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