fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/module.h>
   8 #include <linux/spinlock.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/swap.h>
  11 #include <linux/writeback.h>
  12 #include <linux/pagevec.h>
  13 #include "extent_io.h"
  14 #include "extent_map.h"
  15 #include "compat.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18
  19 static struct kmem_cache *extent_state_cache;
  20 static struct kmem_cache *extent_buffer_cache;
  21
  22 static LIST_HEAD(buffers);
  23 static LIST_HEAD(states);
  24
  25 #define LEAK_DEBUG 0
  26 #if LEAK_DEBUG
  27 static DEFINE_SPINLOCK(leak_lock);
  28 #endif
  29
  30 #define BUFFER_LRU_MAX 64
  31
  32 struct tree_entry {
  33         u64 start;
  34         u64 end;
  35         struct rb_node rb_node;
  36 };
  37
  38 struct extent_page_data {
  39         struct bio *bio;
  40         struct extent_io_tree *tree;
  41         get_extent_t *get_extent;
  42
  43         /* tells writepage not to lock the state bits for this range
  44          * it still does the unlocking
  45          */
  46         unsigned int extent_locked:1;
  47
  48         /* tells the submit_bio code to use a WRITE_SYNC */
  49         unsigned int sync_io:1;
  50 };
  51
  52 int __init extent_io_init(void)
  53 {
  54         extent_state_cache = kmem_cache_create("extent_state",
  55                         sizeof(struct extent_state), 0,
  56                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  57         if (!extent_state_cache)
  58                 return -ENOMEM;
  59
  60         extent_buffer_cache = kmem_cache_create("extent_buffers",
  61                         sizeof(struct extent_buffer), 0,
  62                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  63         if (!extent_buffer_cache)
  64                 goto free_state_cache;
  65         return 0;
  66
  67 free_state_cache:
  68         kmem_cache_destroy(extent_state_cache);
  69         return -ENOMEM;
  70 }
  71
  72 void extent_io_exit(void)
  73 {
  74         struct extent_state *state;
  75         struct extent_buffer *eb;
  76
  77         while (!list_empty(&states)) {
  78                 state = list_entry(states.next, struct extent_state, leak_list);
  79                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  80                        "state %lu in tree %p refs %d\n",
  81                        (unsigned long long)state->start,
  82                        (unsigned long long)state->end,
  83                        state->state, state->tree, atomic_read(&state->refs));
  84                 list_del(&state->leak_list);
  85                 kmem_cache_free(extent_state_cache, state);
  86
  87         }
  88
  89         while (!list_empty(&buffers)) {
  90                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  91                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  92                        "refs %d\n", (unsigned long long)eb->start,
  93                        eb->len, atomic_read(&eb->refs));
  94                 list_del(&eb->leak_list);
  95                 kmem_cache_free(extent_buffer_cache, eb);
  96         }
  97         if (extent_state_cache)
  98                 kmem_cache_destroy(extent_state_cache);
  99         if (extent_buffer_cache)
 100                 kmem_cache_destroy(extent_buffer_cache);
 101 }
 102
 103 void extent_io_tree_init(struct extent_io_tree *tree,
 104                           struct address_space *mapping, gfp_t mask)
 105 {
 106         tree->state = RB_ROOT;
 107         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 108         tree->ops = NULL;
 109         tree->dirty_bytes = 0;
 110         spin_lock_init(&tree->lock);
 111         spin_lock_init(&tree->buffer_lock);
 112         tree->mapping = mapping;
 113 }
 114
 115 static struct extent_state *alloc_extent_state(gfp_t mask)
 116 {
 117         struct extent_state *state;
 118 #if LEAK_DEBUG
 119         unsigned long flags;
 120 #endif
 121
 122         state = kmem_cache_alloc(extent_state_cache, mask);
 123         if (!state)
 124                 return state;
 125         state->state = 0;
 126         state->private = 0;
 127         state->tree = NULL;
 128 #if LEAK_DEBUG
 129         spin_lock_irqsave(&leak_lock, flags);
 130         list_add(&state->leak_list, &states);
 131         spin_unlock_irqrestore(&leak_lock, flags);
 132 #endif
 133         atomic_set(&state->refs, 1);
 134         init_waitqueue_head(&state->wq);
 135         return state;
 136 }
 137
 138 void free_extent_state(struct extent_state *state)
 139 {
 140         if (!state)
 141                 return;
 142         if (atomic_dec_and_test(&state->refs)) {
 143 #if LEAK_DEBUG
 144                 unsigned long flags;
 145 #endif
 146                 WARN_ON(state->tree);
 147 #if LEAK_DEBUG
 148                 spin_lock_irqsave(&leak_lock, flags);
 149                 list_del(&state->leak_list);
 150                 spin_unlock_irqrestore(&leak_lock, flags);
 151 #endif
 152                 kmem_cache_free(extent_state_cache, state);
 153         }
 154 }
 155
 156 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 157                                    struct rb_node *node)
 158 {
 159         struct rb_node **p = &root->rb_node;
 160         struct rb_node *parent = NULL;
 161         struct tree_entry *entry;
 162
 163         while (*p) {
 164                 parent = *p;
 165                 entry = rb_entry(parent, struct tree_entry, rb_node);
 166
 167                 if (offset < entry->start)
 168                         p = &(*p)->rb_left;
 169                 else if (offset > entry->end)
 170                         p = &(*p)->rb_right;
 171                 else
 172                         return parent;
 173         }
 174
 175         entry = rb_entry(node, struct tree_entry, rb_node);
 176         rb_link_node(node, parent, p);
 177         rb_insert_color(node, root);
 178         return NULL;
 179 }
 180
 181 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 182                                      struct rb_node **prev_ret,
 183                                      struct rb_node **next_ret)
 184 {
 185         struct rb_root *root = &tree->state;
 186         struct rb_node *n = root->rb_node;
 187         struct rb_node *prev = NULL;
 188         struct rb_node *orig_prev = NULL;
 189         struct tree_entry *entry;
 190         struct tree_entry *prev_entry = NULL;
 191
 192         while (n) {
 193                 entry = rb_entry(n, struct tree_entry, rb_node);
 194                 prev = n;
 195                 prev_entry = entry;
 196
 197                 if (offset < entry->start)
 198                         n = n->rb_left;
 199                 else if (offset > entry->end)
 200                         n = n->rb_right;
 201                 else
 202                         return n;
 203         }
 204
 205         if (prev_ret) {
 206                 orig_prev = prev;
 207                 while (prev && offset > prev_entry->end) {
 208                         prev = rb_next(prev);
 209                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 210                 }
 211                 *prev_ret = prev;
 212                 prev = orig_prev;
 213         }
 214
 215         if (next_ret) {
 216                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 217                 while (prev && offset < prev_entry->start) {
 218                         prev = rb_prev(prev);
 219                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 220                 }
 221                 *next_ret = prev;
 222         }
 223         return NULL;
 224 }
 225
 226 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 227                                           u64 offset)
 228 {
 229         struct rb_node *prev = NULL;
 230         struct rb_node *ret;
 231
 232         ret = __etree_search(tree, offset, &prev, NULL);
 233         if (!ret)
 234                 return prev;
 235         return ret;
 236 }
 237
 238 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 239                      struct extent_state *other)
 240 {
 241         if (tree->ops && tree->ops->merge_extent_hook)
 242                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 243                                              other);
 244 }
 245
 246 /*
 247  * utility function to look for merge candidates inside a given range.
 248  * Any extents with matching state are merged together into a single
 249  * extent in the tree.  Extents with EXTENT_IO in their state field
 250  * are not merged because the end_io handlers need to be able to do
 251  * operations on them without sleeping (or doing allocations/splits).
 252  *
 253  * This should be called with the tree lock held.
 254  */
 255 static int merge_state(struct extent_io_tree *tree,
 256                        struct extent_state *state)
 257 {
 258         struct extent_state *other;
 259         struct rb_node *other_node;
 260
 261         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 262                 return 0;
 263
 264         other_node = rb_prev(&state->rb_node);
 265         if (other_node) {
 266                 other = rb_entry(other_node, struct extent_state, rb_node);
 267                 if (other->end == state->start - 1 &&
 268                     other->state == state->state) {
 269                         merge_cb(tree, state, other);
 270                         state->start = other->start;
 271                         other->tree = NULL;
 272                         rb_erase(&other->rb_node, &tree->state);
 273                         free_extent_state(other);
 274                 }
 275         }
 276         other_node = rb_next(&state->rb_node);
 277         if (other_node) {
 278                 other = rb_entry(other_node, struct extent_state, rb_node);
 279                 if (other->start == state->end + 1 &&
 280                     other->state == state->state) {
 281                         merge_cb(tree, state, other);
 282                         other->start = state->start;
 283                         state->tree = NULL;
 284                         rb_erase(&state->rb_node, &tree->state);
 285                         free_extent_state(state);
 286                         state = NULL;
 287                 }
 288         }
 289
 290         return 0;
 291 }
 292
 293 static int set_state_cb(struct extent_io_tree *tree,
 294                          struct extent_state *state, int *bits)
 295 {
 296         if (tree->ops && tree->ops->set_bit_hook) {
 297                 return tree->ops->set_bit_hook(tree->mapping->host,
 298                                                state, bits);
 299         }
 300
 301         return 0;
 302 }
 303
 304 static void clear_state_cb(struct extent_io_tree *tree,
 305                            struct extent_state *state, int *bits)
 306 {
 307         if (tree->ops && tree->ops->clear_bit_hook)
 308                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 309 }
 310
 311 /*
 312  * insert an extent_state struct into the tree.  'bits' are set on the
 313  * struct before it is inserted.
 314  *
 315  * This may return -EEXIST if the extent is already there, in which case the
 316  * state struct is freed.
 317  *
 318  * The tree lock is not taken internally.  This is a utility function and
 319  * probably isn't what you want to call (see set/clear_extent_bit).
 320  */
 321 static int insert_state(struct extent_io_tree *tree,
 322                         struct extent_state *state, u64 start, u64 end,
 323                         int *bits)
 324 {
 325         struct rb_node *node;
 326         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 327         int ret;
 328
 329         if (end < start) {
 330                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
 331                        (unsigned long long)end,
 332                        (unsigned long long)start);
 333                 WARN_ON(1);
 334         }
 335         state->start = start;
 336         state->end = end;
 337         ret = set_state_cb(tree, state, bits);
 338         if (ret)
 339                 return ret;
 340
 341         if (bits_to_set & EXTENT_DIRTY)
 342                 tree->dirty_bytes += end - start + 1;
 343         state->state |= bits_to_set;
 344         node = tree_insert(&tree->state, end, &state->rb_node);
 345         if (node) {
 346                 struct extent_state *found;
 347                 found = rb_entry(node, struct extent_state, rb_node);
 348                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 349                        "%llu %llu\n", (unsigned long long)found->start,
 350                        (unsigned long long)found->end,
 351                        (unsigned long long)start, (unsigned long long)end);
 352                 free_extent_state(state);
 353                 return -EEXIST;
 354         }
 355         state->tree = tree;
 356         merge_state(tree, state);
 357         return 0;
 358 }
 359
 360 static int split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 361                      u64 split)
 362 {
 363         if (tree->ops && tree->ops->split_extent_hook)
 364                 return tree->ops->split_extent_hook(tree->mapping->host,
 365                                                     orig, split);
 366         return 0;
 367 }
 368
 369 /*
 370  * split a given extent state struct in two, inserting the preallocated
 371  * struct 'prealloc' as the newly created second half.  'split' indicates an
 372  * offset inside 'orig' where it should be split.
 373  *
 374  * Before calling,
 375  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 376  * are two extent state structs in the tree:
 377  * prealloc: [orig->start, split - 1]
 378  * orig: [ split, orig->end ]
 379  *
 380  * The tree locks are not taken by this function. They need to be held
 381  * by the caller.
 382  */
 383 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 384                        struct extent_state *prealloc, u64 split)
 385 {
 386         struct rb_node *node;
 387
 388         split_cb(tree, orig, split);
 389
 390         prealloc->start = orig->start;
 391         prealloc->end = split - 1;
 392         prealloc->state = orig->state;
 393         orig->start = split;
 394
 395         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 396         if (node) {
 397                 free_extent_state(prealloc);
 398                 return -EEXIST;
 399         }
 400         prealloc->tree = tree;
 401         return 0;
 402 }
 403
 404 /*
 405  * utility function to clear some bits in an extent state struct.
 406  * it will optionally wake up any one waiting on this state (wake == 1), or
 407  * forcibly remove the state from the tree (delete == 1).
 408  *
 409  * If no bits are set on the state struct after clearing things, the
 410  * struct is freed and removed from the tree
 411  */
 412 static int clear_state_bit(struct extent_io_tree *tree,
 413                             struct extent_state *state,
 414                             int *bits, int wake)
 415 {
 416         int bits_to_clear = *bits & ~EXTENT_CTLBITS;
 417         int ret = state->state & bits_to_clear;
 418
 419         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 420                 u64 range = state->end - state->start + 1;
 421                 WARN_ON(range > tree->dirty_bytes);
 422                 tree->dirty_bytes -= range;
 423         }
 424         clear_state_cb(tree, state, bits);
 425         state->state &= ~bits_to_clear;
 426         if (wake)
 427                 wake_up(&state->wq);
 428         if (state->state == 0) {
 429                 if (state->tree) {
 430                         rb_erase(&state->rb_node, &tree->state);
 431                         state->tree = NULL;
 432                         free_extent_state(state);
 433                 } else {
 434                         WARN_ON(1);
 435                 }
 436         } else {
 437                 merge_state(tree, state);
 438         }
 439         return ret;
 440 }
 441
 442 /*
 443  * clear some bits on a range in the tree.  This may require splitting
 444  * or inserting elements in the tree, so the gfp mask is used to
 445  * indicate which allocations or sleeping are allowed.
 446  *
 447  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 448  * the given range from the tree regardless of state (ie for truncate).
 449  *
 450  * the range [start, end] is inclusive.
 451  *
 452  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 453  * bits were already set, or zero if none of the bits were already set.
 454  */
 455 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 456                      int bits, int wake, int delete,
 457                      struct extent_state **cached_state,
 458                      gfp_t mask)
 459 {
 460         struct extent_state *state;
 461         struct extent_state *cached;
 462         struct extent_state *prealloc = NULL;
 463         struct rb_node *next_node;
 464         struct rb_node *node;
 465         u64 last_end;
 466         int err;
 467         int set = 0;
 468         int clear = 0;
 469
 470         if (delete)
 471                 bits |= ~EXTENT_CTLBITS;
 472         bits |= EXTENT_FIRST_DELALLOC;
 473
 474         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 475                 clear = 1;
 476 again:
 477         if (!prealloc && (mask & __GFP_WAIT)) {
 478                 prealloc = alloc_extent_state(mask);
 479                 if (!prealloc)
 480                         return -ENOMEM;
 481         }
 482
 483         spin_lock(&tree->lock);
 484         if (cached_state) {
 485                 cached = *cached_state;
 486
 487                 if (clear) {
 488                         *cached_state = NULL;
 489                         cached_state = NULL;
 490                 }
 491
 492                 if (cached && cached->tree && cached->start == start) {
 493                         if (clear)
 494                                 atomic_dec(&cached->refs);
 495                         state = cached;
 496                         goto hit_next;
 497                 }
 498                 if (clear)
 499                         free_extent_state(cached);
 500         }
 501         /*
 502          * this search will find the extents that end after
 503          * our range starts
 504          */
 505         node = tree_search(tree, start);
 506         if (!node)
 507                 goto out;
 508         state = rb_entry(node, struct extent_state, rb_node);
 509 hit_next:
 510         if (state->start > end)
 511                 goto out;
 512         WARN_ON(state->end < start);
 513         last_end = state->end;
 514
 515         /*
 516          *     | ---- desired range ---- |
 517          *  | state | or
 518          *  | ------------- state -------------- |
 519          *
 520          * We need to split the extent we found, and may flip
 521          * bits on second half.
 522          *
 523          * If the extent we found extends past our range, we
 524          * just split and search again.  It'll get split again
 525          * the next time though.
 526          *
 527          * If the extent we found is inside our range, we clear
 528          * the desired bit on it.
 529          */
 530
 531         if (state->start < start) {
 532                 if (!prealloc)
 533                         prealloc = alloc_extent_state(GFP_ATOMIC);
 534                 err = split_state(tree, state, prealloc, start);
 535                 BUG_ON(err == -EEXIST);
 536                 prealloc = NULL;
 537                 if (err)
 538                         goto out;
 539                 if (state->end <= end) {
 540                         set |= clear_state_bit(tree, state, &bits, wake);
 541                         if (last_end == (u64)-1)
 542                                 goto out;
 543                         start = last_end + 1;
 544                 }
 545                 goto search_again;
 546         }
 547         /*
 548          * | ---- desired range ---- |
 549          *                        | state |
 550          * We need to split the extent, and clear the bit
 551          * on the first half
 552          */
 553         if (state->start <= end && state->end > end) {
 554                 if (!prealloc)
 555                         prealloc = alloc_extent_state(GFP_ATOMIC);
 556                 err = split_state(tree, state, prealloc, end + 1);
 557                 BUG_ON(err == -EEXIST);
 558                 if (wake)
 559                         wake_up(&state->wq);
 560
 561                 set |= clear_state_bit(tree, prealloc, &bits, wake);
 562
 563                 prealloc = NULL;
 564                 goto out;
 565         }
 566
 567         if (state->end < end && prealloc && !need_resched())
 568                 next_node = rb_next(&state->rb_node);
 569         else
 570                 next_node = NULL;
 571
 572         set |= clear_state_bit(tree, state, &bits, wake);
 573         if (last_end == (u64)-1)
 574                 goto out;
 575         start = last_end + 1;
 576         if (start <= end && next_node) {
 577                 state = rb_entry(next_node, struct extent_state,
 578                                  rb_node);
 579                 if (state->start == start)
 580                         goto hit_next;
 581         }
 582         goto search_again;
 583
 584 out:
 585         spin_unlock(&tree->lock);
 586         if (prealloc)
 587                 free_extent_state(prealloc);
 588
 589         return set;
 590
 591 search_again:
 592         if (start > end)
 593                 goto out;
 594         spin_unlock(&tree->lock);
 595         if (mask & __GFP_WAIT)
 596                 cond_resched();
 597         goto again;
 598 }
 599
 600 static int wait_on_state(struct extent_io_tree *tree,
 601                          struct extent_state *state)
 602                 __releases(tree->lock)
 603                 __acquires(tree->lock)
 604 {
 605         DEFINE_WAIT(wait);
 606         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 607         spin_unlock(&tree->lock);
 608         schedule();
 609         spin_lock(&tree->lock);
 610         finish_wait(&state->wq, &wait);
 611         return 0;
 612 }
 613
 614 /*
 615  * waits for one or more bits to clear on a range in the state tree.
 616  * The range [start, end] is inclusive.
 617  * The tree lock is taken by this function
 618  */
 619 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 620 {
 621         struct extent_state *state;
 622         struct rb_node *node;
 623
 624         spin_lock(&tree->lock);
 625 again:
 626         while (1) {
 627                 /*
 628                  * this search will find all the extents that end after
 629                  * our range starts
 630                  */
 631                 node = tree_search(tree, start);
 632                 if (!node)
 633                         break;
 634
 635                 state = rb_entry(node, struct extent_state, rb_node);
 636
 637                 if (state->start > end)
 638                         goto out;
 639
 640                 if (state->state & bits) {
 641                         start = state->start;
 642                         atomic_inc(&state->refs);
 643                         wait_on_state(tree, state);
 644                         free_extent_state(state);
 645                         goto again;
 646                 }
 647                 start = state->end + 1;
 648
 649                 if (start > end)
 650                         break;
 651
 652                 if (need_resched()) {
 653                         spin_unlock(&tree->lock);
 654                         cond_resched();
 655                         spin_lock(&tree->lock);
 656                 }
 657         }
 658 out:
 659         spin_unlock(&tree->lock);
 660         return 0;
 661 }
 662
 663 static int set_state_bits(struct extent_io_tree *tree,
 664                            struct extent_state *state,
 665                            int *bits)
 666 {
 667         int ret;
 668         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 669
 670         ret = set_state_cb(tree, state, bits);
 671         if (ret)
 672                 return ret;
 673         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 674                 u64 range = state->end - state->start + 1;
 675                 tree->dirty_bytes += range;
 676         }
 677         state->state |= bits_to_set;
 678
 679         return 0;
 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 static void uncache_state(struct extent_state **cached_ptr)
 694 {
 695         if (cached_ptr && (*cached_ptr)) {
 696                 struct extent_state *state = *cached_ptr;
 697                 *cached_ptr = NULL;
 698                 free_extent_state(state);
 699         }
 700 }
 701
 702 /*
 703  * set some bits on a range in the tree.  This may require allocations or
 704  * sleeping, so the gfp mask is used to indicate what is allowed.
 705  *
 706  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 707  * part of the range already has the desired bits set.  The start of the
 708  * existing range is returned in failed_start in this case.
 709  *
 710  * [start, end] is inclusive This takes the tree lock.
 711  */
 712
 713 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 714                    int bits, int exclusive_bits, u64 *failed_start,
 715                    struct extent_state **cached_state, gfp_t mask)
 716 {
 717         struct extent_state *state;
 718         struct extent_state *prealloc = NULL;
 719         struct rb_node *node;
 720         int err = 0;
 721         u64 last_start;
 722         u64 last_end;
 723
 724         bits |= EXTENT_FIRST_DELALLOC;
 725 again:
 726         if (!prealloc && (mask & __GFP_WAIT)) {
 727                 prealloc = alloc_extent_state(mask);
 728                 if (!prealloc)
 729                         return -ENOMEM;
 730         }
 731
 732         spin_lock(&tree->lock);
 733         if (cached_state && *cached_state) {
 734                 state = *cached_state;
 735                 if (state->start == start && state->tree) {
 736                         node = &state->rb_node;
 737                         goto hit_next;
 738                 }
 739         }
 740         /*
 741          * this search will find all the extents that end after
 742          * our range starts.
 743          */
 744         node = tree_search(tree, start);
 745         if (!node) {
 746                 err = insert_state(tree, prealloc, start, end, &bits);
 747                 prealloc = NULL;
 748                 BUG_ON(err == -EEXIST);
 749                 goto out;
 750         }
 751         state = rb_entry(node, struct extent_state, rb_node);
 752 hit_next:
 753         last_start = state->start;
 754         last_end = state->end;
 755
 756         /*
 757          * | ---- desired range ---- |
 758          * | state |
 759          *
 760          * Just lock what we found and keep going
 761          */
 762         if (state->start == start && state->end <= end) {
 763                 struct rb_node *next_node;
 764                 if (state->state & exclusive_bits) {
 765                         *failed_start = state->start;
 766                         err = -EEXIST;
 767                         goto out;
 768                 }
 769
 770                 err = set_state_bits(tree, state, &bits);
 771                 if (err)
 772                         goto out;
 773
 774                 cache_state(state, cached_state);
 775                 merge_state(tree, state);
 776                 if (last_end == (u64)-1)
 777                         goto out;
 778
 779                 start = last_end + 1;
 780                 if (start < end && prealloc && !need_resched()) {
 781                         next_node = rb_next(node);
 782                         if (next_node) {
 783                                 state = rb_entry(next_node, struct extent_state,
 784                                                  rb_node);
 785                                 if (state->start == start)
 786                                         goto hit_next;
 787                         }
 788                 }
 789                 goto search_again;
 790         }
 791
 792         /*
 793          *     | ---- desired range ---- |
 794          * | state |
 795          *   or
 796          * | ------------- state -------------- |
 797          *
 798          * We need to split the extent we found, and may flip bits on
 799          * second half.
 800          *
 801          * If the extent we found extends past our
 802          * range, we just split and search again.  It'll get split
 803          * again the next time though.
 804          *
 805          * If the extent we found is inside our range, we set the
 806          * desired bit on it.
 807          */
 808         if (state->start < start) {
 809                 if (state->state & exclusive_bits) {
 810                         *failed_start = start;
 811                         err = -EEXIST;
 812                         goto out;
 813                 }
 814                 err = split_state(tree, state, prealloc, start);
 815                 BUG_ON(err == -EEXIST);
 816                 prealloc = NULL;
 817                 if (err)
 818                         goto out;
 819                 if (state->end <= end) {
 820                         err = set_state_bits(tree, state, &bits);
 821                         if (err)
 822                                 goto out;
 823                         cache_state(state, cached_state);
 824                         merge_state(tree, state);
 825                         if (last_end == (u64)-1)
 826                                 goto out;
 827                         start = last_end + 1;
 828                 }
 829                 goto search_again;
 830         }
 831         /*
 832          * | ---- desired range ---- |
 833          *     | state | or               | state |
 834          *
 835          * There's a hole, we need to insert something in it and
 836          * ignore the extent we found.
 837          */
 838         if (state->start > start) {
 839                 u64 this_end;
 840                 if (end < last_start)
 841                         this_end = end;
 842                 else
 843                         this_end = last_start - 1;
 844                 err = insert_state(tree, prealloc, start, this_end,
 845                                    &bits);
 846                 BUG_ON(err == -EEXIST);
 847                 if (err) {
 848                         prealloc = NULL;
 849                         goto out;
 850                 }
 851                 cache_state(prealloc, cached_state);
 852                 prealloc = NULL;
 853                 start = this_end + 1;
 854                 goto search_again;
 855         }
 856         /*
 857          * | ---- desired range ---- |
 858          *                        | state |
 859          * We need to split the extent, and set the bit
 860          * on the first half
 861          */
 862         if (state->start <= end && state->end > end) {
 863                 if (state->state & exclusive_bits) {
 864                         *failed_start = start;
 865                         err = -EEXIST;
 866                         goto out;
 867                 }
 868                 err = split_state(tree, state, prealloc, end + 1);
 869                 BUG_ON(err == -EEXIST);
 870
 871                 err = set_state_bits(tree, prealloc, &bits);
 872                 if (err) {
 873                         prealloc = NULL;
 874                         goto out;
 875                 }
 876                 cache_state(prealloc, cached_state);
 877                 merge_state(tree, prealloc);
 878                 prealloc = NULL;
 879                 goto out;
 880         }
 881
 882         goto search_again;
 883
 884 out:
 885         spin_unlock(&tree->lock);
 886         if (prealloc)
 887                 free_extent_state(prealloc);
 888
 889         return err;
 890
 891 search_again:
 892         if (start > end)
 893                 goto out;
 894         spin_unlock(&tree->lock);
 895         if (mask & __GFP_WAIT)
 896                 cond_resched();
 897         goto again;
 898 }
 899
 900 /* wrappers around set/clear extent bit */
 901 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 902                      gfp_t mask)
 903 {
 904         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 905                               NULL, mask);
 906 }
 907
 908 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 909                     int bits, gfp_t mask)
 910 {
 911         return set_extent_bit(tree, start, end, bits, 0, NULL,
 912                               NULL, mask);
 913 }
 914
 915 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 916                       int bits, gfp_t mask)
 917 {
 918         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
 919 }
 920
 921 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
 922                         struct extent_state **cached_state, gfp_t mask)
 923 {
 924         return set_extent_bit(tree, start, end,
 925                               EXTENT_DELALLOC | EXTENT_DIRTY | EXTENT_UPTODATE,
 926                               0, NULL, cached_state, mask);
 927 }
 928
 929 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 930                        gfp_t mask)
 931 {
 932         return clear_extent_bit(tree, start, end,
 933                                 EXTENT_DIRTY | EXTENT_DELALLOC |
 934                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
 935 }
 936
 937 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 938                      gfp_t mask)
 939 {
 940         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 941                               NULL, mask);
 942 }
 943
 944 static int clear_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 945                        gfp_t mask)
 946 {
 947         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0,
 948                                 NULL, mask);
 949 }
 950
 951 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 952                         struct extent_state **cached_state, gfp_t mask)
 953 {
 954         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
 955                               NULL, cached_state, mask);
 956 }
 957
 958 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
 959                                  u64 end, struct extent_state **cached_state,
 960                                  gfp_t mask)
 961 {
 962         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
 963                                 cached_state, mask);
 964 }
 965
 966 int wait_on_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 967 {
 968         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 969 }
 970
 971 /*
 972  * either insert or lock state struct between start and end use mask to tell
 973  * us if waiting is desired.
 974  */
 975 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 976                      int bits, struct extent_state **cached_state, gfp_t mask)
 977 {
 978         int err;
 979         u64 failed_start;
 980         while (1) {
 981                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
 982                                      EXTENT_LOCKED, &failed_start,
 983                                      cached_state, mask);
 984                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
 985                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
 986                         start = failed_start;
 987                 } else {
 988                         break;
 989                 }
 990                 WARN_ON(start > end);
 991         }
 992         return err;
 993 }
 994
 995 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
 996 {
 997         return lock_extent_bits(tree, start, end, 0, NULL, mask);
 998 }
 999
1000 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1001                     gfp_t mask)
1002 {
1003         int err;
1004         u64 failed_start;
1005
1006         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1007                              &failed_start, NULL, mask);
1008         if (err == -EEXIST) {
1009                 if (failed_start > start)
1010                         clear_extent_bit(tree, start, failed_start - 1,
1011                                          EXTENT_LOCKED, 1, 0, NULL, mask);
1012                 return 0;
1013         }
1014         return 1;
1015 }
1016
1017 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1018                          struct extent_state **cached, gfp_t mask)
1019 {
1020         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1021                                 mask);
1022 }
1023
1024 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1025 {
1026         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1027                                 mask);
1028 }
1029
1030 /*
1031  * helper function to set pages and extents in the tree dirty
1032  */
1033 int set_range_dirty(struct extent_io_tree *tree, u64 start, u64 end)
1034 {
1035         unsigned long index = start >> PAGE_CACHE_SHIFT;
1036         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1037         struct page *page;
1038
1039         while (index <= end_index) {
1040                 page = find_get_page(tree->mapping, index);
1041                 BUG_ON(!page);
1042                 __set_page_dirty_nobuffers(page);
1043                 page_cache_release(page);
1044                 index++;
1045         }
1046         return 0;
1047 }
1048
1049 /*
1050  * helper function to set both pages and extents in the tree writeback
1051  */
1052 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1053 {
1054         unsigned long index = start >> PAGE_CACHE_SHIFT;
1055         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1056         struct page *page;
1057
1058         while (index <= end_index) {
1059                 page = find_get_page(tree->mapping, index);
1060                 BUG_ON(!page);
1061                 set_page_writeback(page);
1062                 page_cache_release(page);
1063                 index++;
1064         }
1065         return 0;
1066 }
1067
1068 /*
1069  * find the first offset in the io tree with 'bits' set. zero is
1070  * returned if we find something, and *start_ret and *end_ret are
1071  * set to reflect the state struct that was found.
1072  *
1073  * If nothing was found, 1 is returned, < 0 on error
1074  */
1075 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1076                           u64 *start_ret, u64 *end_ret, int bits)
1077 {
1078         struct rb_node *node;
1079         struct extent_state *state;
1080         int ret = 1;
1081
1082         spin_lock(&tree->lock);
1083         /*
1084          * this search will find all the extents that end after
1085          * our range starts.
1086          */
1087         node = tree_search(tree, start);
1088         if (!node)
1089                 goto out;
1090
1091         while (1) {
1092                 state = rb_entry(node, struct extent_state, rb_node);
1093                 if (state->end >= start && (state->state & bits)) {
1094                         *start_ret = state->start;
1095                         *end_ret = state->end;
1096                         ret = 0;
1097                         break;
1098                 }
1099                 node = rb_next(node);
1100                 if (!node)
1101                         break;
1102         }
1103 out:
1104         spin_unlock(&tree->lock);
1105         return ret;
1106 }
1107
1108 /* find the first state struct with 'bits' set after 'start', and
1109  * return it.  tree->lock must be held.  NULL will returned if
1110  * nothing was found after 'start'
1111  */
1112 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1113                                                  u64 start, int bits)
1114 {
1115         struct rb_node *node;
1116         struct extent_state *state;
1117
1118         /*
1119          * this search will find all the extents that end after
1120          * our range starts.
1121          */
1122         node = tree_search(tree, start);
1123         if (!node)
1124                 goto out;
1125
1126         while (1) {
1127                 state = rb_entry(node, struct extent_state, rb_node);
1128                 if (state->end >= start && (state->state & bits))
1129                         return state;
1130
1131                 node = rb_next(node);
1132                 if (!node)
1133                         break;
1134         }
1135 out:
1136         return NULL;
1137 }
1138
1139 /*
1140  * find a contiguous range of bytes in the file marked as delalloc, not
1141  * more than 'max_bytes'.  start and end are used to return the range,
1142  *
1143  * 1 is returned if we find something, 0 if nothing was in the tree
1144  */
1145 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1146                                         u64 *start, u64 *end, u64 max_bytes,
1147                                         struct extent_state **cached_state)
1148 {
1149         struct rb_node *node;
1150         struct extent_state *state;
1151         u64 cur_start = *start;
1152         u64 found = 0;
1153         u64 total_bytes = 0;
1154
1155         spin_lock(&tree->lock);
1156
1157         /*
1158          * this search will find all the extents that end after
1159          * our range starts.
1160          */
1161         node = tree_search(tree, cur_start);
1162         if (!node) {
1163                 if (!found)
1164                         *end = (u64)-1;
1165                 goto out;
1166         }
1167
1168         while (1) {
1169                 state = rb_entry(node, struct extent_state, rb_node);
1170                 if (found && (state->start != cur_start ||
1171                               (state->state & EXTENT_BOUNDARY))) {
1172                         goto out;
1173                 }
1174                 if (!(state->state & EXTENT_DELALLOC)) {
1175                         if (!found)
1176                                 *end = state->end;
1177                         goto out;
1178                 }
1179                 if (!found) {
1180                         *start = state->start;
1181                         *cached_state = state;
1182                         atomic_inc(&state->refs);
1183                 }
1184                 found++;
1185                 *end = state->end;
1186                 cur_start = state->end + 1;
1187                 node = rb_next(node);
1188                 if (!node)
1189                         break;
1190                 total_bytes += state->end - state->start + 1;
1191                 if (total_bytes >= max_bytes)
1192                         break;
1193         }
1194 out:
1195         spin_unlock(&tree->lock);
1196         return found;
1197 }
1198
1199 static noinline int __unlock_for_delalloc(struct inode *inode,
1200                                           struct page *locked_page,
1201                                           u64 start, u64 end)
1202 {
1203         int ret;
1204         struct page *pages[16];
1205         unsigned long index = start >> PAGE_CACHE_SHIFT;
1206         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1207         unsigned long nr_pages = end_index - index + 1;
1208         int i;
1209
1210         if (index == locked_page->index && end_index == index)
1211                 return 0;
1212
1213         while (nr_pages > 0) {
1214                 ret = find_get_pages_contig(inode->i_mapping, index,
1215                                      min_t(unsigned long, nr_pages,
1216                                      ARRAY_SIZE(pages)), pages);
1217                 for (i = 0; i < ret; i++) {
1218                         if (pages[i] != locked_page)
1219                                 unlock_page(pages[i]);
1220                         page_cache_release(pages[i]);
1221                 }
1222                 nr_pages -= ret;
1223                 index += ret;
1224                 cond_resched();
1225         }
1226         return 0;
1227 }
1228
1229 static noinline int lock_delalloc_pages(struct inode *inode,
1230                                         struct page *locked_page,
1231                                         u64 delalloc_start,
1232                                         u64 delalloc_end)
1233 {
1234         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1235         unsigned long start_index = index;
1236         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1237         unsigned long pages_locked = 0;
1238         struct page *pages[16];
1239         unsigned long nrpages;
1240         int ret;
1241         int i;
1242
1243         /* the caller is responsible for locking the start index */
1244         if (index == locked_page->index && index == end_index)
1245                 return 0;
1246
1247         /* skip the page at the start index */
1248         nrpages = end_index - index + 1;
1249         while (nrpages > 0) {
1250                 ret = find_get_pages_contig(inode->i_mapping, index,
1251                                      min_t(unsigned long,
1252                                      nrpages, ARRAY_SIZE(pages)), pages);
1253                 if (ret == 0) {
1254                         ret = -EAGAIN;
1255                         goto done;
1256                 }
1257                 /* now we have an array of pages, lock them all */
1258                 for (i = 0; i < ret; i++) {
1259                         /*
1260                          * the caller is taking responsibility for
1261                          * locked_page
1262                          */
1263                         if (pages[i] != locked_page) {
1264                                 lock_page(pages[i]);
1265                                 if (!PageDirty(pages[i]) ||
1266                                     pages[i]->mapping != inode->i_mapping) {
1267                                         ret = -EAGAIN;
1268                                         unlock_page(pages[i]);
1269                                         page_cache_release(pages[i]);
1270                                         goto done;
1271                                 }
1272                         }
1273                         page_cache_release(pages[i]);
1274                         pages_locked++;
1275                 }
1276                 nrpages -= ret;
1277                 index += ret;
1278                 cond_resched();
1279         }
1280         ret = 0;
1281 done:
1282         if (ret && pages_locked) {
1283                 __unlock_for_delalloc(inode, locked_page,
1284                               delalloc_start,
1285                               ((u64)(start_index + pages_locked - 1)) <<
1286                               PAGE_CACHE_SHIFT);
1287         }
1288         return ret;
1289 }
1290
1291 /*
1292  * find a contiguous range of bytes in the file marked as delalloc, not
1293  * more than 'max_bytes'.  start and end are used to return the range,
1294  *
1295  * 1 is returned if we find something, 0 if nothing was in the tree
1296  */
1297 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1298                                              struct extent_io_tree *tree,
1299                                              struct page *locked_page,
1300                                              u64 *start, u64 *end,
1301                                              u64 max_bytes)
1302 {
1303         u64 delalloc_start;
1304         u64 delalloc_end;
1305         u64 found;
1306         struct extent_state *cached_state = NULL;
1307         int ret;
1308         int loops = 0;
1309
1310 again:
1311         /* step one, find a bunch of delalloc bytes starting at start */
1312         delalloc_start = *start;
1313         delalloc_end = 0;
1314         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1315                                     max_bytes, &cached_state);
1316         if (!found || delalloc_end <= *start) {
1317                 *start = delalloc_start;
1318                 *end = delalloc_end;
1319                 free_extent_state(cached_state);
1320                 return found;
1321         }
1322
1323         /*
1324          * start comes from the offset of locked_page.  We have to lock
1325          * pages in order, so we can't process delalloc bytes before
1326          * locked_page
1327          */
1328         if (delalloc_start < *start)
1329                 delalloc_start = *start;
1330
1331         /*
1332          * make sure to limit the number of pages we try to lock down
1333          * if we're looping.
1334          */
1335         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1336                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1337
1338         /* step two, lock all the pages after the page that has start */
1339         ret = lock_delalloc_pages(inode, locked_page,
1340                                   delalloc_start, delalloc_end);
1341         if (ret == -EAGAIN) {
1342                 /* some of the pages are gone, lets avoid looping by
1343                  * shortening the size of the delalloc range we're searching
1344                  */
1345                 free_extent_state(cached_state);
1346                 if (!loops) {
1347                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1348                         max_bytes = PAGE_CACHE_SIZE - offset;
1349                         loops = 1;
1350                         goto again;
1351                 } else {
1352                         found = 0;
1353                         goto out_failed;
1354                 }
1355         }
1356         BUG_ON(ret);
1357
1358         /* step three, lock the state bits for the whole range */
1359         lock_extent_bits(tree, delalloc_start, delalloc_end,
1360                          0, &cached_state, GFP_NOFS);
1361
1362         /* then test to make sure it is all still delalloc */
1363         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1364                              EXTENT_DELALLOC, 1, cached_state);
1365         if (!ret) {
1366                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1367                                      &cached_state, GFP_NOFS);
1368                 __unlock_for_delalloc(inode, locked_page,
1369                               delalloc_start, delalloc_end);
1370                 cond_resched();
1371                 goto again;
1372         }
1373         free_extent_state(cached_state);
1374         *start = delalloc_start;
1375         *end = delalloc_end;
1376 out_failed:
1377         return found;
1378 }
1379
1380 int extent_clear_unlock_delalloc(struct inode *inode,
1381                                 struct extent_io_tree *tree,
1382                                 u64 start, u64 end, struct page *locked_page,
1383                                 unsigned long op)
1384 {
1385         int ret;
1386         struct page *pages[16];
1387         unsigned long index = start >> PAGE_CACHE_SHIFT;
1388         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1389         unsigned long nr_pages = end_index - index + 1;
1390         int i;
1391         int clear_bits = 0;
1392
1393         if (op & EXTENT_CLEAR_UNLOCK)
1394                 clear_bits |= EXTENT_LOCKED;
1395         if (op & EXTENT_CLEAR_DIRTY)
1396                 clear_bits |= EXTENT_DIRTY;
1397
1398         if (op & EXTENT_CLEAR_DELALLOC)
1399                 clear_bits |= EXTENT_DELALLOC;
1400
1401         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1402         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1403                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1404                     EXTENT_SET_PRIVATE2)))
1405                 return 0;
1406
1407         while (nr_pages > 0) {
1408                 ret = find_get_pages_contig(inode->i_mapping, index,
1409                                      min_t(unsigned long,
1410                                      nr_pages, ARRAY_SIZE(pages)), pages);
1411                 for (i = 0; i < ret; i++) {
1412
1413                         if (op & EXTENT_SET_PRIVATE2)
1414                                 SetPagePrivate2(pages[i]);
1415
1416                         if (pages[i] == locked_page) {
1417                                 page_cache_release(pages[i]);
1418                                 continue;
1419                         }
1420                         if (op & EXTENT_CLEAR_DIRTY)
1421                                 clear_page_dirty_for_io(pages[i]);
1422                         if (op & EXTENT_SET_WRITEBACK)
1423                                 set_page_writeback(pages[i]);
1424                         if (op & EXTENT_END_WRITEBACK)
1425                                 end_page_writeback(pages[i]);
1426                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1427                                 unlock_page(pages[i]);
1428                         page_cache_release(pages[i]);
1429                 }
1430                 nr_pages -= ret;
1431                 index += ret;
1432                 cond_resched();
1433         }
1434         return 0;
1435 }
1436
1437 /*
1438  * count the number of bytes in the tree that have a given bit(s)
1439  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1440  * cached.  The total number found is returned.
1441  */
1442 u64 count_range_bits(struct extent_io_tree *tree,
1443                      u64 *start, u64 search_end, u64 max_bytes,
1444                      unsigned long bits, int contig)
1445 {
1446         struct rb_node *node;
1447         struct extent_state *state;
1448         u64 cur_start = *start;
1449         u64 total_bytes = 0;
1450         u64 last = 0;
1451         int found = 0;
1452
1453         if (search_end <= cur_start) {
1454                 WARN_ON(1);
1455                 return 0;
1456         }
1457
1458         spin_lock(&tree->lock);
1459         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1460                 total_bytes = tree->dirty_bytes;
1461                 goto out;
1462         }
1463         /*
1464          * this search will find all the extents that end after
1465          * our range starts.
1466          */
1467         node = tree_search(tree, cur_start);
1468         if (!node)
1469                 goto out;
1470
1471         while (1) {
1472                 state = rb_entry(node, struct extent_state, rb_node);
1473                 if (state->start > search_end)
1474                         break;
1475                 if (contig && found && state->start > last + 1)
1476                         break;
1477                 if (state->end >= cur_start && (state->state & bits) == bits) {
1478                         total_bytes += min(search_end, state->end) + 1 -
1479                                        max(cur_start, state->start);
1480                         if (total_bytes >= max_bytes)
1481                                 break;
1482                         if (!found) {
1483                                 *start = state->start;
1484                                 found = 1;
1485                         }
1486                         last = state->end;
1487                 } else if (contig && found) {
1488                         break;
1489                 }
1490                 node = rb_next(node);
1491                 if (!node)
1492                         break;
1493         }
1494 out:
1495         spin_unlock(&tree->lock);
1496         return total_bytes;
1497 }
1498
1499 /*
1500  * set the private field for a given byte offset in the tree.  If there isn't
1501  * an extent_state there already, this does nothing.
1502  */
1503 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1504 {
1505         struct rb_node *node;
1506         struct extent_state *state;
1507         int ret = 0;
1508
1509         spin_lock(&tree->lock);
1510         /*
1511          * this search will find all the extents that end after
1512          * our range starts.
1513          */
1514         node = tree_search(tree, start);
1515         if (!node) {
1516                 ret = -ENOENT;
1517                 goto out;
1518         }
1519         state = rb_entry(node, struct extent_state, rb_node);
1520         if (state->start != start) {
1521                 ret = -ENOENT;
1522                 goto out;
1523         }
1524         state->private = private;
1525 out:
1526         spin_unlock(&tree->lock);
1527         return ret;
1528 }
1529
1530 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1531 {
1532         struct rb_node *node;
1533         struct extent_state *state;
1534         int ret = 0;
1535
1536         spin_lock(&tree->lock);
1537         /*
1538          * this search will find all the extents that end after
1539          * our range starts.
1540          */
1541         node = tree_search(tree, start);
1542         if (!node) {
1543                 ret = -ENOENT;
1544                 goto out;
1545         }
1546         state = rb_entry(node, struct extent_state, rb_node);
1547         if (state->start != start) {
1548                 ret = -ENOENT;
1549                 goto out;
1550         }
1551         *private = state->private;
1552 out:
1553         spin_unlock(&tree->lock);
1554         return ret;
1555 }
1556
1557 /*
1558  * searches a range in the state tree for a given mask.
1559  * If 'filled' == 1, this returns 1 only if every extent in the tree
1560  * has the bits set.  Otherwise, 1 is returned if any bit in the
1561  * range is found set.
1562  */
1563 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1564                    int bits, int filled, struct extent_state *cached)
1565 {
1566         struct extent_state *state = NULL;
1567         struct rb_node *node;
1568         int bitset = 0;
1569
1570         spin_lock(&tree->lock);
1571         if (cached && cached->tree && cached->start == start)
1572                 node = &cached->rb_node;
1573         else
1574                 node = tree_search(tree, start);
1575         while (node && start <= end) {
1576                 state = rb_entry(node, struct extent_state, rb_node);
1577
1578                 if (filled && state->start > start) {
1579                         bitset = 0;
1580                         break;
1581                 }
1582
1583                 if (state->start > end)
1584                         break;
1585
1586                 if (state->state & bits) {
1587                         bitset = 1;
1588                         if (!filled)
1589                                 break;
1590                 } else if (filled) {
1591                         bitset = 0;
1592                         break;
1593                 }
1594
1595                 if (state->end == (u64)-1)
1596                         break;
1597
1598                 start = state->end + 1;
1599                 if (start > end)
1600                         break;
1601                 node = rb_next(node);
1602                 if (!node) {
1603                         if (filled)
1604                                 bitset = 0;
1605                         break;
1606                 }
1607         }
1608         spin_unlock(&tree->lock);
1609         return bitset;
1610 }
1611
1612 /*
1613  * helper function to set a given page up to date if all the
1614  * extents in the tree for that page are up to date
1615  */
1616 static int check_page_uptodate(struct extent_io_tree *tree,
1617                                struct page *page)
1618 {
1619         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1620         u64 end = start + PAGE_CACHE_SIZE - 1;
1621         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1622                 SetPageUptodate(page);
1623         return 0;
1624 }
1625
1626 /*
1627  * helper function to unlock a page if all the extents in the tree
1628  * for that page are unlocked
1629  */
1630 static int check_page_locked(struct extent_io_tree *tree,
1631                              struct page *page)
1632 {
1633         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1634         u64 end = start + PAGE_CACHE_SIZE - 1;
1635         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1636                 unlock_page(page);
1637         return 0;
1638 }
1639
1640 /*
1641  * helper function to end page writeback if all the extents
1642  * in the tree for that page are done with writeback
1643  */
1644 static int check_page_writeback(struct extent_io_tree *tree,
1645                              struct page *page)
1646 {
1647         end_page_writeback(page);
1648         return 0;
1649 }
1650
1651 /* lots and lots of room for performance fixes in the end_bio funcs */
1652
1653 /*
1654  * after a writepage IO is done, we need to:
1655  * clear the uptodate bits on error
1656  * clear the writeback bits in the extent tree for this IO
1657  * end_page_writeback if the page has no more pending IO
1658  *
1659  * Scheduling is not allowed, so the extent state tree is expected
1660  * to have one and only one object corresponding to this IO.
1661  */
1662 static void end_bio_extent_writepage(struct bio *bio, int err)
1663 {
1664         int uptodate = err == 0;
1665         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1666         struct extent_io_tree *tree;
1667         u64 start;
1668         u64 end;
1669         int whole_page;
1670         int ret;
1671
1672         do {
1673                 struct page *page = bvec->bv_page;
1674                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1675
1676                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1677                          bvec->bv_offset;
1678                 end = start + bvec->bv_len - 1;
1679
1680                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1681                         whole_page = 1;
1682                 else
1683                         whole_page = 0;
1684
1685                 if (--bvec >= bio->bi_io_vec)
1686                         prefetchw(&bvec->bv_page->flags);
1687                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1688                         ret = tree->ops->writepage_end_io_hook(page, start,
1689                                                        end, NULL, uptodate);
1690                         if (ret)
1691                                 uptodate = 0;
1692                 }
1693
1694                 if (!uptodate && tree->ops &&
1695                     tree->ops->writepage_io_failed_hook) {
1696                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1697                                                          start, end, NULL);
1698                         if (ret == 0) {
1699                                 uptodate = (err == 0);
1700                                 continue;
1701                         }
1702                 }
1703
1704                 if (!uptodate) {
1705                         clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
1706                         ClearPageUptodate(page);
1707                         SetPageError(page);
1708                 }
1709
1710                 if (whole_page)
1711                         end_page_writeback(page);
1712                 else
1713                         check_page_writeback(tree, page);
1714         } while (bvec >= bio->bi_io_vec);
1715
1716         bio_put(bio);
1717 }
1718
1719 /*
1720  * after a readpage IO is done, we need to:
1721  * clear the uptodate bits on error
1722  * set the uptodate bits if things worked
1723  * set the page up to date if all extents in the tree are uptodate
1724  * clear the lock bit in the extent tree
1725  * unlock the page if there are no other extents locked for it
1726  *
1727  * Scheduling is not allowed, so the extent state tree is expected
1728  * to have one and only one object corresponding to this IO.
1729  */
1730 static void end_bio_extent_readpage(struct bio *bio, int err)
1731 {
1732         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1733         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
1734         struct bio_vec *bvec = bio->bi_io_vec;
1735         struct extent_io_tree *tree;
1736         u64 start;
1737         u64 end;
1738         int whole_page;
1739         int ret;
1740
1741         if (err)
1742                 uptodate = 0;
1743
1744         do {
1745                 struct page *page = bvec->bv_page;
1746                 struct extent_state *cached = NULL;
1747                 struct extent_state *state;
1748
1749                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1750
1751                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1752                         bvec->bv_offset;
1753                 end = start + bvec->bv_len - 1;
1754
1755                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1756                         whole_page = 1;
1757                 else
1758                         whole_page = 0;
1759
1760                 if (++bvec <= bvec_end)
1761                         prefetchw(&bvec->bv_page->flags);
1762
1763                 spin_lock(&tree->lock);
1764                 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
1765                 if (state && state->start == start) {
1766                         /*
1767                          * take a reference on the state, unlock will drop
1768                          * the ref
1769                          */
1770                         cache_state(state, &cached);
1771                 }
1772                 spin_unlock(&tree->lock);
1773
1774                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1775                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1776                                                               state);
1777                         if (ret)
1778                                 uptodate = 0;
1779                 }
1780                 if (!uptodate && tree->ops &&
1781                     tree->ops->readpage_io_failed_hook) {
1782                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1783                                                          start, end, NULL);
1784                         if (ret == 0) {
1785                                 uptodate =
1786                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1787                                 if (err)
1788                                         uptodate = 0;
1789                                 uncache_state(&cached);
1790                                 continue;
1791                         }
1792                 }
1793
1794                 if (uptodate) {
1795                         set_extent_uptodate(tree, start, end, &cached,
1796                                             GFP_ATOMIC);
1797                 }
1798                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
1799
1800                 if (whole_page) {
1801                         if (uptodate) {
1802                                 SetPageUptodate(page);
1803                         } else {
1804                                 ClearPageUptodate(page);
1805                                 SetPageError(page);
1806                         }
1807                         unlock_page(page);
1808                 } else {
1809                         if (uptodate) {
1810                                 check_page_uptodate(tree, page);
1811                         } else {
1812                                 ClearPageUptodate(page);
1813                                 SetPageError(page);
1814                         }
1815                         check_page_locked(tree, page);
1816                 }
1817         } while (bvec <= bvec_end);
1818
1819         bio_put(bio);
1820 }
1821
1822 /*
1823  * IO done from prepare_write is pretty simple, we just unlock
1824  * the structs in the extent tree when done, and set the uptodate bits
1825  * as appropriate.
1826  */
1827 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1828 {
1829         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1830         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1831         struct extent_io_tree *tree;
1832         u64 start;
1833         u64 end;
1834
1835         do {
1836                 struct page *page = bvec->bv_page;
1837                 struct extent_state *cached = NULL;
1838                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1839
1840                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1841                         bvec->bv_offset;
1842                 end = start + bvec->bv_len - 1;
1843
1844                 if (--bvec >= bio->bi_io_vec)
1845                         prefetchw(&bvec->bv_page->flags);
1846
1847                 if (uptodate) {
1848                         set_extent_uptodate(tree, start, end, &cached,
1849                                             GFP_ATOMIC);
1850                 } else {
1851                         ClearPageUptodate(page);
1852                         SetPageError(page);
1853                 }
1854
1855                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
1856
1857         } while (bvec >= bio->bi_io_vec);
1858
1859         bio_put(bio);
1860 }
1861
1862 struct bio *
1863 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1864                 gfp_t gfp_flags)
1865 {
1866         struct bio *bio;
1867
1868         bio = bio_alloc(gfp_flags, nr_vecs);
1869
1870         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1871                 while (!bio && (nr_vecs /= 2))
1872                         bio = bio_alloc(gfp_flags, nr_vecs);
1873         }
1874
1875         if (bio) {
1876                 bio->bi_size = 0;
1877                 bio->bi_bdev = bdev;
1878                 bio->bi_sector = first_sector;
1879         }
1880         return bio;
1881 }
1882
1883 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1884                           unsigned long bio_flags)
1885 {
1886         int ret = 0;
1887         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1888         struct page *page = bvec->bv_page;
1889         struct extent_io_tree *tree = bio->bi_private;
1890         u64 start;
1891
1892         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1893
1894         bio->bi_private = NULL;
1895
1896         bio_get(bio);
1897
1898         if (tree->ops && tree->ops->submit_bio_hook)
1899                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1900                                            mirror_num, bio_flags, start);
1901         else
1902                 submit_bio(rw, bio);
1903         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1904                 ret = -EOPNOTSUPP;
1905         bio_put(bio);
1906         return ret;
1907 }
1908
1909 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1910                               struct page *page, sector_t sector,
1911                               size_t size, unsigned long offset,
1912                               struct block_device *bdev,
1913                               struct bio **bio_ret,
1914                               unsigned long max_pages,
1915                               bio_end_io_t end_io_func,
1916                               int mirror_num,
1917                               unsigned long prev_bio_flags,
1918                               unsigned long bio_flags)
1919 {
1920         int ret = 0;
1921         struct bio *bio;
1922         int nr;
1923         int contig = 0;
1924         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1925         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1926         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1927
1928         if (bio_ret && *bio_ret) {
1929                 bio = *bio_ret;
1930                 if (old_compressed)
1931                         contig = bio->bi_sector == sector;
1932                 else
1933                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
1934                                 sector;
1935
1936                 if (prev_bio_flags != bio_flags || !contig ||
1937                     (tree->ops && tree->ops->merge_bio_hook &&
1938                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
1939                                                bio_flags)) ||
1940                     bio_add_page(bio, page, page_size, offset) < page_size) {
1941                         ret = submit_one_bio(rw, bio, mirror_num,
1942                                              prev_bio_flags);
1943                         bio = NULL;
1944                 } else {
1945                         return 0;
1946                 }
1947         }
1948         if (this_compressed)
1949                 nr = BIO_MAX_PAGES;
1950         else
1951                 nr = bio_get_nr_vecs(bdev);
1952
1953         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1954         if (!bio)
1955                 return -ENOMEM;
1956
1957         bio_add_page(bio, page, page_size, offset);
1958         bio->bi_end_io = end_io_func;
1959         bio->bi_private = tree;
1960
1961         if (bio_ret)
1962                 *bio_ret = bio;
1963         else
1964                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1965
1966         return ret;
1967 }
1968
1969 void set_page_extent_mapped(struct page *page)
1970 {
1971         if (!PagePrivate(page)) {
1972                 SetPagePrivate(page);
1973                 page_cache_get(page);
1974                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1975         }
1976 }
1977
1978 static void set_page_extent_head(struct page *page, unsigned long len)
1979 {
1980         WARN_ON(!PagePrivate(page));
1981         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1982 }
1983
1984 /*
1985  * basic readpage implementation.  Locked extent state structs are inserted
1986  * into the tree that are removed when the IO is done (by the end_io
1987  * handlers)
1988  */
1989 static int __extent_read_full_page(struct extent_io_tree *tree,
1990                                    struct page *page,
1991                                    get_extent_t *get_extent,
1992                                    struct bio **bio, int mirror_num,
1993                                    unsigned long *bio_flags)
1994 {
1995         struct inode *inode = page->mapping->host;
1996         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1997         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1998         u64 end;
1999         u64 cur = start;
2000         u64 extent_offset;
2001         u64 last_byte = i_size_read(inode);
2002         u64 block_start;
2003         u64 cur_end;
2004         sector_t sector;
2005         struct extent_map *em;
2006         struct block_device *bdev;
2007         struct btrfs_ordered_extent *ordered;
2008         int ret;
2009         int nr = 0;
2010         size_t page_offset = 0;
2011         size_t iosize;
2012         size_t disk_io_size;
2013         size_t blocksize = inode->i_sb->s_blocksize;
2014         unsigned long this_bio_flag = 0;
2015
2016         set_page_extent_mapped(page);
2017
2018         end = page_end;
2019         while (1) {
2020                 lock_extent(tree, start, end, GFP_NOFS);
2021                 ordered = btrfs_lookup_ordered_extent(inode, start);
2022                 if (!ordered)
2023                         break;
2024                 unlock_extent(tree, start, end, GFP_NOFS);
2025                 btrfs_start_ordered_extent(inode, ordered, 1);
2026                 btrfs_put_ordered_extent(ordered);
2027         }
2028
2029         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2030                 char *userpage;
2031                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2032
2033                 if (zero_offset) {
2034                         iosize = PAGE_CACHE_SIZE - zero_offset;
2035                         userpage = kmap_atomic(page, KM_USER0);
2036                         memset(userpage + zero_offset, 0, iosize);
2037                         flush_dcache_page(page);
2038                         kunmap_atomic(userpage, KM_USER0);
2039                 }
2040         }
2041         while (cur <= end) {
2042                 if (cur >= last_byte) {
2043                         char *userpage;
2044                         struct extent_state *cached = NULL;
2045
2046                         iosize = PAGE_CACHE_SIZE - page_offset;
2047                         userpage = kmap_atomic(page, KM_USER0);
2048                         memset(userpage + page_offset, 0, iosize);
2049                         flush_dcache_page(page);
2050                         kunmap_atomic(userpage, KM_USER0);
2051                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2052                                             &cached, GFP_NOFS);
2053                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2054                                              &cached, GFP_NOFS);
2055                         break;
2056                 }
2057                 em = get_extent(inode, page, page_offset, cur,
2058                                 end - cur + 1, 0);
2059                 if (IS_ERR(em) || !em) {
2060                         SetPageError(page);
2061                         unlock_extent(tree, cur, end, GFP_NOFS);
2062                         break;
2063                 }
2064                 extent_offset = cur - em->start;
2065                 BUG_ON(extent_map_end(em) <= cur);
2066                 BUG_ON(end < cur);
2067
2068                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2069                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2070                         extent_set_compress_type(&this_bio_flag,
2071                                                  em->compress_type);
2072                 }
2073
2074                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2075                 cur_end = min(extent_map_end(em) - 1, end);
2076                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2077                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2078                         disk_io_size = em->block_len;
2079                         sector = em->block_start >> 9;
2080                 } else {
2081                         sector = (em->block_start + extent_offset) >> 9;
2082                         disk_io_size = iosize;
2083                 }
2084                 bdev = em->bdev;
2085                 block_start = em->block_start;
2086                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2087                         block_start = EXTENT_MAP_HOLE;
2088                 free_extent_map(em);
2089                 em = NULL;
2090
2091                 /* we've found a hole, just zero and go on */
2092                 if (block_start == EXTENT_MAP_HOLE) {
2093                         char *userpage;
2094                         struct extent_state *cached = NULL;
2095
2096                         userpage = kmap_atomic(page, KM_USER0);
2097                         memset(userpage + page_offset, 0, iosize);
2098                         flush_dcache_page(page);
2099                         kunmap_atomic(userpage, KM_USER0);
2100
2101                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2102                                             &cached, GFP_NOFS);
2103                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2104                                              &cached, GFP_NOFS);
2105                         cur = cur + iosize;
2106                         page_offset += iosize;
2107                         continue;
2108                 }
2109                 /* the get_extent function already copied into the page */
2110                 if (test_range_bit(tree, cur, cur_end,
2111                                    EXTENT_UPTODATE, 1, NULL)) {
2112                         check_page_uptodate(tree, page);
2113                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2114                         cur = cur + iosize;
2115                         page_offset += iosize;
2116                         continue;
2117                 }
2118                 /* we have an inline extent but it didn't get marked up
2119                  * to date.  Error out
2120                  */
2121                 if (block_start == EXTENT_MAP_INLINE) {
2122                         SetPageError(page);
2123                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2124                         cur = cur + iosize;
2125                         page_offset += iosize;
2126                         continue;
2127                 }
2128
2129                 ret = 0;
2130                 if (tree->ops && tree->ops->readpage_io_hook) {
2131                         ret = tree->ops->readpage_io_hook(page, cur,
2132                                                           cur + iosize - 1);
2133                 }
2134                 if (!ret) {
2135                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2136                         pnr -= page->index;
2137                         ret = submit_extent_page(READ, tree, page,
2138                                          sector, disk_io_size, page_offset,
2139                                          bdev, bio, pnr,
2140                                          end_bio_extent_readpage, mirror_num,
2141                                          *bio_flags,
2142                                          this_bio_flag);
2143                         nr++;
2144                         *bio_flags = this_bio_flag;
2145                 }
2146                 if (ret)
2147                         SetPageError(page);
2148                 cur = cur + iosize;
2149                 page_offset += iosize;
2150         }
2151         if (!nr) {
2152                 if (!PageError(page))
2153                         SetPageUptodate(page);
2154                 unlock_page(page);
2155         }
2156         return 0;
2157 }
2158
2159 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2160                             get_extent_t *get_extent)
2161 {
2162         struct bio *bio = NULL;
2163         unsigned long bio_flags = 0;
2164         int ret;
2165
2166         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2167                                       &bio_flags);
2168         if (bio)
2169                 ret = submit_one_bio(READ, bio, 0, bio_flags);
2170         return ret;
2171 }
2172
2173 static noinline void update_nr_written(struct page *page,
2174                                       struct writeback_control *wbc,
2175                                       unsigned long nr_written)
2176 {
2177         wbc->nr_to_write -= nr_written;
2178         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2179             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2180                 page->mapping->writeback_index = page->index + nr_written;
2181 }
2182
2183 /*
2184  * the writepage semantics are similar to regular writepage.  extent
2185  * records are inserted to lock ranges in the tree, and as dirty areas
2186  * are found, they are marked writeback.  Then the lock bits are removed
2187  * and the end_io handler clears the writeback ranges
2188  */
2189 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2190                               void *data)
2191 {
2192         struct inode *inode = page->mapping->host;
2193         struct extent_page_data *epd = data;
2194         struct extent_io_tree *tree = epd->tree;
2195         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2196         u64 delalloc_start;
2197         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2198         u64 end;
2199         u64 cur = start;
2200         u64 extent_offset;
2201         u64 last_byte = i_size_read(inode);
2202         u64 block_start;
2203         u64 iosize;
2204         sector_t sector;
2205         struct extent_state *cached_state = NULL;
2206         struct extent_map *em;
2207         struct block_device *bdev;
2208         int ret;
2209         int nr = 0;
2210         size_t pg_offset = 0;
2211         size_t blocksize;
2212         loff_t i_size = i_size_read(inode);
2213         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2214         u64 nr_delalloc;
2215         u64 delalloc_end;
2216         int page_started;
2217         int compressed;
2218         int write_flags;
2219         unsigned long nr_written = 0;
2220
2221         if (wbc->sync_mode == WB_SYNC_ALL)
2222                 write_flags = WRITE_SYNC;
2223         else
2224                 write_flags = WRITE;
2225
2226         trace___extent_writepage(page, inode, wbc);
2227
2228         WARN_ON(!PageLocked(page));
2229         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2230         if (page->index > end_index ||
2231            (page->index == end_index && !pg_offset)) {
2232                 page->mapping->a_ops->invalidatepage(page, 0);
2233                 unlock_page(page);
2234                 return 0;
2235         }
2236
2237         if (page->index == end_index) {
2238                 char *userpage;
2239
2240                 userpage = kmap_atomic(page, KM_USER0);
2241                 memset(userpage + pg_offset, 0,
2242                        PAGE_CACHE_SIZE - pg_offset);
2243                 kunmap_atomic(userpage, KM_USER0);
2244                 flush_dcache_page(page);
2245         }
2246         pg_offset = 0;
2247
2248         set_page_extent_mapped(page);
2249
2250         delalloc_start = start;
2251         delalloc_end = 0;
2252         page_started = 0;
2253         if (!epd->extent_locked) {
2254                 u64 delalloc_to_write = 0;
2255                 /*
2256                  * make sure the wbc mapping index is at least updated
2257                  * to this page.
2258                  */
2259                 update_nr_written(page, wbc, 0);
2260
2261                 while (delalloc_end < page_end) {
2262                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2263                                                        page,
2264                                                        &delalloc_start,
2265                                                        &delalloc_end,
2266                                                        128 * 1024 * 1024);
2267                         if (nr_delalloc == 0) {
2268                                 delalloc_start = delalloc_end + 1;
2269                                 continue;
2270                         }
2271                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2272                                                  delalloc_end, &page_started,
2273                                                  &nr_written);
2274                         /*
2275                          * delalloc_end is already one less than the total
2276                          * length, so we don't subtract one from
2277                          * PAGE_CACHE_SIZE
2278                          */
2279                         delalloc_to_write += (delalloc_end - delalloc_start +
2280                                               PAGE_CACHE_SIZE) >>
2281                                               PAGE_CACHE_SHIFT;
2282                         delalloc_start = delalloc_end + 1;
2283                 }
2284                 if (wbc->nr_to_write < delalloc_to_write) {
2285                         int thresh = 8192;
2286
2287                         if (delalloc_to_write < thresh * 2)
2288                                 thresh = delalloc_to_write;
2289                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2290                                                  thresh);
2291                 }
2292
2293                 /* did the fill delalloc function already unlock and start
2294                  * the IO?
2295                  */
2296                 if (page_started) {
2297                         ret = 0;
2298                         /*
2299                          * we've unlocked the page, so we can't update
2300                          * the mapping's writeback index, just update
2301                          * nr_to_write.
2302                          */
2303                         wbc->nr_to_write -= nr_written;
2304                         goto done_unlocked;
2305                 }
2306         }
2307         if (tree->ops && tree->ops->writepage_start_hook) {
2308                 ret = tree->ops->writepage_start_hook(page, start,
2309                                                       page_end);
2310                 if (ret == -EAGAIN) {
2311                         redirty_page_for_writepage(wbc, page);
2312                         update_nr_written(page, wbc, nr_written);
2313                         unlock_page(page);
2314                         ret = 0;
2315                         goto done_unlocked;
2316                 }
2317         }
2318
2319         /*
2320          * we don't want to touch the inode after unlocking the page,
2321          * so we update the mapping writeback index now
2322          */
2323         update_nr_written(page, wbc, nr_written + 1);
2324
2325         end = page_end;
2326         if (last_byte <= start) {
2327                 if (tree->ops && tree->ops->writepage_end_io_hook)
2328                         tree->ops->writepage_end_io_hook(page, start,
2329                                                          page_end, NULL, 1);
2330                 goto done;
2331         }
2332
2333         blocksize = inode->i_sb->s_blocksize;
2334
2335         while (cur <= end) {
2336                 if (cur >= last_byte) {
2337                         if (tree->ops && tree->ops->writepage_end_io_hook)
2338                                 tree->ops->writepage_end_io_hook(page, cur,
2339                                                          page_end, NULL, 1);
2340                         break;
2341                 }
2342                 em = epd->get_extent(inode, page, pg_offset, cur,
2343                                      end - cur + 1, 1);
2344                 if (IS_ERR(em) || !em) {
2345                         SetPageError(page);
2346                         break;
2347                 }
2348
2349                 extent_offset = cur - em->start;
2350                 BUG_ON(extent_map_end(em) <= cur);
2351                 BUG_ON(end < cur);
2352                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2353                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2354                 sector = (em->block_start + extent_offset) >> 9;
2355                 bdev = em->bdev;
2356                 block_start = em->block_start;
2357                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2358                 free_extent_map(em);
2359                 em = NULL;
2360
2361                 /*
2362                  * compressed and inline extents are written through other
2363                  * paths in the FS
2364                  */
2365                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2366                     block_start == EXTENT_MAP_INLINE) {
2367                         /*
2368                          * end_io notification does not happen here for
2369                          * compressed extents
2370                          */
2371                         if (!compressed && tree->ops &&
2372                             tree->ops->writepage_end_io_hook)
2373                                 tree->ops->writepage_end_io_hook(page, cur,
2374                                                          cur + iosize - 1,
2375                                                          NULL, 1);
2376                         else if (compressed) {
2377                                 /* we don't want to end_page_writeback on
2378                                  * a compressed extent.  this happens
2379                                  * elsewhere
2380                                  */
2381                                 nr++;
2382                         }
2383
2384                         cur += iosize;
2385                         pg_offset += iosize;
2386                         continue;
2387                 }
2388                 /* leave this out until we have a page_mkwrite call */
2389                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2390                                    EXTENT_DIRTY, 0, NULL)) {
2391                         cur = cur + iosize;
2392                         pg_offset += iosize;
2393                         continue;
2394                 }
2395
2396                 if (tree->ops && tree->ops->writepage_io_hook) {
2397                         ret = tree->ops->writepage_io_hook(page, cur,
2398                                                 cur + iosize - 1);
2399                 } else {
2400                         ret = 0;
2401                 }
2402                 if (ret) {
2403                         SetPageError(page);
2404                 } else {
2405                         unsigned long max_nr = end_index + 1;
2406
2407                         set_range_writeback(tree, cur, cur + iosize - 1);
2408                         if (!PageWriteback(page)) {
2409                                 printk(KERN_ERR "btrfs warning page %lu not "
2410                                        "writeback, cur %llu end %llu\n",
2411                                        page->index, (unsigned long long)cur,
2412                                        (unsigned long long)end);
2413                         }
2414
2415                         ret = submit_extent_page(write_flags, tree, page,
2416                                                  sector, iosize, pg_offset,
2417                                                  bdev, &epd->bio, max_nr,
2418                                                  end_bio_extent_writepage,
2419                                                  0, 0, 0);
2420                         if (ret)
2421                                 SetPageError(page);
2422                 }
2423                 cur = cur + iosize;
2424                 pg_offset += iosize;
2425                 nr++;
2426         }
2427 done:
2428         if (nr == 0) {
2429                 /* make sure the mapping tag for page dirty gets cleared */
2430                 set_page_writeback(page);
2431                 end_page_writeback(page);
2432         }
2433         unlock_page(page);
2434
2435 done_unlocked:
2436
2437         /* drop our reference on any cached states */
2438         free_extent_state(cached_state);
2439         return 0;
2440 }
2441
2442 /**
2443  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2444  * @mapping: address space structure to write
2445  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2446  * @writepage: function called for each page
2447  * @data: data passed to writepage function
2448  *
2449  * If a page is already under I/O, write_cache_pages() skips it, even
2450  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2451  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2452  * and msync() need to guarantee that all the data which was dirty at the time
2453  * the call was made get new I/O started against them.  If wbc->sync_mode is
2454  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2455  * existing IO to complete.
2456  */
2457 static int extent_write_cache_pages(struct extent_io_tree *tree,
2458                              struct address_space *mapping,
2459                              struct writeback_control *wbc,
2460                              writepage_t writepage, void *data,
2461                              void (*flush_fn)(void *))
2462 {
2463         int ret = 0;
2464         int done = 0;
2465         int nr_to_write_done = 0;
2466         struct pagevec pvec;
2467         int nr_pages;
2468         pgoff_t index;
2469         pgoff_t end;            /* Inclusive */
2470         int scanned = 0;
2471
2472         pagevec_init(&pvec, 0);
2473         if (wbc->range_cyclic) {
2474                 index = mapping->writeback_index; /* Start from prev offset */
2475                 end = -1;
2476         } else {
2477                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2478                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2479                 scanned = 1;
2480         }
2481 retry:
2482         while (!done && !nr_to_write_done && (index <= end) &&
2483                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2484                               PAGECACHE_TAG_DIRTY, min(end - index,
2485                                   (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2486                 unsigned i;
2487
2488                 scanned = 1;
2489                 for (i = 0; i < nr_pages; i++) {
2490                         struct page *page = pvec.pages[i];
2491
2492                         /*
2493                          * At this point we hold neither mapping->tree_lock nor
2494                          * lock on the page itself: the page may be truncated or
2495                          * invalidated (changing page->mapping to NULL), or even
2496                          * swizzled back from swapper_space to tmpfs file
2497                          * mapping
2498                          */
2499                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2500                                 tree->ops->write_cache_pages_lock_hook(page);
2501                         else
2502                                 lock_page(page);
2503
2504                         if (unlikely(page->mapping != mapping)) {
2505                                 unlock_page(page);
2506                                 continue;
2507                         }
2508
2509                         if (!wbc->range_cyclic && page->index > end) {
2510                                 done = 1;
2511                                 unlock_page(page);
2512                                 continue;
2513                         }
2514
2515                         if (wbc->sync_mode != WB_SYNC_NONE) {
2516                                 if (PageWriteback(page))
2517                                         flush_fn(data);
2518                                 wait_on_page_writeback(page);
2519                         }
2520
2521                         if (PageWriteback(page) ||
2522                             !clear_page_dirty_for_io(page)) {
2523                                 unlock_page(page);
2524                                 continue;
2525                         }
2526
2527                         ret = (*writepage)(page, wbc, data);
2528
2529                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2530                                 unlock_page(page);
2531                                 ret = 0;
2532                         }
2533                         if (ret)
2534                                 done = 1;
2535
2536                         /*
2537                          * the filesystem may choose to bump up nr_to_write.
2538                          * We have to make sure to honor the new nr_to_write
2539                          * at any time
2540                          */
2541                         nr_to_write_done = wbc->nr_to_write <= 0;
2542                 }
2543                 pagevec_release(&pvec);
2544                 cond_resched();
2545         }
2546         if (!scanned && !done) {
2547                 /*
2548                  * We hit the last page and there is more work to be done: wrap
2549                  * back to the start of the file
2550                  */
2551                 scanned = 1;
2552                 index = 0;
2553                 goto retry;
2554         }
2555         return ret;
2556 }
2557
2558 static void flush_epd_write_bio(struct extent_page_data *epd)
2559 {
2560         if (epd->bio) {
2561                 if (epd->sync_io)
2562                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2563                 else
2564                         submit_one_bio(WRITE, epd->bio, 0, 0);
2565                 epd->bio = NULL;
2566         }
2567 }
2568
2569 static noinline void flush_write_bio(void *data)
2570 {
2571         struct extent_page_data *epd = data;
2572         flush_epd_write_bio(epd);
2573 }
2574
2575 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2576                           get_extent_t *get_extent,
2577                           struct writeback_control *wbc)
2578 {
2579         int ret;
2580         struct address_space *mapping = page->mapping;
2581         struct extent_page_data epd = {
2582                 .bio = NULL,
2583                 .tree = tree,
2584                 .get_extent = get_extent,
2585                 .extent_locked = 0,
2586                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2587         };
2588         struct writeback_control wbc_writepages = {
2589                 .sync_mode      = wbc->sync_mode,
2590                 .older_than_this = NULL,
2591                 .nr_to_write    = 64,
2592                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2593                 .range_end      = (loff_t)-1,
2594         };
2595
2596         ret = __extent_writepage(page, wbc, &epd);
2597
2598         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2599                                  __extent_writepage, &epd, flush_write_bio);
2600         flush_epd_write_bio(&epd);
2601         return ret;
2602 }
2603
2604 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2605                               u64 start, u64 end, get_extent_t *get_extent,
2606                               int mode)
2607 {
2608         int ret = 0;
2609         struct address_space *mapping = inode->i_mapping;
2610         struct page *page;
2611         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2612                 PAGE_CACHE_SHIFT;
2613
2614         struct extent_page_data epd = {
2615                 .bio = NULL,
2616                 .tree = tree,
2617                 .get_extent = get_extent,
2618                 .extent_locked = 1,
2619                 .sync_io = mode == WB_SYNC_ALL,
2620         };
2621         struct writeback_control wbc_writepages = {
2622                 .sync_mode      = mode,
2623                 .older_than_this = NULL,
2624                 .nr_to_write    = nr_pages * 2,
2625                 .range_start    = start,
2626                 .range_end      = end + 1,
2627         };
2628
2629         while (start <= end) {
2630                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2631                 if (clear_page_dirty_for_io(page))
2632                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2633                 else {
2634                         if (tree->ops && tree->ops->writepage_end_io_hook)
2635                                 tree->ops->writepage_end_io_hook(page, start,
2636                                                  start + PAGE_CACHE_SIZE - 1,
2637                                                  NULL, 1);
2638                         unlock_page(page);
2639                 }
2640                 page_cache_release(page);
2641                 start += PAGE_CACHE_SIZE;
2642         }
2643
2644         flush_epd_write_bio(&epd);
2645         return ret;
2646 }
2647
2648 int extent_writepages(struct extent_io_tree *tree,
2649                       struct address_space *mapping,
2650                       get_extent_t *get_extent,
2651                       struct writeback_control *wbc)
2652 {
2653         int ret = 0;
2654         struct extent_page_data epd = {
2655                 .bio = NULL,
2656                 .tree = tree,
2657                 .get_extent = get_extent,
2658                 .extent_locked = 0,
2659                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2660         };
2661
2662         ret = extent_write_cache_pages(tree, mapping, wbc,
2663                                        __extent_writepage, &epd,
2664                                        flush_write_bio);
2665         flush_epd_write_bio(&epd);
2666         return ret;
2667 }
2668
2669 int extent_readpages(struct extent_io_tree *tree,
2670                      struct address_space *mapping,
2671                      struct list_head *pages, unsigned nr_pages,
2672                      get_extent_t get_extent)
2673 {
2674         struct bio *bio = NULL;
2675         unsigned page_idx;
2676         unsigned long bio_flags = 0;
2677
2678         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2679                 struct page *page = list_entry(pages->prev, struct page, lru);
2680
2681                 prefetchw(&page->flags);
2682                 list_del(&page->lru);
2683                 if (!add_to_page_cache_lru(page, mapping,
2684                                         page->index, GFP_KERNEL)) {
2685                         __extent_read_full_page(tree, page, get_extent,
2686                                                 &bio, 0, &bio_flags);
2687                 }
2688                 page_cache_release(page);
2689         }
2690         BUG_ON(!list_empty(pages));
2691         if (bio)
2692                 submit_one_bio(READ, bio, 0, bio_flags);
2693         return 0;
2694 }
2695
2696 /*
2697  * basic invalidatepage code, this waits on any locked or writeback
2698  * ranges corresponding to the page, and then deletes any extent state
2699  * records from the tree
2700  */
2701 int extent_invalidatepage(struct extent_io_tree *tree,
2702                           struct page *page, unsigned long offset)
2703 {
2704         struct extent_state *cached_state = NULL;
2705         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2706         u64 end = start + PAGE_CACHE_SIZE - 1;
2707         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2708
2709         start += (offset + blocksize - 1) & ~(blocksize - 1);
2710         if (start > end)
2711                 return 0;
2712
2713         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
2714         wait_on_page_writeback(page);
2715         clear_extent_bit(tree, start, end,
2716                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2717                          EXTENT_DO_ACCOUNTING,
2718                          1, 1, &cached_state, GFP_NOFS);
2719         return 0;
2720 }
2721
2722 /*
2723  * simple commit_write call, set_range_dirty is used to mark both
2724  * the pages and the extent records as dirty
2725  */
2726 int extent_commit_write(struct extent_io_tree *tree,
2727                         struct inode *inode, struct page *page,
2728                         unsigned from, unsigned to)
2729 {
2730         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2731
2732         set_page_extent_mapped(page);
2733         set_page_dirty(page);
2734
2735         if (pos > inode->i_size) {
2736                 i_size_write(inode, pos);
2737                 mark_inode_dirty(inode);
2738         }
2739         return 0;
2740 }
2741
2742 int extent_prepare_write(struct extent_io_tree *tree,
2743                          struct inode *inode, struct page *page,
2744                          unsigned from, unsigned to, get_extent_t *get_extent)
2745 {
2746         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2747         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2748         u64 block_start;
2749         u64 orig_block_start;
2750         u64 block_end;
2751         u64 cur_end;
2752         struct extent_map *em;
2753         unsigned blocksize = 1 << inode->i_blkbits;
2754         size_t page_offset = 0;
2755         size_t block_off_start;
2756         size_t block_off_end;
2757         int err = 0;
2758         int iocount = 0;
2759         int ret = 0;
2760         int isnew;
2761
2762         set_page_extent_mapped(page);
2763
2764         block_start = (page_start + from) & ~((u64)blocksize - 1);
2765         block_end = (page_start + to - 1) | (blocksize - 1);
2766         orig_block_start = block_start;
2767
2768         lock_extent(tree, page_start, page_end, GFP_NOFS);
2769         while (block_start <= block_end) {
2770                 em = get_extent(inode, page, page_offset, block_start,
2771                                 block_end - block_start + 1, 1);
2772                 if (IS_ERR(em) || !em)
2773                         goto err;
2774
2775                 cur_end = min(block_end, extent_map_end(em) - 1);
2776                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2777                 block_off_end = block_off_start + blocksize;
2778                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2779
2780                 if (!PageUptodate(page) && isnew &&
2781                     (block_off_end > to || block_off_start < from)) {
2782                         void *kaddr;
2783
2784                         kaddr = kmap_atomic(page, KM_USER0);
2785                         if (block_off_end > to)
2786                                 memset(kaddr + to, 0, block_off_end - to);
2787                         if (block_off_start < from)
2788                                 memset(kaddr + block_off_start, 0,
2789                                        from - block_off_start);
2790                         flush_dcache_page(page);
2791                         kunmap_atomic(kaddr, KM_USER0);
2792                 }
2793                 if ((em->block_start != EXTENT_MAP_HOLE &&
2794                      em->block_start != EXTENT_MAP_INLINE) &&
2795                     !isnew && !PageUptodate(page) &&
2796                     (block_off_end > to || block_off_start < from) &&
2797                     !test_range_bit(tree, block_start, cur_end,
2798                                     EXTENT_UPTODATE, 1, NULL)) {
2799                         u64 sector;
2800                         u64 extent_offset = block_start - em->start;
2801                         size_t iosize;
2802                         sector = (em->block_start + extent_offset) >> 9;
2803                         iosize = (cur_end - block_start + blocksize) &
2804                                 ~((u64)blocksize - 1);
2805                         /*
2806                          * we've already got the extent locked, but we
2807                          * need to split the state such that our end_bio
2808                          * handler can clear the lock.
2809                          */
2810                         set_extent_bit(tree, block_start,
2811                                        block_start + iosize - 1,
2812                                        EXTENT_LOCKED, 0, NULL, NULL, GFP_NOFS);
2813                         ret = submit_extent_page(READ, tree, page,
2814                                          sector, iosize, page_offset, em->bdev,
2815                                          NULL, 1,
2816                                          end_bio_extent_preparewrite, 0,
2817                                          0, 0);
2818                         if (ret && !err)
2819                                 err = ret;
2820                         iocount++;
2821                         block_start = block_start + iosize;
2822                 } else {
2823                         struct extent_state *cached = NULL;
2824
2825                         set_extent_uptodate(tree, block_start, cur_end, &cached,
2826                                             GFP_NOFS);
2827                         unlock_extent_cached(tree, block_start, cur_end,
2828                                              &cached, GFP_NOFS);
2829                         block_start = cur_end + 1;
2830                 }
2831                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2832                 free_extent_map(em);
2833         }
2834         if (iocount) {
2835                 wait_extent_bit(tree, orig_block_start,
2836                                 block_end, EXTENT_LOCKED);
2837         }
2838         check_page_uptodate(tree, page);
2839 err:
2840         /* FIXME, zero out newly allocated blocks on error */
2841         return err;
2842 }
2843
2844 /*
2845  * a helper for releasepage, this tests for areas of the page that
2846  * are locked or under IO and drops the related state bits if it is safe
2847  * to drop the page.
2848  */
2849 int try_release_extent_state(struct extent_map_tree *map,
2850                              struct extent_io_tree *tree, struct page *page,
2851                              gfp_t mask)
2852 {
2853         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2854         u64 end = start + PAGE_CACHE_SIZE - 1;
2855         int ret = 1;
2856
2857         if (test_range_bit(tree, start, end,
2858                            EXTENT_IOBITS, 0, NULL))
2859                 ret = 0;
2860         else {
2861                 if ((mask & GFP_NOFS) == GFP_NOFS)
2862                         mask = GFP_NOFS;
2863                 /*
2864                  * at this point we can safely clear everything except the
2865                  * locked bit and the nodatasum bit
2866                  */
2867                 ret = clear_extent_bit(tree, start, end,
2868                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2869                                  0, 0, NULL, mask);
2870
2871                 /* if clear_extent_bit failed for enomem reasons,
2872                  * we can't allow the release to continue.
2873                  */
2874                 if (ret < 0)
2875                         ret = 0;
2876                 else
2877                         ret = 1;
2878         }
2879         return ret;
2880 }
2881
2882 /*
2883  * a helper for releasepage.  As long as there are no locked extents
2884  * in the range corresponding to the page, both state records and extent
2885  * map records are removed
2886  */
2887 int try_release_extent_mapping(struct extent_map_tree *map,
2888                                struct extent_io_tree *tree, struct page *page,
2889                                gfp_t mask)
2890 {
2891         struct extent_map *em;
2892         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2893         u64 end = start + PAGE_CACHE_SIZE - 1;
2894
2895         if ((mask & __GFP_WAIT) &&
2896             page->mapping->host->i_size > 16 * 1024 * 1024) {
2897                 u64 len;
2898                 while (start <= end) {
2899                         len = end - start + 1;
2900                         write_lock(&map->lock);
2901                         em = lookup_extent_mapping(map, start, len);
2902                         if (!em || IS_ERR(em)) {
2903                                 write_unlock(&map->lock);
2904                                 break;
2905                         }
2906                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2907                             em->start != start) {
2908                                 write_unlock(&map->lock);
2909                                 free_extent_map(em);
2910                                 break;
2911                         }
2912                         if (!test_range_bit(tree, em->start,
2913                                             extent_map_end(em) - 1,
2914                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
2915                                             0, NULL)) {
2916                                 remove_extent_mapping(map, em);
2917                                 /* once for the rb tree */
2918                                 free_extent_map(em);
2919                         }
2920                         start = extent_map_end(em);
2921                         write_unlock(&map->lock);
2922
2923                         /* once for us */
2924                         free_extent_map(em);
2925                 }
2926         }
2927         return try_release_extent_state(map, tree, page, mask);
2928 }
2929
2930 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2931                 get_extent_t *get_extent)
2932 {
2933         struct inode *inode = mapping->host;
2934         struct extent_state *cached_state = NULL;
2935         u64 start = iblock << inode->i_blkbits;
2936         sector_t sector = 0;
2937         size_t blksize = (1 << inode->i_blkbits);
2938         struct extent_map *em;
2939
2940         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2941                          0, &cached_state, GFP_NOFS);
2942         em = get_extent(inode, NULL, 0, start, blksize, 0);
2943         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start,
2944                              start + blksize - 1, &cached_state, GFP_NOFS);
2945         if (!em || IS_ERR(em))
2946                 return 0;
2947
2948         if (em->block_start > EXTENT_MAP_LAST_BYTE)
2949                 goto out;
2950
2951         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2952 out:
2953         free_extent_map(em);
2954         return sector;
2955 }
2956
2957 /*
2958  * helper function for fiemap, which doesn't want to see any holes.
2959  * This maps until we find something past 'last'
2960  */
2961 static struct extent_map *get_extent_skip_holes(struct inode *inode,
2962                                                 u64 offset,
2963                                                 u64 last,
2964                                                 get_extent_t *get_extent)
2965 {
2966         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
2967         struct extent_map *em;
2968         u64 len;
2969
2970         if (offset >= last)
2971                 return NULL;
2972
2973         while(1) {
2974                 len = last - offset;
2975                 if (len == 0)
2976                         break;
2977                 len = (len + sectorsize - 1) & ~(sectorsize - 1);
2978                 em = get_extent(inode, NULL, 0, offset, len, 0);
2979                 if (!em || IS_ERR(em))
2980                         return em;
2981
2982                 /* if this isn't a hole return it */
2983                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
2984                     em->block_start != EXTENT_MAP_HOLE) {
2985                         return em;
2986                 }
2987
2988                 /* this is a hole, advance to the next extent */
2989                 offset = extent_map_end(em);
2990                 free_extent_map(em);
2991                 if (offset >= last)
2992                         break;
2993         }
2994         return NULL;
2995 }
2996
2997 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2998                 __u64 start, __u64 len, get_extent_t *get_extent)
2999 {
3000         int ret = 0;
3001         u64 off = start;
3002         u64 max = start + len;
3003         u32 flags = 0;
3004         u32 found_type;
3005         u64 last;
3006         u64 last_for_get_extent = 0;
3007         u64 disko = 0;
3008         u64 isize = i_size_read(inode);
3009         struct btrfs_key found_key;
3010         struct extent_map *em = NULL;
3011         struct extent_state *cached_state = NULL;
3012         struct btrfs_path *path;
3013         struct btrfs_file_extent_item *item;
3014         int end = 0;
3015         u64 em_start = 0;
3016         u64 em_len = 0;
3017         u64 em_end = 0;
3018         unsigned long emflags;
3019
3020         if (len == 0)
3021                 return -EINVAL;
3022
3023         path = btrfs_alloc_path();
3024         if (!path)
3025                 return -ENOMEM;
3026         path->leave_spinning = 1;
3027
3028         /*
3029          * lookup the last file extent.  We're not using i_size here
3030          * because there might be preallocation past i_size
3031          */
3032         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3033                                        path, inode->i_ino, -1, 0);
3034         if (ret < 0) {
3035                 btrfs_free_path(path);
3036                 return ret;
3037         }
3038         WARN_ON(!ret);
3039         path->slots[0]--;
3040         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3041                               struct btrfs_file_extent_item);
3042         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3043         found_type = btrfs_key_type(&found_key);
3044
3045         /* No extents, but there might be delalloc bits */
3046         if (found_key.objectid != inode->i_ino ||
3047             found_type != BTRFS_EXTENT_DATA_KEY) {
3048                 /* have to trust i_size as the end */
3049                 last = (u64)-1;
3050                 last_for_get_extent = isize;
3051         } else {
3052                 /*
3053                  * remember the start of the last extent.  There are a
3054                  * bunch of different factors that go into the length of the
3055                  * extent, so its much less complex to remember where it started
3056                  */
3057                 last = found_key.offset;
3058                 last_for_get_extent = last + 1;
3059         }
3060         btrfs_free_path(path);
3061
3062         /*
3063          * we might have some extents allocated but more delalloc past those
3064          * extents.  so, we trust isize unless the start of the last extent is
3065          * beyond isize
3066          */
3067         if (last < isize) {
3068                 last = (u64)-1;
3069                 last_for_get_extent = isize;
3070         }
3071
3072         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3073                          &cached_state, GFP_NOFS);
3074
3075         em = get_extent_skip_holes(inode, off, last_for_get_extent,
3076                                    get_extent);
3077         if (!em)
3078                 goto out;
3079         if (IS_ERR(em)) {
3080                 ret = PTR_ERR(em);
3081                 goto out;
3082         }
3083
3084         while (!end) {
3085                 u64 offset_in_extent;
3086
3087                 /* break if the extent we found is outside the range */
3088                 if (em->start >= max || extent_map_end(em) < off)
3089                         break;
3090
3091                 /*
3092                  * get_extent may return an extent that starts before our
3093                  * requested range.  We have to make sure the ranges
3094                  * we return to fiemap always move forward and don't
3095                  * overlap, so adjust the offsets here
3096                  */
3097                 em_start = max(em->start, off);
3098
3099                 /*
3100                  * record the offset from the start of the extent
3101                  * for adjusting the disk offset below
3102                  */
3103                 offset_in_extent = em_start - em->start;
3104                 em_end = extent_map_end(em);
3105                 em_len = em_end - em_start;
3106                 emflags = em->flags;
3107                 disko = 0;
3108                 flags = 0;
3109
3110                 /*
3111                  * bump off for our next call to get_extent
3112                  */
3113                 off = extent_map_end(em);
3114                 if (off >= max)
3115                         end = 1;
3116
3117                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3118                         end = 1;
3119                         flags |= FIEMAP_EXTENT_LAST;
3120                 } else if (em->block_start == EXTENT_MAP_INLINE) {
3121                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
3122                                   FIEMAP_EXTENT_NOT_ALIGNED);
3123                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3124                         flags |= (FIEMAP_EXTENT_DELALLOC |
3125                                   FIEMAP_EXTENT_UNKNOWN);
3126                 } else {
3127                         disko = em->block_start + offset_in_extent;
3128                 }
3129                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3130                         flags |= FIEMAP_EXTENT_ENCODED;
3131
3132                 free_extent_map(em);
3133                 em = NULL;
3134                 if ((em_start >= last) || em_len == (u64)-1 ||
3135                    (last == (u64)-1 && isize <= em_end)) {
3136                         flags |= FIEMAP_EXTENT_LAST;
3137                         end = 1;
3138                 }
3139
3140                 /* now scan forward to see if this is really the last extent. */
3141                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
3142                                            get_extent);
3143                 if (IS_ERR(em)) {
3144                         ret = PTR_ERR(em);
3145                         goto out;
3146                 }
3147                 if (!em) {
3148                         flags |= FIEMAP_EXTENT_LAST;
3149                         end = 1;
3150                 }
3151                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3152                                               em_len, flags);
3153                 if (ret)
3154                         goto out_free;
3155         }
3156 out_free:
3157         free_extent_map(em);
3158 out:
3159         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3160                              &cached_state, GFP_NOFS);
3161         return ret;
3162 }
3163
3164 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
3165                                               unsigned long i)
3166 {
3167         struct page *p;
3168         struct address_space *mapping;
3169
3170         if (i == 0)
3171                 return eb->first_page;
3172         i += eb->start >> PAGE_CACHE_SHIFT;
3173         mapping = eb->first_page->mapping;
3174         if (!mapping)
3175                 return NULL;
3176
3177         /*
3178          * extent_buffer_page is only called after pinning the page
3179          * by increasing the reference count.  So we know the page must
3180          * be in the radix tree.
3181          */
3182         rcu_read_lock();
3183         p = radix_tree_lookup(&mapping->page_tree, i);
3184         rcu_read_unlock();
3185
3186         return p;
3187 }
3188
3189 static inline unsigned long num_extent_pages(u64 start, u64 len)
3190 {
3191         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3192                 (start >> PAGE_CACHE_SHIFT);
3193 }
3194
3195 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3196                                                    u64 start,
3197                                                    unsigned long len,
3198                                                    gfp_t mask)
3199 {
3200         struct extent_buffer *eb = NULL;
3201 #if LEAK_DEBUG
3202         unsigned long flags;
3203 #endif
3204
3205         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3206         if (eb == NULL)
3207                 return NULL;
3208         eb->start = start;
3209         eb->len = len;
3210         spin_lock_init(&eb->lock);
3211         init_waitqueue_head(&eb->lock_wq);
3212
3213 #if LEAK_DEBUG
3214         spin_lock_irqsave(&leak_lock, flags);
3215         list_add(&eb->leak_list, &buffers);
3216         spin_unlock_irqrestore(&leak_lock, flags);
3217 #endif
3218         atomic_set(&eb->refs, 1);
3219
3220         return eb;
3221 }
3222
3223 static void __free_extent_buffer(struct extent_buffer *eb)
3224 {
3225 #if LEAK_DEBUG
3226         unsigned long flags;
3227         spin_lock_irqsave(&leak_lock, flags);
3228         list_del(&eb->leak_list);
3229         spin_unlock_irqrestore(&leak_lock, flags);
3230 #endif
3231         kmem_cache_free(extent_buffer_cache, eb);
3232 }
3233
3234 /*
3235  * Helper for releasing extent buffer page.
3236  */
3237 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3238                                                 unsigned long start_idx)
3239 {
3240         unsigned long index;
3241         struct page *page;
3242
3243         if (!eb->first_page)
3244                 return;
3245
3246         index = num_extent_pages(eb->start, eb->len);
3247         if (start_idx >= index)
3248                 return;
3249
3250         do {
3251                 index--;
3252                 page = extent_buffer_page(eb, index);
3253                 if (page)
3254                         page_cache_release(page);
3255         } while (index != start_idx);
3256 }
3257
3258 /*
3259  * Helper for releasing the extent buffer.
3260  */
3261 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3262 {
3263         btrfs_release_extent_buffer_page(eb, 0);
3264         __free_extent_buffer(eb);
3265 }
3266
3267 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3268                                           u64 start, unsigned long len,
3269                                           struct page *page0,
3270                                           gfp_t mask)
3271 {
3272         unsigned long num_pages = num_extent_pages(start, len);
3273         unsigned long i;
3274         unsigned long index = start >> PAGE_CACHE_SHIFT;
3275         struct extent_buffer *eb;
3276         struct extent_buffer *exists = NULL;
3277         struct page *p;
3278         struct address_space *mapping = tree->mapping;
3279         int uptodate = 1;
3280         int ret;
3281
3282         rcu_read_lock();
3283         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3284         if (eb && atomic_inc_not_zero(&eb->refs)) {
3285                 rcu_read_unlock();
3286                 mark_page_accessed(eb->first_page);
3287                 return eb;
3288         }
3289         rcu_read_unlock();
3290
3291         eb = __alloc_extent_buffer(tree, start, len, mask);
3292         if (!eb)
3293                 return NULL;
3294
3295         if (page0) {
3296                 eb->first_page = page0;
3297                 i = 1;
3298                 index++;
3299                 page_cache_get(page0);
3300                 mark_page_accessed(page0);
3301                 set_page_extent_mapped(page0);
3302                 set_page_extent_head(page0, len);
3303                 uptodate = PageUptodate(page0);
3304         } else {
3305                 i = 0;
3306         }
3307         for (; i < num_pages; i++, index++) {
3308                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
3309                 if (!p) {
3310                         WARN_ON(1);
3311                         goto free_eb;
3312                 }
3313                 set_page_extent_mapped(p);
3314                 mark_page_accessed(p);
3315                 if (i == 0) {
3316                         eb->first_page = p;
3317                         set_page_extent_head(p, len);
3318                 } else {
3319                         set_page_private(p, EXTENT_PAGE_PRIVATE);
3320                 }
3321                 if (!PageUptodate(p))
3322                         uptodate = 0;
3323
3324                 /*
3325                  * see below about how we avoid a nasty race with release page
3326                  * and why we unlock later
3327                  */
3328                 if (i != 0)
3329                         unlock_page(p);
3330         }
3331         if (uptodate)
3332                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3333
3334         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
3335         if (ret)
3336                 goto free_eb;
3337
3338         spin_lock(&tree->buffer_lock);
3339         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
3340         if (ret == -EEXIST) {
3341                 exists = radix_tree_lookup(&tree->buffer,
3342                                                 start >> PAGE_CACHE_SHIFT);
3343                 /* add one reference for the caller */
3344                 atomic_inc(&exists->refs);
3345                 spin_unlock(&tree->buffer_lock);
3346                 radix_tree_preload_end();
3347                 goto free_eb;
3348         }
3349         /* add one reference for the tree */
3350         atomic_inc(&eb->refs);
3351         spin_unlock(&tree->buffer_lock);
3352         radix_tree_preload_end();
3353
3354         /*
3355          * there is a race where release page may have
3356          * tried to find this extent buffer in the radix
3357          * but failed.  It will tell the VM it is safe to
3358          * reclaim the, and it will clear the page private bit.
3359          * We must make sure to set the page private bit properly
3360          * after the extent buffer is in the radix tree so
3361          * it doesn't get lost
3362          */
3363         set_page_extent_mapped(eb->first_page);
3364         set_page_extent_head(eb->first_page, eb->len);
3365         if (!page0)
3366                 unlock_page(eb->first_page);
3367         return eb;
3368
3369 free_eb:
3370         if (eb->first_page && !page0)
3371                 unlock_page(eb->first_page);
3372
3373         if (!atomic_dec_and_test(&eb->refs))
3374                 return exists;
3375         btrfs_release_extent_buffer(eb);
3376         return exists;
3377 }
3378
3379 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3380                                          u64 start, unsigned long len,
3381                                           gfp_t mask)
3382 {
3383         struct extent_buffer *eb;
3384
3385         rcu_read_lock();
3386         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3387         if (eb && atomic_inc_not_zero(&eb->refs)) {
3388                 rcu_read_unlock();
3389                 mark_page_accessed(eb->first_page);
3390                 return eb;
3391         }
3392         rcu_read_unlock();
3393
3394         return NULL;
3395 }
3396
3397 void free_extent_buffer(struct extent_buffer *eb)
3398 {
3399         if (!eb)
3400                 return;
3401
3402         if (!atomic_dec_and_test(&eb->refs))
3403                 return;
3404
3405         WARN_ON(1);
3406 }
3407
3408 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3409                               struct extent_buffer *eb)
3410 {
3411         unsigned long i;
3412         unsigned long num_pages;
3413         struct page *page;
3414
3415         num_pages = num_extent_pages(eb->start, eb->len);
3416
3417         for (i = 0; i < num_pages; i++) {
3418                 page = extent_buffer_page(eb, i);
3419                 if (!PageDirty(page))
3420                         continue;
3421
3422                 lock_page(page);
3423                 WARN_ON(!PagePrivate(page));
3424
3425                 set_page_extent_mapped(page);
3426                 if (i == 0)
3427                         set_page_extent_head(page, eb->len);
3428
3429                 clear_page_dirty_for_io(page);
3430                 spin_lock_irq(&page->mapping->tree_lock);
3431                 if (!PageDirty(page)) {
3432                         radix_tree_tag_clear(&page->mapping->page_tree,
3433                                                 page_index(page),
3434                                                 PAGECACHE_TAG_DIRTY);
3435                 }
3436                 spin_unlock_irq(&page->mapping->tree_lock);
3437                 unlock_page(page);
3438         }
3439         return 0;
3440 }
3441
3442 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
3443                                     struct extent_buffer *eb)
3444 {
3445         return wait_on_extent_writeback(tree, eb->start,
3446                                         eb->start + eb->len - 1);
3447 }
3448
3449 int set_extent_buffer_dirty(struct extent_io_tree *tree,
3450                              struct extent_buffer *eb)
3451 {
3452         unsigned long i;
3453         unsigned long num_pages;
3454         int was_dirty = 0;
3455
3456         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3457         num_pages = num_extent_pages(eb->start, eb->len);
3458         for (i = 0; i < num_pages; i++)
3459                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3460         return was_dirty;
3461 }
3462
3463 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3464                                 struct extent_buffer *eb,
3465                                 struct extent_state **cached_state)
3466 {
3467         unsigned long i;
3468         struct page *page;
3469         unsigned long num_pages;
3470
3471         num_pages = num_extent_pages(eb->start, eb->len);
3472         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3473
3474         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3475                               cached_state, GFP_NOFS);
3476         for (i = 0; i < num_pages; i++) {
3477                 page = extent_buffer_page(eb, i);
3478                 if (page)
3479                         ClearPageUptodate(page);
3480         }
3481         return 0;
3482 }
3483
3484 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3485                                 struct extent_buffer *eb)
3486 {
3487         unsigned long i;
3488         struct page *page;
3489         unsigned long num_pages;
3490
3491         num_pages = num_extent_pages(eb->start, eb->len);
3492
3493         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3494                             NULL, GFP_NOFS);
3495         for (i = 0; i < num_pages; i++) {
3496                 page = extent_buffer_page(eb, i);
3497                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3498                     ((i == num_pages - 1) &&
3499                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3500                         check_page_uptodate(tree, page);
3501                         continue;
3502                 }
3503                 SetPageUptodate(page);
3504         }
3505         return 0;
3506 }
3507
3508 int extent_range_uptodate(struct extent_io_tree *tree,
3509                           u64 start, u64 end)
3510 {
3511         struct page *page;
3512         int ret;
3513         int pg_uptodate = 1;
3514         int uptodate;
3515         unsigned long index;
3516
3517         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL);
3518         if (ret)
3519                 return 1;
3520         while (start <= end) {
3521                 index = start >> PAGE_CACHE_SHIFT;
3522                 page = find_get_page(tree->mapping, index);
3523                 uptodate = PageUptodate(page);
3524                 page_cache_release(page);
3525                 if (!uptodate) {
3526                         pg_uptodate = 0;
3527                         break;
3528                 }
3529                 start += PAGE_CACHE_SIZE;
3530         }
3531         return pg_uptodate;
3532 }
3533
3534 int extent_buffer_uptodate(struct extent_io_tree *tree,
3535                            struct extent_buffer *eb,
3536                            struct extent_state *cached_state)
3537 {
3538         int ret = 0;
3539         unsigned long num_pages;
3540         unsigned long i;
3541         struct page *page;
3542         int pg_uptodate = 1;
3543
3544         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3545                 return 1;
3546
3547         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3548                            EXTENT_UPTODATE, 1, cached_state);
3549         if (ret)
3550                 return ret;
3551
3552         num_pages = num_extent_pages(eb->start, eb->len);
3553         for (i = 0; i < num_pages; i++) {
3554                 page = extent_buffer_page(eb, i);
3555                 if (!PageUptodate(page)) {
3556                         pg_uptodate = 0;
3557                         break;
3558                 }
3559         }
3560         return pg_uptodate;
3561 }
3562
3563 int read_extent_buffer_pages(struct extent_io_tree *tree,
3564                              struct extent_buffer *eb,
3565                              u64 start, int wait,
3566                              get_extent_t *get_extent, int mirror_num)
3567 {
3568         unsigned long i;
3569         unsigned long start_i;
3570         struct page *page;
3571         int err;
3572         int ret = 0;
3573         int locked_pages = 0;
3574         int all_uptodate = 1;
3575         int inc_all_pages = 0;
3576         unsigned long num_pages;
3577         struct bio *bio = NULL;
3578         unsigned long bio_flags = 0;
3579
3580         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3581                 return 0;
3582
3583         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3584                            EXTENT_UPTODATE, 1, NULL)) {
3585                 return 0;
3586         }
3587
3588         if (start) {
3589                 WARN_ON(start < eb->start);
3590                 start_i = (start >> PAGE_CACHE_SHIFT) -
3591                         (eb->start >> PAGE_CACHE_SHIFT);
3592         } else {
3593                 start_i = 0;
3594         }
3595
3596         num_pages = num_extent_pages(eb->start, eb->len);
3597         for (i = start_i; i < num_pages; i++) {
3598                 page = extent_buffer_page(eb, i);
3599                 if (!wait) {
3600                         if (!trylock_page(page))
3601                                 goto unlock_exit;
3602                 } else {
3603                         lock_page(page);
3604                 }
3605                 locked_pages++;
3606                 if (!PageUptodate(page))
3607                         all_uptodate = 0;
3608         }
3609         if (all_uptodate) {
3610                 if (start_i == 0)
3611                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3612                 goto unlock_exit;
3613         }
3614
3615         for (i = start_i; i < num_pages; i++) {
3616                 page = extent_buffer_page(eb, i);
3617
3618                 WARN_ON(!PagePrivate(page));
3619
3620                 set_page_extent_mapped(page);
3621                 if (i == 0)
3622                         set_page_extent_head(page, eb->len);
3623
3624                 if (inc_all_pages)
3625                         page_cache_get(page);
3626                 if (!PageUptodate(page)) {
3627                         if (start_i == 0)
3628                                 inc_all_pages = 1;
3629                         ClearPageError(page);
3630                         err = __extent_read_full_page(tree, page,
3631                                                       get_extent, &bio,
3632                                                       mirror_num, &bio_flags);
3633                         if (err)
3634                                 ret = err;
3635                 } else {
3636                         unlock_page(page);
3637                 }
3638         }
3639
3640         if (bio)
3641                 submit_one_bio(READ, bio, mirror_num, bio_flags);
3642
3643         if (ret || !wait)
3644                 return ret;
3645
3646         for (i = start_i; i < num_pages; i++) {
3647                 page = extent_buffer_page(eb, i);
3648                 wait_on_page_locked(page);
3649                 if (!PageUptodate(page))
3650                         ret = -EIO;
3651         }
3652
3653         if (!ret)
3654                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3655         return ret;
3656
3657 unlock_exit:
3658         i = start_i;
3659         while (locked_pages > 0) {
3660                 page = extent_buffer_page(eb, i);
3661                 i++;
3662                 unlock_page(page);
3663                 locked_pages--;
3664         }
3665         return ret;
3666 }
3667
3668 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3669                         unsigned long start,
3670                         unsigned long len)
3671 {
3672         size_t cur;
3673         size_t offset;
3674         struct page *page;
3675         char *kaddr;
3676         char *dst = (char *)dstv;
3677         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3678         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3679
3680         WARN_ON(start > eb->len);
3681         WARN_ON(start + len > eb->start + eb->len);
3682
3683         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3684
3685         while (len > 0) {
3686                 page = extent_buffer_page(eb, i);
3687
3688                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3689                 kaddr = kmap_atomic(page, KM_USER1);
3690                 memcpy(dst, kaddr + offset, cur);
3691                 kunmap_atomic(kaddr, KM_USER1);
3692
3693                 dst += cur;
3694                 len -= cur;
3695                 offset = 0;
3696                 i++;
3697         }
3698 }
3699
3700 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3701                                unsigned long min_len, char **token, char **map,
3702                                unsigned long *map_start,
3703                                unsigned long *map_len, int km)
3704 {
3705         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3706         char *kaddr;
3707         struct page *p;
3708         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3709         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3710         unsigned long end_i = (start_offset + start + min_len - 1) >>
3711                 PAGE_CACHE_SHIFT;
3712
3713         if (i != end_i)
3714                 return -EINVAL;
3715
3716         if (i == 0) {
3717                 offset = start_offset;
3718                 *map_start = 0;
3719         } else {
3720                 offset = 0;
3721                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3722         }
3723
3724         if (start + min_len > eb->len) {
3725                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3726                        "wanted %lu %lu\n", (unsigned long long)eb->start,
3727                        eb->len, start, min_len);
3728                 WARN_ON(1);
3729                 return -EINVAL;
3730         }
3731
3732         p = extent_buffer_page(eb, i);
3733         kaddr = kmap_atomic(p, km);
3734         *token = kaddr;
3735         *map = kaddr + offset;
3736         *map_len = PAGE_CACHE_SIZE - offset;
3737         return 0;
3738 }
3739
3740 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3741                       unsigned long min_len,
3742                       char **token, char **map,
3743                       unsigned long *map_start,
3744                       unsigned long *map_len, int km)
3745 {
3746         int err;
3747         int save = 0;
3748         if (eb->map_token) {
3749                 unmap_extent_buffer(eb, eb->map_token, km);
3750                 eb->map_token = NULL;
3751                 save = 1;
3752         }
3753         err = map_private_extent_buffer(eb, start, min_len, token, map,
3754                                        map_start, map_len, km);
3755         if (!err && save) {
3756                 eb->map_token = *token;
3757                 eb->kaddr = *map;
3758                 eb->map_start = *map_start;
3759                 eb->map_len = *map_len;
3760         }
3761         return err;
3762 }
3763
3764 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3765 {
3766         kunmap_atomic(token, km);
3767 }
3768
3769 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3770                           unsigned long start,
3771                           unsigned long len)
3772 {
3773         size_t cur;
3774         size_t offset;
3775         struct page *page;
3776         char *kaddr;
3777         char *ptr = (char *)ptrv;
3778         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3779         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3780         int ret = 0;
3781
3782         WARN_ON(start > eb->len);
3783         WARN_ON(start + len > eb->start + eb->len);
3784
3785         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3786
3787         while (len > 0) {
3788                 page = extent_buffer_page(eb, i);
3789
3790                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3791
3792                 kaddr = kmap_atomic(page, KM_USER0);
3793                 ret = memcmp(ptr, kaddr + offset, cur);
3794                 kunmap_atomic(kaddr, KM_USER0);
3795                 if (ret)
3796                         break;
3797
3798                 ptr += cur;
3799                 len -= cur;
3800                 offset = 0;
3801                 i++;
3802         }
3803         return ret;
3804 }
3805
3806 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3807                          unsigned long start, unsigned long len)
3808 {
3809         size_t cur;
3810         size_t offset;
3811         struct page *page;
3812         char *kaddr;
3813         char *src = (char *)srcv;
3814         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3815         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3816
3817         WARN_ON(start > eb->len);
3818         WARN_ON(start + len > eb->start + eb->len);
3819
3820         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3821
3822         while (len > 0) {
3823                 page = extent_buffer_page(eb, i);
3824                 WARN_ON(!PageUptodate(page));
3825
3826                 cur = min(len, PAGE_CACHE_SIZE - offset);
3827                 kaddr = kmap_atomic(page, KM_USER1);
3828                 memcpy(kaddr + offset, src, cur);
3829                 kunmap_atomic(kaddr, KM_USER1);
3830
3831                 src += cur;
3832                 len -= cur;
3833                 offset = 0;
3834                 i++;
3835         }
3836 }
3837
3838 void memset_extent_buffer(struct extent_buffer *eb, char c,
3839                           unsigned long start, unsigned long len)
3840 {
3841         size_t cur;
3842         size_t offset;
3843         struct page *page;
3844         char *kaddr;
3845         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3846         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3847
3848         WARN_ON(start > eb->len);
3849         WARN_ON(start + len > eb->start + eb->len);
3850
3851         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3852
3853         while (len > 0) {
3854                 page = extent_buffer_page(eb, i);
3855                 WARN_ON(!PageUptodate(page));
3856
3857                 cur = min(len, PAGE_CACHE_SIZE - offset);
3858                 kaddr = kmap_atomic(page, KM_USER0);
3859                 memset(kaddr + offset, c, cur);
3860                 kunmap_atomic(kaddr, KM_USER0);
3861
3862                 len -= cur;
3863                 offset = 0;
3864                 i++;
3865         }
3866 }
3867
3868 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3869                         unsigned long dst_offset, unsigned long src_offset,
3870                         unsigned long len)
3871 {
3872         u64 dst_len = dst->len;
3873         size_t cur;
3874         size_t offset;
3875         struct page *page;
3876         char *kaddr;
3877         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3878         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3879
3880         WARN_ON(src->len != dst_len);
3881
3882         offset = (start_offset + dst_offset) &
3883                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3884
3885         while (len > 0) {
3886                 page = extent_buffer_page(dst, i);
3887                 WARN_ON(!PageUptodate(page));
3888
3889                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3890
3891                 kaddr = kmap_atomic(page, KM_USER0);
3892                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3893                 kunmap_atomic(kaddr, KM_USER0);
3894
3895                 src_offset += cur;
3896                 len -= cur;
3897                 offset = 0;
3898                 i++;
3899         }
3900 }
3901
3902 static void move_pages(struct page *dst_page, struct page *src_page,
3903                        unsigned long dst_off, unsigned long src_off,
3904                        unsigned long len)
3905 {
3906         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3907         if (dst_page == src_page) {
3908                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3909         } else {
3910                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3911                 char *p = dst_kaddr + dst_off + len;
3912                 char *s = src_kaddr + src_off + len;
3913
3914                 while (len--)
3915                         *--p = *--s;
3916
3917                 kunmap_atomic(src_kaddr, KM_USER1);
3918         }
3919         kunmap_atomic(dst_kaddr, KM_USER0);
3920 }
3921
3922 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
3923 {
3924         unsigned long distance = (src > dst) ? src - dst : dst - src;
3925         return distance < len;
3926 }
3927
3928 static void copy_pages(struct page *dst_page, struct page *src_page,
3929                        unsigned long dst_off, unsigned long src_off,
3930                        unsigned long len)
3931 {
3932         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3933         char *src_kaddr;
3934
3935         if (dst_page != src_page) {
3936                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3937         } else {
3938                 src_kaddr = dst_kaddr;
3939                 BUG_ON(areas_overlap(src_off, dst_off, len));
3940         }
3941
3942         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3943         kunmap_atomic(dst_kaddr, KM_USER0);
3944         if (dst_page != src_page)
3945                 kunmap_atomic(src_kaddr, KM_USER1);
3946 }
3947
3948 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3949                            unsigned long src_offset, unsigned long len)
3950 {
3951         size_t cur;
3952         size_t dst_off_in_page;
3953         size_t src_off_in_page;
3954         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3955         unsigned long dst_i;
3956         unsigned long src_i;
3957
3958         if (src_offset + len > dst->len) {
3959                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3960                        "len %lu dst len %lu\n", src_offset, len, dst->len);
3961                 BUG_ON(1);
3962         }
3963         if (dst_offset + len > dst->len) {
3964                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3965                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
3966                 BUG_ON(1);
3967         }
3968
3969         while (len > 0) {
3970                 dst_off_in_page = (start_offset + dst_offset) &
3971                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3972                 src_off_in_page = (start_offset + src_offset) &
3973                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3974
3975                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3976                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3977
3978                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3979                                                src_off_in_page));
3980                 cur = min_t(unsigned long, cur,
3981                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3982
3983                 copy_pages(extent_buffer_page(dst, dst_i),
3984                            extent_buffer_page(dst, src_i),
3985                            dst_off_in_page, src_off_in_page, cur);
3986
3987                 src_offset += cur;
3988                 dst_offset += cur;
3989                 len -= cur;
3990         }
3991 }
3992
3993 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3994                            unsigned long src_offset, unsigned long len)
3995 {
3996         size_t cur;
3997         size_t dst_off_in_page;
3998         size_t src_off_in_page;
3999         unsigned long dst_end = dst_offset + len - 1;
4000         unsigned long src_end = src_offset + len - 1;
4001         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4002         unsigned long dst_i;
4003         unsigned long src_i;
4004
4005         if (src_offset + len > dst->len) {
4006                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4007                        "len %lu len %lu\n", src_offset, len, dst->len);
4008                 BUG_ON(1);
4009         }
4010         if (dst_offset + len > dst->len) {
4011                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4012                        "len %lu len %lu\n", dst_offset, len, dst->len);
4013                 BUG_ON(1);
4014         }
4015         if (!areas_overlap(src_offset, dst_offset, len)) {
4016                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4017                 return;
4018         }
4019         while (len > 0) {
4020                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4021                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4022
4023                 dst_off_in_page = (start_offset + dst_end) &
4024                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4025                 src_off_in_page = (start_offset + src_end) &
4026                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4027
4028                 cur = min_t(unsigned long, len, src_off_in_page + 1);
4029                 cur = min(cur, dst_off_in_page + 1);
4030                 move_pages(extent_buffer_page(dst, dst_i),
4031                            extent_buffer_page(dst, src_i),
4032                            dst_off_in_page - cur + 1,
4033                            src_off_in_page - cur + 1, cur);
4034
4035                 dst_end -= cur;
4036                 src_end -= cur;
4037                 len -= cur;
4038         }
4039 }
4040
4041 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4042 {
4043         struct extent_buffer *eb =
4044                         container_of(head, struct extent_buffer, rcu_head);
4045
4046         btrfs_release_extent_buffer(eb);
4047 }
4048
4049 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
4050 {
4051         u64 start = page_offset(page);
4052         struct extent_buffer *eb;
4053         int ret = 1;
4054
4055         spin_lock(&tree->buffer_lock);
4056         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4057         if (!eb) {
4058                 spin_unlock(&tree->buffer_lock);
4059                 return ret;
4060         }
4061
4062         if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
4063                 ret = 0;
4064                 goto out;
4065         }
4066
4067         /*
4068          * set @eb->refs to 0 if it is already 1, and then release the @eb.
4069          * Or go back.
4070          */
4071         if (atomic_cmpxchg(&eb->refs, 1, 0) != 1) {
4072                 ret = 0;
4073                 goto out;
4074         }
4075
4076         radix_tree_delete(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4077 out:
4078         spin_unlock(&tree->buffer_lock);
4079
4080         /* at this point we can safely release the extent buffer */
4081         if (atomic_read(&eb->refs) == 0)
4082                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4083         return ret;
4084 }