2027                         userpage = kmap_atomic(page, KM_USER0);
2028                         memset(userpage + zero_offset, 0, iosize);
2029                         flush_dcache_page(page);
2030                         kunmap_atomic(userpage, KM_USER0);
2031                 }
2032         }
2033         while (cur <= end) {
2034                 if (cur >= last_byte) {
2035                         char *userpage;
2036                         iosize = PAGE_CACHE_SIZE - page_offset;
2037                         userpage = kmap_atomic(page, KM_USER0);
2038                         memset(userpage + page_offset, 0, iosize);
2039                         flush_dcache_page(page);
2040                         kunmap_atomic(userpage, KM_USER0);
2041                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2042                                             GFP_NOFS);
2043                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2044                         break;
2045                 }
2046                 em = get_extent(inode, page, page_offset, cur,
2047                                 end - cur + 1, 0);
2048                 if (IS_ERR(em) || !em) {
2049                         SetPageError(page);
2050                         unlock_extent(tree, cur, end, GFP_NOFS);
2051                         break;
2052                 }
2053                 extent_offset = cur - em->start;
2054                 BUG_ON(extent_map_end(em) <= cur);
2055                 BUG_ON(end < cur);
2056
2057                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2058                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2059
2060                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2061                 cur_end = min(extent_map_end(em) - 1, end);
2062                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2063                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2064                         disk_io_size = em->block_len;
2065                         sector = em->block_start >> 9;
2066                 } else {
2067                         sector = (em->block_start + extent_offset) >> 9;
2068                         disk_io_size = iosize;
2069                 }
2070                 bdev = em->bdev;
2071                 block_start = em->block_start;
2072                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2073                         block_start = EXTENT_MAP_HOLE;
2074                 free_extent_map(em);
2075                 em = NULL;
2076
2077                 /* we've found a hole, just zero and go on */
2078                 if (block_start == EXTENT_MAP_HOLE) {
2079                         char *userpage;
2080                         userpage = kmap_atomic(page, KM_USER0);
2081                         memset(userpage + page_offset, 0, iosize);
2082                         flush_dcache_page(page);
2083                         kunmap_atomic(userpage, KM_USER0);
2084
2085                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2086                                             GFP_NOFS);
2087                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2088                         cur = cur + iosize;
2089                         page_offset += iosize;
2090                         continue;
2091                 }
2092                 /* the get_extent function already copied into the page */
2093                 if (test_range_bit(tree, cur, cur_end,
2094                                    EXTENT_UPTODATE, 1, NULL)) {
2095                         check_page_uptodate(tree, page);
2096                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2097                         cur = cur + iosize;
2098                         page_offset += iosize;
2099                         continue;
2100                 }
2101                 /* we have an inline extent but it didn't get marked up
2102                  * to date.  Error out
2103                  */
2104                 if (block_start == EXTENT_MAP_INLINE) {
2105                         SetPageError(page);
2106                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2107                         cur = cur + iosize;
2108                         page_offset += iosize;
2109                         continue;
2110                 }
2111
2112                 ret = 0;
2113                 if (tree->ops && tree->ops->readpage_io_hook) {
2114                         ret = tree->ops->readpage_io_hook(page, cur,
2115                                                           cur + iosize - 1);
2116                 }
2117                 if (!ret) {
2118                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2119                         pnr -= page->index;
2120                         ret = submit_extent_page(READ, tree, page,
2121                                          sector, disk_io_size, page_offset,
2122                                          bdev, bio, pnr,
2123                                          end_bio_extent_readpage, mirror_num,
2124                                          *bio_flags,
2125                                          this_bio_flag);
2126                         nr++;
2127                         *bio_flags = this_bio_flag;
2128                 }
2129                 if (ret)
2130                         SetPageError(page);
2131                 cur = cur + iosize;
2132                 page_offset += iosize;
2133         }
2134         if (!nr) {
2135                 if (!PageError(page))
2136                         SetPageUptodate(page);
2137                 unlock_page(page);
2138         }
2139         return 0;
2140 }
2141
2142 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2143                             get_extent_t *get_extent)
2144 {
2145         struct bio *bio = NULL;
2146         unsigned long bio_flags = 0;
2147         int ret;
2148
2149         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2150                                       &bio_flags);
2151         if (bio)
2152                 submit_one_bio(READ, bio, 0, bio_flags);
2153         return ret;
2154 }
2155
2156 static noinline void update_nr_written(struct page *page,
2157                                       struct writeback_control *wbc,
2158                                       unsigned long nr_written)
2159 {
2160         wbc->nr_to_write -= nr_written;
2161         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2162             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2163                 page->mapping->writeback_index = page->index + nr_written;
2164 }
2165
2166 /*
2167  * the writepage semantics are similar to regular writepage.  extent
2168  * records are inserted to lock ranges in the tree, and as dirty areas
2169  * are found, they are marked writeback.  Then the lock bits are removed
2170  * and the end_io handler clears the writeback ranges
2171  */
2172 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2173                               void *data)
2174 {
2175         struct inode *inode = page->mapping->host;
2176         struct extent_page_data *epd = data;
2177         struct extent_io_tree *tree = epd->tree;
2178         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2179         u64 delalloc_start;
2180         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2181         u64 end;
2182         u64 cur = start;
2183         u64 extent_offset;
2184         u64 last_byte = i_size_read(inode);
2185         u64 block_start;
2186         u64 iosize;
2187         u64 unlock_start;
2188         sector_t sector;
2189         struct extent_state *cached_state = NULL;
2190         struct extent_map *em;
2191         struct block_device *bdev;
2192         int ret;
2193         int nr = 0;
2194         size_t pg_offset = 0;
2195         size_t blocksize;
2196         loff_t i_size = i_size_read(inode);
2197         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2198         u64 nr_delalloc;
2199         u64 delalloc_end;
2200         int page_started;
2201         int compressed;
2202         int write_flags;
2203         unsigned long nr_written = 0;
2204
2205         if (wbc->sync_mode == WB_SYNC_ALL)
2206                 write_flags = WRITE_SYNC_PLUG;
2207         else
2208                 write_flags = WRITE;
2209
2210         WARN_ON(!PageLocked(page));
2211         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2212         if (page->index > end_index ||
2213            (page->index == end_index && !pg_offset)) {
2214                 page->mapping->a_ops->invalidatepage(page, 0);
2215                 unlock_page(page);
2216                 return 0;
2217         }
2218
2219         if (page->index == end_index) {
2220                 char *userpage;
2221
2222                 userpage = kmap_atomic(page, KM_USER0);
2223                 memset(userpage + pg_offset, 0,
2224                        PAGE_CACHE_SIZE - pg_offset);
2225                 kunmap_atomic(userpage, KM_USER0);
2226                 flush_dcache_page(page);
2227         }
2228         pg_offset = 0;
2229
2230         set_page_extent_mapped(page);
2231
2232         delalloc_start = start;
2233         delalloc_end = 0;
2234         page_started = 0;
2235         if (!epd->extent_locked) {
2236                 u64 delalloc_to_write = 0;
2237                 /*
2238                  * make sure the wbc mapping index is at least updated
2239                  * to this page.
2240                  */
2241                 update_nr_written(page, wbc, 0);
2242
2243                 while (delalloc_end < page_end) {
2244                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2245                                                        page,
2246                                                        &delalloc_start,
2247                                                        &delalloc_end,
2248                                                        128 * 1024 * 1024);
2249                         if (nr_delalloc == 0) {
2250                                 delalloc_start = delalloc_end + 1;
2251                                 continue;
2252                         }
2253                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2254                                                  delalloc_end, &page_started,
2255                                                  &nr_written);
2256                         /*
2257                          * delalloc_end is already one less than the total
2258                          * length, so we don't subtract one from
2259                          * PAGE_CACHE_SIZE
2260                          */
2261                         delalloc_to_write += (delalloc_end - delalloc_start +
2262                                               PAGE_CACHE_SIZE) >>
2263                                               PAGE_CACHE_SHIFT;
2264                         delalloc_start = delalloc_end + 1;
2265                 }
2266                 if (wbc->nr_to_write < delalloc_to_write) {
2267                         int thresh = 8192;
2268
2269                         if (delalloc_to_write < thresh * 2)
2270                                 thresh = delalloc_to_write;
2271                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2272                                                  thresh);
2273                 }
2274
2275                 /* did the fill delalloc function already unlock and start
2276                  * the IO?
2277                  */
2278                 if (page_started) {
2279                         ret = 0;
2280                         /*
2281                          * we've unlocked the page, so we can't update
2282                          * the mapping's writeback index, just update
2283                          * nr_to_write.
2284                          */
2285                         wbc->nr_to_write -= nr_written;
2286                         goto done_unlocked;
2287                 }
2288         }
2289         if (tree->ops && tree->ops->writepage_start_hook) {
2290                 ret = tree->ops->writepage_start_hook(page, start,
2291                                                       page_end);
2292                 if (ret == -EAGAIN) {
2293                         redirty_page_for_writepage(wbc, page);
2294                         update_nr_written(page, wbc, nr_written);
2295                         unlock_page(page);
2296                         ret = 0;
2297                         goto done_unlocked;
2298                 }
2299         }
2300
2301         /*
2302          * we don't want to touch the inode after unlocking the page,
2303          * so we update the mapping writeback index now
2304          */
2305         update_nr_written(page, wbc, nr_written + 1);
2306
2307         end = page_end;
2308         if (last_byte <= start) {
2309                 if (tree->ops && tree->ops->writepage_end_io_hook)
2310                         tree->ops->writepage_end_io_hook(page, start,
2311                                                          page_end, NULL, 1);
2312                 unlock_start = page_end + 1;
2313                 goto done;
2314         }
2315
2316         blocksize = inode->i_sb->s_blocksize;
2317
2318         while (cur <= end) {
2319                 if (cur >= last_byte) {
2320                         if (tree->ops && tree->ops->writepage_end_io_hook)
2321                                 tree->ops->writepage_end_io_hook(page, cur,
2322                                                          page_end, NULL, 1);
2323                         unlock_start = page_end + 1;
2324                         break;
2325                 }
2326                 em = epd->get_extent(inode, page, pg_offset, cur,
2327                                      end - cur + 1, 1);
2328                 if (IS_ERR(em) || !em) {
2329                         SetPageError(page);
2330                         break;
2331                 }
2332
2333                 extent_offset = cur - em->start;
2334                 BUG_ON(extent_map_end(em) <= cur);
2335                 BUG_ON(end < cur);
2336                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2337                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2338                 sector = (em->block_start + extent_offset) >> 9;
2339                 bdev = em->bdev;
2340                 block_start = em->block_start;
2341                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2342                 free_extent_map(em);
2343                 em = NULL;
2344
2345                 /*
2346                  * compressed and inline extents are written through other
2347                  * paths in the FS
2348                  */
2349                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2350                     block_start == EXTENT_MAP_INLINE) {
2351                         /*
2352                          * end_io notification does not happen here for
2353                          * compressed extents
2354                          */
2355                         if (!compressed && tree->ops &&
2356                             tree->ops->writepage_end_io_hook)
2357                                 tree->ops->writepage_end_io_hook(page, cur,
2358                                                          cur + iosize - 1,
2359                                                          NULL, 1);
2360                         else if (compressed) {
2361                                 /* we don't want to end_page_writeback on
2362                                  * a compressed extent.  this happens
2363                                  * elsewhere
2364                                  */
2365                                 nr++;
2366                         }
2367
2368                         cur += iosize;
2369                         pg_offset += iosize;
2370                         unlock_start = cur;
2371                         continue;
2372                 }
2373                 /* leave this out until we have a page_mkwrite call */
2374                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2375                                    EXTENT_DIRTY, 0, NULL)) {
2376                         cur = cur + iosize;
2377                         pg_offset += iosize;
2378                         continue;
2379                 }
2380
2381                 if (tree->ops && tree->ops->writepage_io_hook) {
2382                         ret = tree->ops->writepage_io_hook(page, cur,
2383                                                 cur + iosize - 1);
2384                 } else {
2385                         ret = 0;
2386                 }
2387                 if (ret) {
2388                         SetPageError(page);
2389                 } else {
2390                         unsigned long max_nr = end_index + 1;
2391
2392                         set_range_writeback(tree, cur, cur + iosize - 1);
2393                         if (!PageWriteback(page)) {
2394                                 printk(KERN_ERR "btrfs warning page %lu not "
2395                                        "writeback, cur %llu end %llu\n",
2396                                        page->index, (unsigned long long)cur,
2397                                        (unsigned long long)end);
2398                         }
2399
2400                         ret = submit_extent_page(write_flags, tree, page,
2401                                                  sector, iosize, pg_offset,
2402                                                  bdev, &epd->bio, max_nr,
2403                                                  end_bio_extent_writepage,
2404                                                  0, 0, 0);
2405                         if (ret)
2406                                 SetPageError(page);
2407                 }
2408                 cur = cur + iosize;
2409                 pg_offset += iosize;
2410                 nr++;
2411         }
2412 done:
2413         if (nr == 0) {
2414                 /* make sure the mapping tag for page dirty gets cleared */
2415                 set_page_writeback(page);
2416                 end_page_writeback(page);
2417         }
2418         unlock_page(page);
2419
2420 done_unlocked:
2421
2422         /* drop our reference on any cached states */
2423         free_extent_state(cached_state);
2424         return 0;
2425 }
2426
2427 /**
2428  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2429  * @mapping: address space structure to write
2430  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2431  * @writepage: function called for each page
2432  * @data: data passed to writepage function
2433  *
2434  * If a page is already under I/O, write_cache_pages() skips it, even
2435  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2436  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2437  * and msync() need to guarantee that all the data which was dirty at the time
2438  * the call was made get new I/O started against them.  If wbc->sync_mode is
2439  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2440  * existing IO to complete.
2441  */
2442 static int extent_write_cache_pages(struct extent_io_tree *tree,
2443                              struct address_space *mapping,
2444                              struct writeback_control *wbc,
2445                              writepage_t writepage, void *data,
2446                              void (*flush_fn)(void *))
2447 {
2448         int ret = 0;
2449         int done = 0;
2450         int nr_to_write_done = 0;
2451         struct pagevec pvec;
2452         int nr_pages;
2453         pgoff_t index;
2454         pgoff_t end;            /* Inclusive */
2455         int scanned = 0;
2456         int range_whole = 0;
2457
2458         pagevec_init(&pvec, 0);
2459         if (wbc->range_cyclic) {
2460                 index = mapping->writeback_index; /* Start from prev offset */
2461                 end = -1;
2462         } else {
2463                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2464                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2465                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2466                         range_whole = 1;
2467                 scanned = 1;
2468         }
2469 retry:
2470         while (!done && !nr_to_write_done && (index <= end) &&
2471                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2472                               PAGECACHE_TAG_DIRTY, min(end - index,
2473                                   (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2474                 unsigned i;
2475
2476                 scanned = 1;
2477                 for (i = 0; i < nr_pages; i++) {
2478                         struct page *page = pvec.pages[i];
2479
2480                         /*
2481                          * At this point we hold neither mapping->tree_lock nor
2482                          * lock on the page itself: the page may be truncated or
2483                          * invalidated (changing page->mapping to NULL), or even
2484                          * swizzled back from swapper_space to tmpfs file
2485                          * mapping
2486                          */
2487                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2488                                 tree->ops->write_cache_pages_lock_hook(page);
2489                         else
2490                                 lock_page(page);
2491
2492                         if (unlikely(page->mapping != mapping)) {
2493                                 unlock_page(page);
2494                                 continue;
2495                         }
2496
2497                         if (!wbc->range_cyclic && page->index > end) {
2498                                 done = 1;
2499                                 unlock_page(page);
2500                                 continue;
2501                         }
2502
2503                         if (wbc->sync_mode != WB_SYNC_NONE) {
2504                                 if (PageWriteback(page))
2505                                         flush_fn(data);
2506                                 wait_on_page_writeback(page);
2507                         }
2508
2509                         if (PageWriteback(page) ||
2510                             !clear_page_dirty_for_io(page)) {
2511                                 unlock_page(page);
2512                                 continue;
2513                         }
2514
2515                         ret = (*writepage)(page, wbc, data);
2516
2517                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2518                                 unlock_page(page);
2519                                 ret = 0;
2520                         }
2521                         if (ret)
2522                                 done = 1;
2523
2524                         /*
2525                          * the filesystem may choose to bump up nr_to_write.
2526                          * We have to make sure to honor the new nr_to_write
2527                          * at any time
2528                          */
2529                         nr_to_write_done = wbc->nr_to_write <= 0;
2530                 }
2531                 pagevec_release(&pvec);
2532                 cond_resched();
2533         }
2534         if (!scanned && !done) {
2535                 /*
2536                  * We hit the last page and there is more work to be done: wrap
2537                  * back to the start of the file
2538                  */
2539                 scanned = 1;
2540                 index = 0;
2541                 goto retry;
2542         }
2543         return ret;
2544 }
2545
2546 static void flush_epd_write_bio(struct extent_page_data *epd)
2547 {
2548         if (epd->bio) {
2549                 if (epd->sync_io)
2550                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2551                 else
2552                         submit_one_bio(WRITE, epd->bio, 0, 0);
2553                 epd->bio = NULL;
2554         }
2555 }
2556
2557 static noinline void flush_write_bio(void *data)
2558 {
2559         struct extent_page_data *epd = data;
2560         flush_epd_write_bio(epd);
2561 }
2562
2563 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2564                           get_extent_t *get_extent,
2565                           struct writeback_control *wbc)
2566 {
2567         int ret;
2568         struct address_space *mapping = page->mapping;
2569         struct extent_page_data epd = {
2570                 .bio = NULL,
2571                 .tree = tree,
2572                 .get_extent = get_extent,
2573                 .extent_locked = 0,
2574                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2575         };
2576         struct writeback_control wbc_writepages = {
2577                 .bdi            = wbc->bdi,
2578                 .sync_mode      = wbc->sync_mode,
2579                 .older_than_this = NULL,
2580                 .nr_to_write    = 64,
2581                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2582                 .range_end      = (loff_t)-1,
2583         };
2584
2585         ret = __extent_writepage(page, wbc, &epd);
2586
2587         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2588                                  __extent_writepage, &epd, flush_write_bio);
2589         flush_epd_write_bio(&epd);
2590         return ret;
2591 }
2592
2593 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2594                               u64 start, u64 end, get_extent_t *get_extent,
2595                               int mode)
2596 {
2597         int ret = 0;
2598         struct address_space *mapping = inode->i_mapping;
2599         struct page *page;
2600         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2601                 PAGE_CACHE_SHIFT;
2602
2603         struct extent_page_data epd = {
2604                 .bio = NULL,
2605                 .tree = tree,
2606                 .get_extent = get_extent,
2607                 .extent_locked = 1,
2608                 .sync_io = mode == WB_SYNC_ALL,
2609         };
2610         struct writeback_control wbc_writepages = {
2611                 .bdi            = inode->i_mapping->backing_dev_info,
2612                 .sync_mode      = mode,
2613                 .older_than_this = NULL,
2614                 .nr_to_write    = nr_pages * 2,
2615                 .range_start    = start,
2616                 .range_end      = end + 1,
2617         };
2618
2619         while (start <= end) {
2620                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2621                 if (clear_page_dirty_for_io(page))
2622                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2623                 else {
2624                         if (tree->ops && tree->ops->writepage_end_io_hook)
2625                                 tree->ops->writepage_end_io_hook(page, start,
2626                                                  start + PAGE_CACHE_SIZE - 1,
2627                                                  NULL, 1);
2628                         unlock_page(page);
2629                 }
2630                 page_cache_release(page);
2631                 start += PAGE_CACHE_SIZE;
2632         }
2633
2634         flush_epd_write_bio(&epd);
2635         return ret;
2636 }
2637
2638 int extent_writepages(struct extent_io_tree *tree,
2639                       struct address_space *mapping,
2640                       get_extent_t *get_extent,
2641                       struct writeback_control *wbc)
2642 {
2643         int ret = 0;
2644         struct extent_page_data epd = {
2645                 .bio = NULL,
2646                 .tree = tree,
2647                 .get_extent = get_extent,
2648                 .extent_locked = 0,
2649                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2650         };
2651
2652         ret = extent_write_cache_pages(tree, mapping, wbc,
2653                                        __extent_writepage, &epd,
2654                                        flush_write_bio);
2655         flush_epd_write_bio(&epd);
2656         return ret;
2657 }
2658
2659 int extent_readpages(struct extent_io_tree *tree,
2660                      struct address_space *mapping,
2661                      struct list_head *pages, unsigned nr_pages,
2662                      get_extent_t get_extent)
2663 {
2664         struct bio *bio = NULL;
2665         unsigned page_idx;
2666         struct pagevec pvec;
2667         unsigned long bio_flags = 0;
2668
2669         pagevec_init(&pvec, 0);
2670         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2671                 struct page *page = list_entry(pages->prev, struct page, lru);
2672
2673                 prefetchw(&page->flags);
2674                 list_del(&page->lru);
2675                 /*
2676                  * what we want to do here is call add_to_page_cache_lru,
2677                  * but that isn't exported, so we reproduce it here
2678                  */
2679                 if (!add_to_page_cache(page, mapping,
2680                                         page->index, GFP_KERNEL)) {
2681
2682                         /* open coding of lru_cache_add, also not exported */
2683                         page_cache_get(page);
2684                         if (!pagevec_add(&pvec, page))
2685                                 __pagevec_lru_add_file(&pvec);
2686                         __extent_read_full_page(tree, page, get_extent,
2687                                                 &bio, 0, &bio_flags);
2688                 }
2689                 page_cache_release(page);
2690         }
2691         if (pagevec_count(&pvec))
2692                 __pagevec_lru_add_file(&pvec);
2693         BUG_ON(!list_empty(pages));
2694         if (bio)
2695                 submit_one_bio(READ, bio, 0, bio_flags);
2696         return 0;
2697 }
2698
2699 /*
2700  * basic invalidatepage code, this waits on any locked or writeback
2701  * ranges corresponding to the page, and then deletes any extent state
2702  * records from the tree
2703  */
2704 int extent_invalidatepage(struct extent_io_tree *tree,
2705                           struct page *page, unsigned long offset)
2706 {
2707         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2708         u64 end = start + PAGE_CACHE_SIZE - 1;
2709         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2710
2711         start += (offset + blocksize - 1) & ~(blocksize - 1);
2712         if (start > end)
2713                 return 0;
2714
2715         lock_extent(tree, start, end, GFP_NOFS);
2716         wait_on_page_writeback(page);
2717         clear_extent_bit(tree, start, end,
2718                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2719                          EXTENT_DO_ACCOUNTING,
2720                          1, 1, NULL, GFP_NOFS);
2721         return 0;
2722 }
2723
2724 /*
2725  * simple commit_write call, set_range_dirty is used to mark both
2726  * the pages and the extent records as dirty
2727  */
2728 int extent_commit_write(struct extent_io_tree *tree,
2729                         struct inode *inode, struct page *page,
2730                         unsigned from, unsigned to)
2731 {
2732         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2733
2734         set_page_extent_mapped(page);
2735         set_page_dirty(page);
2736
2737         if (pos > inode->i_size) {
2738                 i_size_write(inode, pos);
2739                 mark_inode_dirty(inode);
2740         }
2741         return 0;
2742 }
2743
2744 int extent_prepare_write(struct extent_io_tree *tree,
2745                          struct inode *inode, struct page *page,
2746                          unsigned from, unsigned to, get_extent_t *get_extent)
2747 {
2748         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2749         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2750         u64 block_start;
2751         u64 orig_block_start;
2752         u64 block_end;
2753         u64 cur_end;
2754         struct extent_map *em;
2755         unsigned blocksize = 1 << inode->i_blkbits;
2756         size_t page_offset = 0;
2757         size_t block_off_start;
2758         size_t block_off_end;
2759         int err = 0;
2760         int iocount = 0;
2761         int ret = 0;
2762         int isnew;
2763
2764         set_page_extent_mapped(page);
2765
2766         block_start = (page_start + from) & ~((u64)blocksize - 1);
2767         block_end = (page_start + to - 1) | (blocksize - 1);
2768         orig_block_start = block_start;
2769
2770         lock_extent(tree, page_start, page_end, GFP_NOFS);
2771         while (block_start <= block_end) {
2772                 em = get_extent(inode, page, page_offset, block_start,
2773                                 block_end - block_start + 1, 1);
2774                 if (IS_ERR(em) || !em)
2775                         goto err;
2776
2777                 cur_end = min(block_end, extent_map_end(em) - 1);
2778                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2779                 block_off_end = block_off_start + blocksize;
2780                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2781
2782                 if (!PageUptodate(page) && isnew &&
2783                     (block_off_end > to || block_off_start < from)) {
2784                         void *kaddr;
2785
2786                         kaddr = kmap_atomic(page, KM_USER0);
2787                         if (block_off_end > to)
2788                                 memset(kaddr + to, 0, block_off_end - to);
2789                         if (block_off_start < from)
2790                                 memset(kaddr + block_off_start, 0,
2791                                        from - block_off_start);
2792                         flush_dcache_page(page);
2793                         kunmap_atomic(kaddr, KM_USER0);
2794                 }
2795                 if ((em->block_start != EXTENT_MAP_HOLE &&
2796                      em->block_start != EXTENT_MAP_INLINE) &&
2797                     !isnew && !PageUptodate(page) &&
2798                     (block_off_end > to || block_off_start < from) &&
2799                     !test_range_bit(tree, block_start, cur_end,
2800                                     EXTENT_UPTODATE, 1, NULL)) {
2801                         u64 sector;
2802                         u64 extent_offset = block_start - em->start;
2803                         size_t iosize;
2804                         sector = (em->block_start + extent_offset) >> 9;
2805                         iosize = (cur_end - block_start + blocksize) &
2806                                 ~((u64)blocksize - 1);
2807                         /*
2808                          * we've already got the extent locked, but we
2809                          * need to split the state such that our end_bio
2810                          * handler can clear the lock.
2811                          */
2812                         set_extent_bit(tree, block_start,
2813                                        block_start + iosize - 1,
2814                                        EXTENT_LOCKED, 0, NULL, NULL, GFP_NOFS);
2815                         ret = submit_extent_page(READ, tree, page,
2816                                          sector, iosize, page_offset, em->bdev,
2817                                          NULL, 1,
2818                                          end_bio_extent_preparewrite, 0,
2819                                          0, 0);
2820                         iocount++;
2821                         block_start = block_start + iosize;
2822                 } else {
2823                         set_extent_uptodate(tree, block_start, cur_end,
2824                                             GFP_NOFS);
2825                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2826                         block_start = cur_end + 1;
2827                 }
2828                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2829                 free_extent_map(em);
2830         }
2831         if (iocount) {
2832                 wait_extent_bit(tree, orig_block_start,
2833                                 block_end, EXTENT_LOCKED);
2834         }
2835         check_page_uptodate(tree, page);
2836 err:
2837         /* FIXME, zero out newly allocated blocks on error */
2838         return err;
2839 }
2840
2841 /*
2842  * a helper for releasepage, this tests for areas of the page that
2843  * are locked or under IO and drops the related state bits if it is safe
2844  * to drop the page.
2845  */
2846 int try_release_extent_state(struct extent_map_tree *map,
2847                              struct extent_io_tree *tree, struct page *page,
2848                              gfp_t mask)
2849 {
2850         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2851         u64 end = start + PAGE_CACHE_SIZE - 1;
2852         int ret = 1;
2853
2854         if (test_range_bit(tree, start, end,
2855                            EXTENT_IOBITS, 0, NULL))
2856                 ret = 0;
2857         else {
2858                 if ((mask & GFP_NOFS) == GFP_NOFS)
2859                         mask = GFP_NOFS;
2860                 /*
2861                  * at this point we can safely clear everything except the
2862                  * locked bit and the nodatasum bit
2863                  */
2864                 clear_extent_bit(tree, start, end,
2865                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2866                                  0, 0, NULL, mask);
2867         }
2868         return ret;
2869 }
2870
2871 /*
2872  * a helper for releasepage.  As long as there are no locked extents
2873  * in the range corresponding to the page, both state records and extent
2874  * map records are removed
2875  */
2876 int try_release_extent_mapping(struct extent_map_tree *map,
2877                                struct extent_io_tree *tree, struct page *page,
2878                                gfp_t mask)
2879 {
2880         struct extent_map *em;
2881         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2882         u64 end = start + PAGE_CACHE_SIZE - 1;
2883
2884         if ((mask & __GFP_WAIT) &&
2885             page->mapping->host->i_size > 16 * 1024 * 1024) {
2886                 u64 len;
2887                 while (start <= end) {
2888                         len = end - start + 1;
2889                         write_lock(&map->lock);
2890                         em = lookup_extent_mapping(map, start, len);
2891                         if (!em || IS_ERR(em)) {
2892                                 write_unlock(&map->lock);
2893                                 break;
2894                         }
2895                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2896                             em->start != start) {
2897                                 write_unlock(&map->lock);
2898                                 free_extent_map(em);
2899                                 break;
2900                         }
2901                         if (!test_range_bit(tree, em->start,
2902                                             extent_map_end(em) - 1,
2903                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
2904                                             0, NULL)) {
2905                                 remove_extent_mapping(map, em);
2906                                 /* once for the rb tree */
2907                                 free_extent_map(em);
2908                         }
2909                         start = extent_map_end(em);
2910                         write_unlock(&map->lock);
2911
2912                         /* once for us */
2913                         free_extent_map(em);
2914                 }
2915         }
2916         return try_release_extent_state(map, tree, page, mask);
2917 }
2918
2919 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2920                 get_extent_t *get_extent)
2921 {
2922         struct inode *inode = mapping->host;
2923         u64 start = iblock << inode->i_blkbits;
2924         sector_t sector = 0;
2925         size_t blksize = (1 << inode->i_blkbits);
2926         struct extent_map *em;
2927
2928         lock_extent(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2929                     GFP_NOFS);
2930         em = get_extent(inode, NULL, 0, start, blksize, 0);
2931         unlock_extent(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2932                       GFP_NOFS);
2933         if (!em || IS_ERR(em))
2934                 return 0;
2935
2936         if (em->block_start > EXTENT_MAP_LAST_BYTE)
2937                 goto out;
2938
2939         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2940 out:
2941         free_extent_map(em);
2942         return sector;
2943 }
2944
2945 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2946                 __u64 start, __u64 len, get_extent_t *get_extent)
2947 {
2948         int ret;
2949         u64 off = start;
2950         u64 max = start + len;
2951         u32 flags = 0;
2952         u64 disko = 0;
2953         struct extent_map *em = NULL;
2954         int end = 0;
2955         u64 em_start = 0, em_len = 0;
2956         unsigned long emflags;
2957         ret = 0;
2958
2959         if (len == 0)
2960                 return -EINVAL;
2961
2962         lock_extent(&BTRFS_I(inode)->io_tree, start, start + len,
2963                 GFP_NOFS);
2964         em = get_extent(inode, NULL, 0, off, max - off, 0);
2965         if (!em)
2966                 goto out;
2967         if (IS_ERR(em)) {
2968                 ret = PTR_ERR(em);
2969                 goto out;
2970         }
2971         while (!end) {
2972                 off = em->start + em->len;
2973                 if (off >= max)
2974                         end = 1;
2975
2976                 em_start = em->start;
2977                 em_len = em->len;
2978
2979                 disko = 0;
2980                 flags = 0;
2981
2982                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
2983                         end = 1;
2984                         flags |= FIEMAP_EXTENT_LAST;
2985                 } else if (em->block_start == EXTENT_MAP_HOLE) {
2986                         flags |= FIEMAP_EXTENT_UNWRITTEN;
2987                 } else if (em->block_start == EXTENT_MAP_INLINE) {
2988                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
2989                                   FIEMAP_EXTENT_NOT_ALIGNED);
2990                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
2991                         flags |= (FIEMAP_EXTENT_DELALLOC |
2992                                   FIEMAP_EXTENT_UNKNOWN);
2993                 } else {
2994                         disko = em->block_start;
2995                 }
2996                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2997                         flags |= FIEMAP_EXTENT_ENCODED;
2998
2999                 emflags = em->flags;
3000                 free_extent_map(em);
3001                 em = NULL;
3002
3003                 if (!end) {
3004                         em = get_extent(inode, NULL, 0, off, max - off, 0);
3005                         if (!em)
3006                                 goto out;
3007                         if (IS_ERR(em)) {
3008                                 ret = PTR_ERR(em);
3009                                 goto out;
3010                         }
3011                         emflags = em->flags;
3012                 }
3013                 if (test_bit(EXTENT_FLAG_VACANCY, &emflags)) {
3014                         flags |= FIEMAP_EXTENT_LAST;
3015                         end = 1;
3016                 }
3017
3018                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3019                                         em_len, flags);
3020                 if (ret)
3021                         goto out_free;
3022         }
3023 out_free:
3024         free_extent_map(em);
3025 out:
3026         unlock_extent(&BTRFS_I(inode)->io_tree, start, start + len,
3027                         GFP_NOFS);
3028         return ret;
3029 }
3030
3031 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
3032                                               unsigned long i)
3033 {
3034         struct page *p;
3035         struct address_space *mapping;
3036
3037         if (i == 0)
3038                 return eb->first_page;
3039         i += eb->start >> PAGE_CACHE_SHIFT;
3040         mapping = eb->first_page->mapping;
3041         if (!mapping)
3042                 return NULL;
3043
3044         /*
3045          * extent_buffer_page is only called after pinning the page
3046          * by increasing the reference count.  So we know the page must
3047          * be in the radix tree.
3048          */
3049         rcu_read_lock();
3050         p = radix_tree_lookup(&mapping->page_tree, i);
3051         rcu_read_unlock();
3052
3053         return p;
3054 }
3055
3056 static inline unsigned long num_extent_pages(u64 start, u64 len)
3057 {
3058         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3059                 (start >> PAGE_CACHE_SHIFT);
3060 }
3061
3062 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3063                                                    u64 start,
3064                                                    unsigned long len,
3065                                                    gfp_t mask)
3066 {
3067         struct extent_buffer *eb = NULL;
3068 #if LEAK_DEBUG
3069         unsigned long flags;
3070 #endif
3071
3072         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3073         eb->start = start;
3074         eb->len = len;
3075         spin_lock_init(&eb->lock);
3076         init_waitqueue_head(&eb->lock_wq);
3077
3078 #if LEAK_DEBUG
3079         spin_lock_irqsave(&leak_lock, flags);
3080         list_add(&eb->leak_list, &buffers);
3081         spin_unlock_irqrestore(&leak_lock, flags);
3082 #endif
3083         atomic_set(&eb->refs, 1);
3084
3085         return eb;
3086 }
3087
3088 static void __free_extent_buffer(struct extent_buffer *eb)
3089 {
3090 #if LEAK_DEBUG
3091         unsigned long flags;
3092         spin_lock_irqsave(&leak_lock, flags);
3093         list_del(&eb->leak_list);
3094         spin_unlock_irqrestore(&leak_lock, flags);
3095 #endif
3096         kmem_cache_free(extent_buffer_cache, eb);
3097 }
3098
3099 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3100                                           u64 start, unsigned long len,
3101                                           struct page *page0,
3102                                           gfp_t mask)
3103 {
3104         unsigned long num_pages = num_extent_pages(start, len);
3105         unsigned long i;
3106         unsigned long index = start >> PAGE_CACHE_SHIFT;
3107         struct extent_buffer *eb;
3108         struct extent_buffer *exists = NULL;
3109         struct page *p;
3110         struct address_space *mapping = tree->mapping;
3111         int uptodate = 1;
3112
3113         spin_lock(&tree->buffer_lock);
3114         eb = buffer_search(tree, start);
3115         if (eb) {
3116                 atomic_inc(&eb->refs);
3117                 spin_unlock(&tree->buffer_lock);
3118                 mark_page_accessed(eb->first_page);
3119                 return eb;
3120         }
3121         spin_unlock(&tree->buffer_lock);
3122
3123         eb = __alloc_extent_buffer(tree, start, len, mask);
3124         if (!eb)
3125                 return NULL;
3126
3127         if (page0) {
3128                 eb->first_page = page0;
3129                 i = 1;
3130                 index++;
3131                 page_cache_get(page0);
3132                 mark_page_accessed(page0);
3133                 set_page_extent_mapped(page0);
3134                 set_page_extent_head(page0, len);
3135                 uptodate = PageUptodate(page0);
3136         } else {
3137                 i = 0;
3138         }
3139         for (; i < num_pages; i++, index++) {
3140                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
3141                 if (!p) {
3142                         WARN_ON(1);
3143                         goto free_eb;
3144                 }
3145                 set_page_extent_mapped(p);
3146                 mark_page_accessed(p);
3147                 if (i == 0) {
3148                         eb->first_page = p;
3149                         set_page_extent_head(p, len);
3150                 } else {
3151                         set_page_private(p, EXTENT_PAGE_PRIVATE);
3152                 }
3153                 if (!PageUptodate(p))
3154                         uptodate = 0;
3155                 unlock_page(p);
3156         }
3157         if (uptodate)
3158                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3159
3160         spin_lock(&tree->buffer_lock);
3161         exists = buffer_tree_insert(tree, start, &eb->rb_node);
3162         if (exists) {
3163                 /* add one reference for the caller */
3164                 atomic_inc(&exists->refs);
3165                 spin_unlock(&tree->buffer_lock);
3166                 goto free_eb;
3167         }
3168         /* add one reference for the tree */
3169         atomic_inc(&eb->refs);
3170         spin_unlock(&tree->buffer_lock);
3171         return eb;
3172
3173 free_eb:
3174         if (!atomic_dec_and_test(&eb->refs))
3175                 return exists;
3176         for (index = 1; index < i; index++)
3177                 page_cache_release(extent_buffer_page(eb, index));
3178         page_cache_release(extent_buffer_page(eb, 0));
3179         __free_extent_buffer(eb);
3180         return exists;
3181 }
3182
3183 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3184                                          u64 start, unsigned long len,
3185                                           gfp_t mask)
3186 {
3187         struct extent_buffer *eb;
3188
3189         spin_lock(&tree->buffer_lock);
3190         eb = buffer_search(tree, start);
3191         if (eb)
3192                 atomic_inc(&eb->refs);
3193         spin_unlock(&tree->buffer_lock);
3194
3195         if (eb)
3196                 mark_page_accessed(eb->first_page);
3197
3198         return eb;
3199 }
3200
3201 void free_extent_buffer(struct extent_buffer *eb)
3202 {
3203         if (!eb)
3204                 return;
3205
3206         if (!atomic_dec_and_test(&eb->refs))
3207                 return;
3208
3209         WARN_ON(1);
3210 }
3211
3212 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3213                               struct extent_buffer *eb)
3214 {
3215         unsigned long i;
3216         unsigned long num_pages;
3217         struct page *page;
3218
3219         num_pages = num_extent_pages(eb->start, eb->len);
3220
3221         for (i = 0; i < num_pages; i++) {
3222                 page = extent_buffer_page(eb, i);
3223                 if (!PageDirty(page))
3224                         continue;
3225
3226                 lock_page(page);
3227                 if (i == 0)
3228                         set_page_extent_head(page, eb->len);
3229                 else
3230                         set_page_private(page, EXTENT_PAGE_PRIVATE);
3231
3232                 clear_page_dirty_for_io(page);
3233                 spin_lock_irq(&page->mapping->tree_lock);
3234                 if (!PageDirty(page)) {
3235                         radix_tree_tag_clear(&page->mapping->page_tree,
3236                                                 page_index(page),
3237                                                 PAGECACHE_TAG_DIRTY);
3238                 }
3239                 spin_unlock_irq(&page->mapping->tree_lock);
3240                 unlock_page(page);
3241         }
3242         return 0;
3243 }
3244
3245 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
3246                                     struct extent_buffer *eb)
3247 {
3248         return wait_on_extent_writeback(tree, eb->start,
3249                                         eb->start + eb->len - 1);
3250 }
3251
3252 int set_extent_buffer_dirty(struct extent_io_tree *tree,
3253                              struct extent_buffer *eb)
3254 {
3255         unsigned long i;
3256         unsigned long num_pages;
3257         int was_dirty = 0;
3258
3259         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3260         num_pages = num_extent_pages(eb->start, eb->len);
3261         for (i = 0; i < num_pages; i++)
3262                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3263         return was_dirty;
3264 }
3265
3266 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3267                                 struct extent_buffer *eb)
3268 {
3269         unsigned long i;
3270         struct page *page;
3271         unsigned long num_pages;
3272
3273         num_pages = num_extent_pages(eb->start, eb->len);
3274         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3275
3276         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3277                               GFP_NOFS);
3278         for (i = 0; i < num_pages; i++) {
3279                 page = extent_buffer_page(eb, i);
3280                 if (page)
3281                         ClearPageUptodate(page);
3282         }
3283         return 0;
3284 }
3285
3286 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3287                                 struct extent_buffer *eb)
3288 {
3289         unsigned long i;
3290         struct page *page;
3291         unsigned long num_pages;
3292
3293         num_pages = num_extent_pages(eb->start, eb->len);
3294
3295         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3296                             GFP_NOFS);
3297         for (i = 0; i < num_pages; i++) {
3298                 page = extent_buffer_page(eb, i);
3299                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3300                     ((i == num_pages - 1) &&
3301                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3302                         check_page_uptodate(tree, page);
3303                         continue;
3304                 }
3305                 SetPageUptodate(page);
3306         }
3307         return 0;
3308 }
3309
3310 int extent_range_uptodate(struct extent_io_tree *tree,
3311                           u64 start, u64 end)
3312 {
3313         struct page *page;
3314         int ret;
3315         int pg_uptodate = 1;
3316         int uptodate;
3317         unsigned long index;
3318
3319         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL);
3320         if (ret)
3321                 return 1;
3322         while (start <= end) {
3323                 index = start >> PAGE_CACHE_SHIFT;
3324                 page = find_get_page(tree->mapping, index);
3325                 uptodate = PageUptodate(page);
3326                 page_cache_release(page);
3327                 if (!uptodate) {
3328                         pg_uptodate = 0;
3329                         break;
3330                 }
3331                 start += PAGE_CACHE_SIZE;
3332         }
3333         return pg_uptodate;
3334 }
3335
3336 int extent_buffer_uptodate(struct extent_io_tree *tree,
3337                            struct extent_buffer *eb)
3338 {
3339         int ret = 0;
3340         unsigned long num_pages;
3341         unsigned long i;
3342         struct page *page;
3343         int pg_uptodate = 1;
3344
3345         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3346                 return 1;
3347
3348         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3349                            EXTENT_UPTODATE, 1, NULL);
3350         if (ret)
3351                 return ret;
3352
3353         num_pages = num_extent_pages(eb->start, eb->len);
3354         for (i = 0; i < num_pages; i++) {
3355                 page = extent_buffer_page(eb, i);
3356                 if (!PageUptodate(page)) {
3357                         pg_uptodate = 0;
3358                         break;
3359                 }
3360         }
3361         return pg_uptodate;
3362 }
3363
3364 int read_extent_buffer_pages(struct extent_io_tree *tree,
3365                              struct extent_buffer *eb,
3366                              u64 start, int wait,
3367                              get_extent_t *get_extent, int mirror_num)
3368 {
3369         unsigned long i;
3370         unsigned long start_i;
3371         struct page *page;
3372         int err;
3373         int ret = 0;
3374         int locked_pages = 0;
3375         int all_uptodate = 1;
3376         int inc_all_pages = 0;
3377         unsigned long num_pages;
3378         struct bio *bio = NULL;
3379         unsigned long bio_flags = 0;
3380
3381         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3382                 return 0;
3383
3384         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3385                            EXTENT_UPTODATE, 1, NULL)) {
3386                 return 0;
3387         }
3388
3389         if (start) {
3390                 WARN_ON(start < eb->start);
3391                 start_i = (start >> PAGE_CACHE_SHIFT) -
3392                         (eb->start >> PAGE_CACHE_SHIFT);
3393         } else {
3394                 start_i = 0;
3395         }
3396
3397         num_pages = num_extent_pages(eb->start, eb->len);
3398         for (i = start_i; i < num_pages; i++) {
3399                 page = extent_buffer_page(eb, i);
3400                 if (!wait) {
3401                         if (!trylock_page(page))
3402                                 goto unlock_exit;
3403                 } else {
3404                         lock_page(page);
3405                 }
3406                 locked_pages++;
3407                 if (!PageUptodate(page))
3408                         all_uptodate = 0;
3409         }
3410         if (all_uptodate) {
3411                 if (start_i == 0)
3412                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3413                 goto unlock_exit;
3414         }
3415
3416         for (i = start_i; i < num_pages; i++) {
3417                 page = extent_buffer_page(eb, i);
3418                 if (inc_all_pages)
3419                         page_cache_get(page);
3420                 if (!PageUptodate(page)) {
3421                         if (start_i == 0)
3422                                 inc_all_pages = 1;
3423                         ClearPageError(page);
3424                         err = __extent_read_full_page(tree, page,
3425                                                       get_extent, &bio,
3426                                                       mirror_num, &bio_flags);
3427                         if (err)
3428                                 ret = err;
3429                 } else {
3430                         unlock_page(page);
3431                 }
3432         }
3433
3434         if (bio)
3435                 submit_one_bio(READ, bio, mirror_num, bio_flags);
3436
3437         if (ret || !wait)
3438                 return ret;
3439
3440         for (i = start_i; i < num_pages; i++) {
3441                 page = extent_buffer_page(eb, i);
3442                 wait_on_page_locked(page);
3443                 if (!PageUptodate(page))
3444                         ret = -EIO;
3445         }
3446
3447         if (!ret)
3448                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3449         return ret;
3450
3451 unlock_exit:
3452         i = start_i;
3453         while (locked_pages > 0) {
3454                 page = extent_buffer_page(eb, i);
3455                 i++;
3456                 unlock_page(page);
3457                 locked_pages--;
3458         }
3459         return ret;
3460 }
3461
3462 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3463                         unsigned long start,
3464                         unsigned long len)
3465 {
3466         size_t cur;
3467         size_t offset;
3468         struct page *page;
3469         char *kaddr;
3470         char *dst = (char *)dstv;
3471         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3472         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3473
3474         WARN_ON(start > eb->len);
3475         WARN_ON(start + len > eb->start + eb->len);
3476
3477         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3478
3479         while (len > 0) {
3480                 page = extent_buffer_page(eb, i);
3481
3482                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3483                 kaddr = kmap_atomic(page, KM_USER1);
3484                 memcpy(dst, kaddr + offset, cur);
3485                 kunmap_atomic(kaddr, KM_USER1);
3486
3487                 dst += cur;
3488                 len -= cur;
3489                 offset = 0;
3490                 i++;
3491         }
3492 }
3493
3494 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3495                                unsigned long min_len, char **token, char **map,
3496                                unsigned long *map_start,
3497                                unsigned long *map_len, int km)
3498 {
3499         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3500         char *kaddr;
3501         struct page *p;
3502         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3503         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3504         unsigned long end_i = (start_offset + start + min_len - 1) >>
3505                 PAGE_CACHE_SHIFT;
3506
3507         if (i != end_i)
3508                 return -EINVAL;
3509
3510         if (i == 0) {
3511                 offset = start_offset;
3512                 *map_start = 0;
3513         } else {
3514                 offset = 0;
3515                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3516         }
3517
3518         if (start + min_len > eb->len) {
3519                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3520                        "wanted %lu %lu\n", (unsigned long long)eb->start,
3521                        eb->len, start, min_len);
3522                 WARN_ON(1);
3523         }
3524
3525         p = extent_buffer_page(eb, i);
3526         kaddr = kmap_atomic(p, km);
3527         *token = kaddr;
3528         *map = kaddr + offset;
3529         *map_len = PAGE_CACHE_SIZE - offset;
3530         return 0;
3531 }
3532
3533 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3534                       unsigned long min_len,
3535                       char **token, char **map,
3536                       unsigned long *map_start,
3537                       unsigned long *map_len, int km)
3538 {
3539         int err;
3540         int save = 0;
3541         if (eb->map_token) {
3542                 unmap_extent_buffer(eb, eb->map_token, km);
3543                 eb->map_token = NULL;
3544                 save = 1;
3545         }
3546         err = map_private_extent_buffer(eb, start, min_len, token, map,
3547                                        map_start, map_len, km);
3548         if (!err && save) {
3549                 eb->map_token = *token;
3550                 eb->kaddr = *map;
3551                 eb->map_start = *map_start;
3552                 eb->map_len = *map_len;
3553         }
3554         return err;
3555 }
3556
3557 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3558 {
3559         kunmap_atomic(token, km);
3560 }
3561
3562 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3563                           unsigned long start,
3564                           unsigned long len)
3565 {
3566         size_t cur;
3567         size_t offset;
3568         struct page *page;
3569         char *kaddr;
3570         char *ptr = (char *)ptrv;
3571         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3572         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3573         int ret = 0;
3574
3575         WARN_ON(start > eb->len);
3576         WARN_ON(start + len > eb->start + eb->len);
3577
3578         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3579
3580         while (len > 0) {
3581                 page = extent_buffer_page(eb, i);
3582
3583                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3584
3585                 kaddr = kmap_atomic(page, KM_USER0);
3586                 ret = memcmp(ptr, kaddr + offset, cur);
3587                 kunmap_atomic(kaddr, KM_USER0);
3588                 if (ret)
3589                         break;
3590
3591                 ptr += cur;
3592                 len -= cur;
3593                 offset = 0;
3594                 i++;
3595         }
3596         return ret;
3597 }
3598
3599 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3600                          unsigned long start, unsigned long len)
3601 {
3602         size_t cur;
3603         size_t offset;
3604         struct page *page;
3605         char *kaddr;
3606         char *src = (char *)srcv;
3607         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3608         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3609
3610         WARN_ON(start > eb->len);
3611         WARN_ON(start + len > eb->start + eb->len);
3612
3613         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3614
3615         while (len > 0) {
3616                 page = extent_buffer_page(eb, i);
3617                 WARN_ON(!PageUptodate(page));
3618
3619                 cur = min(len, PAGE_CACHE_SIZE - offset);
3620                 kaddr = kmap_atomic(page, KM_USER1);
3621                 memcpy(kaddr + offset, src, cur);
3622                 kunmap_atomic(kaddr, KM_USER1);
3623
3624                 src += cur;
3625                 len -= cur;
3626                 offset = 0;
3627                 i++;
3628         }
3629 }
3630
3631 void memset_extent_buffer(struct extent_buffer *eb, char c,
3632                           unsigned long start, unsigned long len)
3633 {
3634         size_t cur;
3635         size_t offset;
3636         struct page *page;
3637         char *kaddr;
3638         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3639         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3640
3641         WARN_ON(start > eb->len);
3642         WARN_ON(start + len > eb->start + eb->len);
3643
3644         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3645
3646         while (len > 0) {
3647                 page = extent_buffer_page(eb, i);
3648                 WARN_ON(!PageUptodate(page));
3649
3650                 cur = min(len, PAGE_CACHE_SIZE - offset);
3651                 kaddr = kmap_atomic(page, KM_USER0);
3652                 memset(kaddr + offset, c, cur);
3653                 kunmap_atomic(kaddr, KM_USER0);
3654
3655                 len -= cur;
3656                 offset = 0;
3657                 i++;
3658         }
3659 }
3660
3661 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3662                         unsigned long dst_offset, unsigned long src_offset,
3663                         unsigned long len)
3664 {
3665         u64 dst_len = dst->len;
3666         size_t cur;
3667         size_t offset;
3668         struct page *page;
3669         char *kaddr;
3670         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3671         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3672
3673         WARN_ON(src->len != dst_len);
3674
3675         offset = (start_offset + dst_offset) &
3676                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3677
3678         while (len > 0) {
3679                 page = extent_buffer_page(dst, i);
3680                 WARN_ON(!PageUptodate(page));
3681
3682                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3683
3684                 kaddr = kmap_atomic(page, KM_USER0);
3685                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3686                 kunmap_atomic(kaddr, KM_USER0);
3687
3688                 src_offset += cur;
3689                 len -= cur;
3690                 offset = 0;
3691                 i++;
3692         }
3693 }
3694
3695 static void move_pages(struct page *dst_page, struct page *src_page,
3696                        unsigned long dst_off, unsigned long src_off,
3697                        unsigned long len)
3698 {
3699         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3700         if (dst_page == src_page) {
3701                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3702         } else {
3703                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3704                 char *p = dst_kaddr + dst_off + len;
3705                 char *s = src_kaddr + src_off + len;
3706
3707                 while (len--)
3708                         *--p = *--s;
3709
3710                 kunmap_atomic(src_kaddr, KM_USER1);
3711         }
3712         kunmap_atomic(dst_kaddr, KM_USER0);
3713 }
3714
3715 static void copy_pages(struct page *dst_page, struct page *src_page,
3716                        unsigned long dst_off, unsigned long src_off,
3717                        unsigned long len)
3718 {
3719         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3720         char *src_kaddr;
3721
3722         if (dst_page != src_page)
3723                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3724         else
3725                 src_kaddr = dst_kaddr;
3726
3727         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3728         kunmap_atomic(dst_kaddr, KM_USER0);
3729         if (dst_page != src_page)
3730                 kunmap_atomic(src_kaddr, KM_USER1);
3731 }
3732
3733 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3734                            unsigned long src_offset, unsigned long len)
3735 {
3736         size_t cur;
3737         size_t dst_off_in_page;
3738         size_t src_off_in_page;
3739         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3740         unsigned long dst_i;
3741         unsigned long src_i;
3742
3743         if (src_offset + len > dst->len) {
3744                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3745                        "len %lu dst len %lu\n", src_offset, len, dst->len);
3746                 BUG_ON(1);
3747         }
3748         if (dst_offset + len > dst->len) {
3749                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3750                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
3751                 BUG_ON(1);
3752         }
3753
3754         while (len > 0) {
3755                 dst_off_in_page = (start_offset + dst_offset) &
3756                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3757                 src_off_in_page = (start_offset + src_offset) &
3758                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3759
3760                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3761                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3762
3763                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3764                                                src_off_in_page));
3765                 cur = min_t(unsigned long, cur,
3766                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3767
3768                 copy_pages(extent_buffer_page(dst, dst_i),
3769                            extent_buffer_page(dst, src_i),
3770                            dst_off_in_page, src_off_in_page, cur);
3771
3772                 src_offset += cur;
3773                 dst_offset += cur;
3774                 len -= cur;
3775         }
3776 }
3777
3778 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3779                            unsigned long src_offset, unsigned long len)
3780 {
3781         size_t cur;
3782         size_t dst_off_in_page;
3783         size_t src_off_in_page;
3784         unsigned long dst_end = dst_offset + len - 1;
3785         unsigned long src_end = src_offset + len - 1;
3786         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3787         unsigned long dst_i;
3788         unsigned long src_i;
3789
3790         if (src_offset + len > dst->len) {
3791                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3792                        "len %lu len %lu\n", src_offset, len, dst->len);
3793                 BUG_ON(1);
3794         }
3795         if (dst_offset + len > dst->len) {
3796                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3797                        "len %lu len %lu\n", dst_offset, len, dst->len);
3798                 BUG_ON(1);
3799         }
3800         if (dst_offset < src_offset) {
3801                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3802                 return;
3803         }
3804         while (len > 0) {
3805                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3806                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3807
3808                 dst_off_in_page = (start_offset + dst_end) &
3809                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3810                 src_off_in_page = (start_offset + src_end) &
3811                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3812
3813                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3814                 cur = min(cur, dst_off_in_page + 1);
3815                 move_pages(extent_buffer_page(dst, dst_i),
3816                            extent_buffer_page(dst, src_i),
3817                            dst_off_in_page - cur + 1,
3818                            src_off_in_page - cur + 1, cur);
3819
3820                 dst_end -= cur;
3821                 src_end -= cur;
3822                 len -= cur;
3823         }
3824 }
3825
3826 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3827 {
3828         u64 start = page_offset(page);
3829         struct extent_buffer *eb;
3830         int ret = 1;
3831         unsigned long i;
3832         unsigned long num_pages;
3833
3834         spin_lock(&tree->buffer_lock);
3835         eb = buffer_search(tree, start);
3836         if (!eb)
3837                 goto out;
3838
3839         if (atomic_read(&eb->refs) > 1) {
3840                 ret = 0;
3841                 goto out;
3842         }
3843         if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3844                 ret = 0;
3845                 goto out;
3846         }
3847         /* at this point we can safely release the extent buffer */
3848         num_pages = num_extent_pages(eb->start, eb->len);
3849         for (i = 0; i < num_pages; i++)
3850                 page_cache_release(extent_buffer_page(eb, i));
3851         rb_erase(&eb->rb_node, &tree->buffer);
3852         __free_extent_buffer(eb);
3853 out:
3854         spin_unlock(&tree->buffer_lock);
3855         return ret;
3856 }