fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/module.h>
   8 #include <linux/spinlock.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/swap.h>
  11 #include <linux/writeback.h>
  12 #include <linux/pagevec.h>
  13 #include <linux/prefetch.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "compat.h"
  17 #include "ctree.h"
  18 #include "btrfs_inode.h"
  19
  20 static struct kmem_cache *extent_state_cache;
  21 static struct kmem_cache *extent_buffer_cache;
  22
  23 static LIST_HEAD(buffers);
  24 static LIST_HEAD(states);
  25
  26 #define LEAK_DEBUG 0
  27 #if LEAK_DEBUG
  28 static DEFINE_SPINLOCK(leak_lock);
  29 #endif
  30
  31 #define BUFFER_LRU_MAX 64
  32
  33 struct tree_entry {
  34         u64 start;
  35         u64 end;
  36         struct rb_node rb_node;
  37 };
  38
  39 struct extent_page_data {
  40         struct bio *bio;
  41         struct extent_io_tree *tree;
  42         get_extent_t *get_extent;
  43
  44         /* tells writepage not to lock the state bits for this range
  45          * it still does the unlocking
  46          */
  47         unsigned int extent_locked:1;
  48
  49         /* tells the submit_bio code to use a WRITE_SYNC */
  50         unsigned int sync_io:1;
  51 };
  52
  53 int __init extent_io_init(void)
  54 {
  55         extent_state_cache = kmem_cache_create("extent_state",
  56                         sizeof(struct extent_state), 0,
  57                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  58         if (!extent_state_cache)
  59                 return -ENOMEM;
  60
  61         extent_buffer_cache = kmem_cache_create("extent_buffers",
  62                         sizeof(struct extent_buffer), 0,
  63                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  64         if (!extent_buffer_cache)
  65                 goto free_state_cache;
  66         return 0;
  67
  68 free_state_cache:
  69         kmem_cache_destroy(extent_state_cache);
  70         return -ENOMEM;
  71 }
  72
  73 void extent_io_exit(void)
  74 {
  75         struct extent_state *state;
  76         struct extent_buffer *eb;
  77
  78         while (!list_empty(&states)) {
  79                 state = list_entry(states.next, struct extent_state, leak_list);
  80                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  81                        "state %lu in tree %p refs %d\n",
  82                        (unsigned long long)state->start,
  83                        (unsigned long long)state->end,
  84                        state->state, state->tree, atomic_read(&state->refs));
  85                 list_del(&state->leak_list);
  86                 kmem_cache_free(extent_state_cache, state);
  87
  88         }
  89
  90         while (!list_empty(&buffers)) {
  91                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  92                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  93                        "refs %d\n", (unsigned long long)eb->start,
  94                        eb->len, atomic_read(&eb->refs));
  95                 list_del(&eb->leak_list);
  96                 kmem_cache_free(extent_buffer_cache, eb);
  97         }
  98         if (extent_state_cache)
  99                 kmem_cache_destroy(extent_state_cache);
 100         if (extent_buffer_cache)
 101                 kmem_cache_destroy(extent_buffer_cache);
 102 }
 103
 104 void extent_io_tree_init(struct extent_io_tree *tree,
 105                           struct address_space *mapping, gfp_t mask)
 106 {
 107         tree->state = RB_ROOT;
 108         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 109         tree->ops = NULL;
 110         tree->dirty_bytes = 0;
 111         spin_lock_init(&tree->lock);
 112         spin_lock_init(&tree->buffer_lock);
 113         tree->mapping = mapping;
 114 }
 115
 116 static struct extent_state *alloc_extent_state(gfp_t mask)
 117 {
 118         struct extent_state *state;
 119 #if LEAK_DEBUG
 120         unsigned long flags;
 121 #endif
 122
 123         state = kmem_cache_alloc(extent_state_cache, mask);
 124         if (!state)
 125                 return state;
 126         state->state = 0;
 127         state->private = 0;
 128         state->tree = NULL;
 129 #if LEAK_DEBUG
 130         spin_lock_irqsave(&leak_lock, flags);
 131         list_add(&state->leak_list, &states);
 132         spin_unlock_irqrestore(&leak_lock, flags);
 133 #endif
 134         atomic_set(&state->refs, 1);
 135         init_waitqueue_head(&state->wq);
 136         return state;
 137 }
 138
 139 void free_extent_state(struct extent_state *state)
 140 {
 141         if (!state)
 142                 return;
 143         if (atomic_dec_and_test(&state->refs)) {
 144 #if LEAK_DEBUG
 145                 unsigned long flags;
 146 #endif
 147                 WARN_ON(state->tree);
 148 #if LEAK_DEBUG
 149                 spin_lock_irqsave(&leak_lock, flags);
 150                 list_del(&state->leak_list);
 151                 spin_unlock_irqrestore(&leak_lock, flags);
 152 #endif
 153                 kmem_cache_free(extent_state_cache, state);
 154         }
 155 }
 156
 157 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 158                                    struct rb_node *node)
 159 {
 160         struct rb_node **p = &root->rb_node;
 161         struct rb_node *parent = NULL;
 162         struct tree_entry *entry;
 163
 164         while (*p) {
 165                 parent = *p;
 166                 entry = rb_entry(parent, struct tree_entry, rb_node);
 167
 168                 if (offset < entry->start)
 169                         p = &(*p)->rb_left;
 170                 else if (offset > entry->end)
 171                         p = &(*p)->rb_right;
 172                 else
 173                         return parent;
 174         }
 175
 176         entry = rb_entry(node, struct tree_entry, rb_node);
 177         rb_link_node(node, parent, p);
 178         rb_insert_color(node, root);
 179         return NULL;
 180 }
 181
 182 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 183                                      struct rb_node **prev_ret,
 184                                      struct rb_node **next_ret)
 185 {
 186         struct rb_root *root = &tree->state;
 187         struct rb_node *n = root->rb_node;
 188         struct rb_node *prev = NULL;
 189         struct rb_node *orig_prev = NULL;
 190         struct tree_entry *entry;
 191         struct tree_entry *prev_entry = NULL;
 192
 193         while (n) {
 194                 entry = rb_entry(n, struct tree_entry, rb_node);
 195                 prev = n;
 196                 prev_entry = entry;
 197
 198                 if (offset < entry->start)
 199                         n = n->rb_left;
 200                 else if (offset > entry->end)
 201                         n = n->rb_right;
 202                 else
 203                         return n;
 204         }
 205
 206         if (prev_ret) {
 207                 orig_prev = prev;
 208                 while (prev && offset > prev_entry->end) {
 209                         prev = rb_next(prev);
 210                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 211                 }
 212                 *prev_ret = prev;
 213                 prev = orig_prev;
 214         }
 215
 216         if (next_ret) {
 217                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 218                 while (prev && offset < prev_entry->start) {
 219                         prev = rb_prev(prev);
 220                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 221                 }
 222                 *next_ret = prev;
 223         }
 224         return NULL;
 225 }
 226
 227 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 228                                           u64 offset)
 229 {
 230         struct rb_node *prev = NULL;
 231         struct rb_node *ret;
 232
 233         ret = __etree_search(tree, offset, &prev, NULL);
 234         if (!ret)
 235                 return prev;
 236         return ret;
 237 }
 238
 239 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 240                      struct extent_state *other)
 241 {
 242         if (tree->ops && tree->ops->merge_extent_hook)
 243                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 244                                              other);
 245 }
 246
 247 /*
 248  * utility function to look for merge candidates inside a given range.
 249  * Any extents with matching state are merged together into a single
 250  * extent in the tree.  Extents with EXTENT_IO in their state field
 251  * are not merged because the end_io handlers need to be able to do
 252  * operations on them without sleeping (or doing allocations/splits).
 253  *
 254  * This should be called with the tree lock held.
 255  */
 256 static int merge_state(struct extent_io_tree *tree,
 257                        struct extent_state *state)
 258 {
 259         struct extent_state *other;
 260         struct rb_node *other_node;
 261
 262         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 263                 return 0;
 264
 265         other_node = rb_prev(&state->rb_node);
 266         if (other_node) {
 267                 other = rb_entry(other_node, struct extent_state, rb_node);
 268                 if (other->end == state->start - 1 &&
 269                     other->state == state->state) {
 270                         merge_cb(tree, state, other);
 271                         state->start = other->start;
 272                         other->tree = NULL;
 273                         rb_erase(&other->rb_node, &tree->state);
 274                         free_extent_state(other);
 275                 }
 276         }
 277         other_node = rb_next(&state->rb_node);
 278         if (other_node) {
 279                 other = rb_entry(other_node, struct extent_state, rb_node);
 280                 if (other->start == state->end + 1 &&
 281                     other->state == state->state) {
 282                         merge_cb(tree, state, other);
 283                         other->start = state->start;
 284                         state->tree = NULL;
 285                         rb_erase(&state->rb_node, &tree->state);
 286                         free_extent_state(state);
 287                         state = NULL;
 288                 }
 289         }
 290
 291         return 0;
 292 }
 293
 294 static int set_state_cb(struct extent_io_tree *tree,
 295                          struct extent_state *state, int *bits)
 296 {
 297         if (tree->ops && tree->ops->set_bit_hook) {
 298                 return tree->ops->set_bit_hook(tree->mapping->host,
 299                                                state, bits);
 300         }
 301
 302         return 0;
 303 }
 304
 305 static void clear_state_cb(struct extent_io_tree *tree,
 306                            struct extent_state *state, int *bits)
 307 {
 308         if (tree->ops && tree->ops->clear_bit_hook)
 309                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 310 }
 311
 312 /*
 313  * insert an extent_state struct into the tree.  'bits' are set on the
 314  * struct before it is inserted.
 315  *
 316  * This may return -EEXIST if the extent is already there, in which case the
 317  * state struct is freed.
 318  *
 319  * The tree lock is not taken internally.  This is a utility function and
 320  * probably isn't what you want to call (see set/clear_extent_bit).
 321  */
 322 static int insert_state(struct extent_io_tree *tree,
 323                         struct extent_state *state, u64 start, u64 end,
 324                         int *bits)
 325 {
 326         struct rb_node *node;
 327         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 328         int ret;
 329
 330         if (end < start) {
 331                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
 332                        (unsigned long long)end,
 333                        (unsigned long long)start);
 334                 WARN_ON(1);
 335         }
 336         state->start = start;
 337         state->end = end;
 338         ret = set_state_cb(tree, state, bits);
 339         if (ret)
 340                 return ret;
 341
 342         if (bits_to_set & EXTENT_DIRTY)
 343                 tree->dirty_bytes += end - start + 1;
 344         state->state |= bits_to_set;
 345         node = tree_insert(&tree->state, end, &state->rb_node);
 346         if (node) {
 347                 struct extent_state *found;
 348                 found = rb_entry(node, struct extent_state, rb_node);
 349                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 350                        "%llu %llu\n", (unsigned long long)found->start,
 351                        (unsigned long long)found->end,
 352                        (unsigned long long)start, (unsigned long long)end);
 353                 free_extent_state(state);
 354                 return -EEXIST;
 355         }
 356         state->tree = tree;
 357         merge_state(tree, state);
 358         return 0;
 359 }
 360
 361 static int split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 362                      u64 split)
 363 {
 364         if (tree->ops && tree->ops->split_extent_hook)
 365                 return tree->ops->split_extent_hook(tree->mapping->host,
 366                                                     orig, split);
 367         return 0;
 368 }
 369
 370 /*
 371  * split a given extent state struct in two, inserting the preallocated
 372  * struct 'prealloc' as the newly created second half.  'split' indicates an
 373  * offset inside 'orig' where it should be split.
 374  *
 375  * Before calling,
 376  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 377  * are two extent state structs in the tree:
 378  * prealloc: [orig->start, split - 1]
 379  * orig: [ split, orig->end ]
 380  *
 381  * The tree locks are not taken by this function. They need to be held
 382  * by the caller.
 383  */
 384 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 385                        struct extent_state *prealloc, u64 split)
 386 {
 387         struct rb_node *node;
 388
 389         split_cb(tree, orig, split);
 390
 391         prealloc->start = orig->start;
 392         prealloc->end = split - 1;
 393         prealloc->state = orig->state;
 394         orig->start = split;
 395
 396         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 397         if (node) {
 398                 free_extent_state(prealloc);
 399                 return -EEXIST;
 400         }
 401         prealloc->tree = tree;
 402         return 0;
 403 }
 404
 405 /*
 406  * utility function to clear some bits in an extent state struct.
 407  * it will optionally wake up any one waiting on this state (wake == 1), or
 408  * forcibly remove the state from the tree (delete == 1).
 409  *
 410  * If no bits are set on the state struct after clearing things, the
 411  * struct is freed and removed from the tree
 412  */
 413 static int clear_state_bit(struct extent_io_tree *tree,
 414                             struct extent_state *state,
 415                             int *bits, int wake)
 416 {
 417         int bits_to_clear = *bits & ~EXTENT_CTLBITS;
 418         int ret = state->state & bits_to_clear;
 419
 420         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 421                 u64 range = state->end - state->start + 1;
 422                 WARN_ON(range > tree->dirty_bytes);
 423                 tree->dirty_bytes -= range;
 424         }
 425         clear_state_cb(tree, state, bits);
 426         state->state &= ~bits_to_clear;
 427         if (wake)
 428                 wake_up(&state->wq);
 429         if (state->state == 0) {
 430                 if (state->tree) {
 431                         rb_erase(&state->rb_node, &tree->state);
 432                         state->tree = NULL;
 433                         free_extent_state(state);
 434                 } else {
 435                         WARN_ON(1);
 436                 }
 437         } else {
 438                 merge_state(tree, state);
 439         }
 440         return ret;
 441 }
 442
 443 /*
 444  * clear some bits on a range in the tree.  This may require splitting
 445  * or inserting elements in the tree, so the gfp mask is used to
 446  * indicate which allocations or sleeping are allowed.
 447  *
 448  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 449  * the given range from the tree regardless of state (ie for truncate).
 450  *
 451  * the range [start, end] is inclusive.
 452  *
 453  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 454  * bits were already set, or zero if none of the bits were already set.
 455  */
 456 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 457                      int bits, int wake, int delete,
 458                      struct extent_state **cached_state,
 459                      gfp_t mask)
 460 {
 461         struct extent_state *state;
 462         struct extent_state *cached;
 463         struct extent_state *prealloc = NULL;
 464         struct rb_node *next_node;
 465         struct rb_node *node;
 466         u64 last_end;
 467         int err;
 468         int set = 0;
 469         int clear = 0;
 470
 471         if (delete)
 472                 bits |= ~EXTENT_CTLBITS;
 473         bits |= EXTENT_FIRST_DELALLOC;
 474
 475         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 476                 clear = 1;
 477 again:
 478         if (!prealloc && (mask & __GFP_WAIT)) {
 479                 prealloc = alloc_extent_state(mask);
 480                 if (!prealloc)
 481                         return -ENOMEM;
 482         }
 483
 484         spin_lock(&tree->lock);
 485         if (cached_state) {
 486                 cached = *cached_state;
 487
 488                 if (clear) {
 489                         *cached_state = NULL;
 490                         cached_state = NULL;
 491                 }
 492
 493                 if (cached && cached->tree && cached->start == start) {
 494                         if (clear)
 495                                 atomic_dec(&cached->refs);
 496                         state = cached;
 497                         goto hit_next;
 498                 }
 499                 if (clear)
 500                         free_extent_state(cached);
 501         }
 502         /*
 503          * this search will find the extents that end after
 504          * our range starts
 505          */
 506         node = tree_search(tree, start);
 507         if (!node)
 508                 goto out;
 509         state = rb_entry(node, struct extent_state, rb_node);
 510 hit_next:
 511         if (state->start > end)
 512                 goto out;
 513         WARN_ON(state->end < start);
 514         last_end = state->end;
 515
 516         /*
 517          *     | ---- desired range ---- |
 518          *  | state | or
 519          *  | ------------- state -------------- |
 520          *
 521          * We need to split the extent we found, and may flip
 522          * bits on second half.
 523          *
 524          * If the extent we found extends past our range, we
 525          * just split and search again.  It'll get split again
 526          * the next time though.
 527          *
 528          * If the extent we found is inside our range, we clear
 529          * the desired bit on it.
 530          */
 531
 532         if (state->start < start) {
 533                 if (!prealloc)
 534                         prealloc = alloc_extent_state(GFP_ATOMIC);
 535                 err = split_state(tree, state, prealloc, start);
 536                 BUG_ON(err == -EEXIST);
 537                 prealloc = NULL;
 538                 if (err)
 539                         goto out;
 540                 if (state->end <= end) {
 541                         set |= clear_state_bit(tree, state, &bits, wake);
 542                         if (last_end == (u64)-1)
 543                                 goto out;
 544                         start = last_end + 1;
 545                 }
 546                 goto search_again;
 547         }
 548         /*
 549          * | ---- desired range ---- |
 550          *                        | state |
 551          * We need to split the extent, and clear the bit
 552          * on the first half
 553          */
 554         if (state->start <= end && state->end > end) {
 555                 if (!prealloc)
 556                         prealloc = alloc_extent_state(GFP_ATOMIC);
 557                 err = split_state(tree, state, prealloc, end + 1);
 558                 BUG_ON(err == -EEXIST);
 559                 if (wake)
 560                         wake_up(&state->wq);
 561
 562                 set |= clear_state_bit(tree, prealloc, &bits, wake);
 563
 564                 prealloc = NULL;
 565                 goto out;
 566         }
 567
 568         if (state->end < end && prealloc && !need_resched())
 569                 next_node = rb_next(&state->rb_node);
 570         else
 571                 next_node = NULL;
 572
 573         set |= clear_state_bit(tree, state, &bits, wake);
 574         if (last_end == (u64)-1)
 575                 goto out;
 576         start = last_end + 1;
 577         if (start <= end && next_node) {
 578                 state = rb_entry(next_node, struct extent_state,
 579                                  rb_node);
 580                 if (state->start == start)
 581                         goto hit_next;
 582         }
 583         goto search_again;
 584
 585 out:
 586         spin_unlock(&tree->lock);
 587         if (prealloc)
 588                 free_extent_state(prealloc);
 589
 590         return set;
 591
 592 search_again:
 593         if (start > end)
 594                 goto out;
 595         spin_unlock(&tree->lock);
 596         if (mask & __GFP_WAIT)
 597                 cond_resched();
 598         goto again;
 599 }
 600
 601 static int wait_on_state(struct extent_io_tree *tree,
 602                          struct extent_state *state)
 603                 __releases(tree->lock)
 604                 __acquires(tree->lock)
 605 {
 606         DEFINE_WAIT(wait);
 607         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 608         spin_unlock(&tree->lock);
 609         schedule();
 610         spin_lock(&tree->lock);
 611         finish_wait(&state->wq, &wait);
 612         return 0;
 613 }
 614
 615 /*
 616  * waits for one or more bits to clear on a range in the state tree.
 617  * The range [start, end] is inclusive.
 618  * The tree lock is taken by this function
 619  */
 620 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 621 {
 622         struct extent_state *state;
 623         struct rb_node *node;
 624
 625         spin_lock(&tree->lock);
 626 again:
 627         while (1) {
 628                 /*
 629                  * this search will find all the extents that end after
 630                  * our range starts
 631                  */
 632                 node = tree_search(tree, start);
 633                 if (!node)
 634                         break;
 635
 636                 state = rb_entry(node, struct extent_state, rb_node);
 637
 638                 if (state->start > end)
 639                         goto out;
 640
 641                 if (state->state & bits) {
 642                         start = state->start;
 643                         atomic_inc(&state->refs);
 644                         wait_on_state(tree, state);
 645                         free_extent_state(state);
 646                         goto again;
 647                 }
 648                 start = state->end + 1;
 649
 650                 if (start > end)
 651                         break;
 652
 653                 if (need_resched()) {
 654                         spin_unlock(&tree->lock);
 655                         cond_resched();
 656                         spin_lock(&tree->lock);
 657                 }
 658         }
 659 out:
 660         spin_unlock(&tree->lock);
 661         return 0;
 662 }
 663
 664 static int set_state_bits(struct extent_io_tree *tree,
 665                            struct extent_state *state,
 666                            int *bits)
 667 {
 668         int ret;
 669         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 670
 671         ret = set_state_cb(tree, state, bits);
 672         if (ret)
 673                 return ret;
 674         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 675                 u64 range = state->end - state->start + 1;
 676                 tree->dirty_bytes += range;
 677         }
 678         state->state |= bits_to_set;
 679
 680         return 0;
 681 }
 682
 683 static void cache_state(struct extent_state *state,
 684                         struct extent_state **cached_ptr)
 685 {
 686         if (cached_ptr && !(*cached_ptr)) {
 687                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 688                         *cached_ptr = state;
 689                         atomic_inc(&state->refs);
 690                 }
 691         }
 692 }
 693
 694 static void uncache_state(struct extent_state **cached_ptr)
 695 {
 696         if (cached_ptr && (*cached_ptr)) {
 697                 struct extent_state *state = *cached_ptr;
 698                 *cached_ptr = NULL;
 699                 free_extent_state(state);
 700         }
 701 }
 702
 703 /*
 704  * set some bits on a range in the tree.  This may require allocations or
 705  * sleeping, so the gfp mask is used to indicate what is allowed.
 706  *
 707  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 708  * part of the range already has the desired bits set.  The start of the
 709  * existing range is returned in failed_start in this case.
 710  *
 711  * [start, end] is inclusive This takes the tree lock.
 712  */
 713
 714 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 715                    int bits, int exclusive_bits, u64 *failed_start,
 716                    struct extent_state **cached_state, gfp_t mask)
 717 {
 718         struct extent_state *state;
 719         struct extent_state *prealloc = NULL;
 720         struct rb_node *node;
 721         int err = 0;
 722         u64 last_start;
 723         u64 last_end;
 724
 725         bits |= EXTENT_FIRST_DELALLOC;
 726 again:
 727         if (!prealloc && (mask & __GFP_WAIT)) {
 728                 prealloc = alloc_extent_state(mask);
 729                 if (!prealloc)
 730                         return -ENOMEM;
 731         }
 732
 733         spin_lock(&tree->lock);
 734         if (cached_state && *cached_state) {
 735                 state = *cached_state;
 736                 if (state->start == start && state->tree) {
 737                         node = &state->rb_node;
 738                         goto hit_next;
 739                 }
 740         }
 741         /*
 742          * this search will find all the extents that end after
 743          * our range starts.
 744          */
 745         node = tree_search(tree, start);
 746         if (!node) {
 747                 err = insert_state(tree, prealloc, start, end, &bits);
 748                 prealloc = NULL;
 749                 BUG_ON(err == -EEXIST);
 750                 goto out;
 751         }
 752         state = rb_entry(node, struct extent_state, rb_node);
 753 hit_next:
 754         last_start = state->start;
 755         last_end = state->end;
 756
 757         /*
 758          * | ---- desired range ---- |
 759          * | state |
 760          *
 761          * Just lock what we found and keep going
 762          */
 763         if (state->start == start && state->end <= end) {
 764                 struct rb_node *next_node;
 765                 if (state->state & exclusive_bits) {
 766                         *failed_start = state->start;
 767                         err = -EEXIST;
 768                         goto out;
 769                 }
 770
 771                 err = set_state_bits(tree, state, &bits);
 772                 if (err)
 773                         goto out;
 774
 775                 cache_state(state, cached_state);
 776                 merge_state(tree, state);
 777                 if (last_end == (u64)-1)
 778                         goto out;
 779
 780                 start = last_end + 1;
 781                 if (start < end && prealloc && !need_resched()) {
 782                         next_node = rb_next(node);
 783                         if (next_node) {
 784                                 state = rb_entry(next_node, struct extent_state,
 785                                                  rb_node);
 786                                 if (state->start == start)
 787                                         goto hit_next;
 788                         }
 789                 }
 790                 goto search_again;
 791         }
 792
 793         /*
 794          *     | ---- desired range ---- |
 795          * | state |
 796          *   or
 797          * | ------------- state -------------- |
 798          *
 799          * We need to split the extent we found, and may flip bits on
 800          * second half.
 801          *
 802          * If the extent we found extends past our
 803          * range, we just split and search again.  It'll get split
 804          * again the next time though.
 805          *
 806          * If the extent we found is inside our range, we set the
 807          * desired bit on it.
 808          */
 809         if (state->start < start) {
 810                 if (state->state & exclusive_bits) {
 811                         *failed_start = start;
 812                         err = -EEXIST;
 813                         goto out;
 814                 }
 815                 err = split_state(tree, state, prealloc, start);
 816                 BUG_ON(err == -EEXIST);
 817                 prealloc = NULL;
 818                 if (err)
 819                         goto out;
 820                 if (state->end <= end) {
 821                         err = set_state_bits(tree, state, &bits);
 822                         if (err)
 823                                 goto out;
 824                         cache_state(state, cached_state);
 825                         merge_state(tree, state);
 826                         if (last_end == (u64)-1)
 827                                 goto out;
 828                         start = last_end + 1;
 829                 }
 830                 goto search_again;
 831         }
 832         /*
 833          * | ---- desired range ---- |
 834          *     | state | or               | state |
 835          *
 836          * There's a hole, we need to insert something in it and
 837          * ignore the extent we found.
 838          */
 839         if (state->start > start) {
 840                 u64 this_end;
 841                 if (end < last_start)
 842                         this_end = end;
 843                 else
 844                         this_end = last_start - 1;
 845                 err = insert_state(tree, prealloc, start, this_end,
 846                                    &bits);
 847                 BUG_ON(err == -EEXIST);
 848                 if (err) {
 849                         prealloc = NULL;
 850                         goto out;
 851                 }
 852                 cache_state(prealloc, cached_state);
 853                 prealloc = NULL;
 854                 start = this_end + 1;
 855                 goto search_again;
 856         }
 857         /*
 858          * | ---- desired range ---- |
 859          *                        | state |
 860          * We need to split the extent, and set the bit
 861          * on the first half
 862          */
 863         if (state->start <= end && state->end > end) {
 864                 if (state->state & exclusive_bits) {
 865                         *failed_start = start;
 866                         err = -EEXIST;
 867                         goto out;
 868                 }
 869                 err = split_state(tree, state, prealloc, end + 1);
 870                 BUG_ON(err == -EEXIST);
 871
 872                 err = set_state_bits(tree, prealloc, &bits);
 873                 if (err) {
 874                         prealloc = NULL;
 875                         goto out;
 876                 }
 877                 cache_state(prealloc, cached_state);
 878                 merge_state(tree, prealloc);
 879                 prealloc = NULL;
 880                 goto out;
 881         }
 882
 883         goto search_again;
 884
 885 out:
 886         spin_unlock(&tree->lock);
 887         if (prealloc)
 888                 free_extent_state(prealloc);
 889
 890         return err;
 891
 892 search_again:
 893         if (start > end)
 894                 goto out;
 895         spin_unlock(&tree->lock);
 896         if (mask & __GFP_WAIT)
 897                 cond_resched();
 898         goto again;
 899 }
 900
 901 /* wrappers around set/clear extent bit */
 902 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 903                      gfp_t mask)
 904 {
 905         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 906                               NULL, mask);
 907 }
 908
 909 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 910                     int bits, gfp_t mask)
 911 {
 912         return set_extent_bit(tree, start, end, bits, 0, NULL,
 913                               NULL, mask);
 914 }
 915
 916 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 917                       int bits, gfp_t mask)
 918 {
 919         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
 920 }
 921
 922 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
 923                         struct extent_state **cached_state, gfp_t mask)
 924 {
 925         return set_extent_bit(tree, start, end,
 926                               EXTENT_DELALLOC | EXTENT_DIRTY | EXTENT_UPTODATE,
 927                               0, NULL, cached_state, mask);
 928 }
 929
 930 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 931                        gfp_t mask)
 932 {
 933         return clear_extent_bit(tree, start, end,
 934                                 EXTENT_DIRTY | EXTENT_DELALLOC |
 935                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
 936 }
 937
 938 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 939                      gfp_t mask)
 940 {
 941         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 942                               NULL, mask);
 943 }
 944
 945 static int clear_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 946                        gfp_t mask)
 947 {
 948         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0,
 949                                 NULL, mask);
 950 }
 951
 952 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 953                         struct extent_state **cached_state, gfp_t mask)
 954 {
 955         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
 956                               NULL, cached_state, mask);
 957 }
 958
 959 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
 960                                  u64 end, struct extent_state **cached_state,
 961                                  gfp_t mask)
 962 {
 963         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
 964                                 cached_state, mask);
 965 }
 966
 967 int wait_on_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 968 {
 969         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 970 }
 971
 972 /*
 973  * either insert or lock state struct between start and end use mask to tell
 974  * us if waiting is desired.
 975  */
 976 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 977                      int bits, struct extent_state **cached_state, gfp_t mask)
 978 {
 979         int err;
 980         u64 failed_start;
 981         while (1) {
 982                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
 983                                      EXTENT_LOCKED, &failed_start,
 984                                      cached_state, mask);
 985                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
 986                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
 987                         start = failed_start;
 988                 } else {
 989                         break;
 990                 }
 991                 WARN_ON(start > end);
 992         }
 993         return err;
 994 }
 995
 996 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
 997 {
 998         return lock_extent_bits(tree, start, end, 0, NULL, mask);
 999 }
1000
1001 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1002                     gfp_t mask)
1003 {
1004         int err;
1005         u64 failed_start;
1006
1007         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1008                              &failed_start, NULL, mask);
1009         if (err == -EEXIST) {
1010                 if (failed_start > start)
1011                         clear_extent_bit(tree, start, failed_start - 1,
1012                                          EXTENT_LOCKED, 1, 0, NULL, mask);
1013                 return 0;
1014         }
1015         return 1;
1016 }
1017
1018 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1019                          struct extent_state **cached, gfp_t mask)
1020 {
1021         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1022                                 mask);
1023 }
1024
1025 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1026 {
1027         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1028                                 mask);
1029 }
1030
1031 /*
1032  * helper function to set pages and extents in the tree dirty
1033  */
1034 int set_range_dirty(struct extent_io_tree *tree, u64 start, u64 end)
1035 {
1036         unsigned long index = start >> PAGE_CACHE_SHIFT;
1037         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1038         struct page *page;
1039
1040         while (index <= end_index) {
1041                 page = find_get_page(tree->mapping, index);
1042                 BUG_ON(!page);
1043                 __set_page_dirty_nobuffers(page);
1044                 page_cache_release(page);
1045                 index++;
1046         }
1047         return 0;
1048 }
1049
1050 /*
1051  * helper function to set both pages and extents in the tree writeback
1052  */
1053 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1054 {
1055         unsigned long index = start >> PAGE_CACHE_SHIFT;
1056         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1057         struct page *page;
1058
1059         while (index <= end_index) {
1060                 page = find_get_page(tree->mapping, index);
1061                 BUG_ON(!page);
1062                 set_page_writeback(page);
1063                 page_cache_release(page);
1064                 index++;
1065         }
1066         return 0;
1067 }
1068
1069 /*
1070  * find the first offset in the io tree with 'bits' set. zero is
1071  * returned if we find something, and *start_ret and *end_ret are
1072  * set to reflect the state struct that was found.
1073  *
1074  * If nothing was found, 1 is returned, < 0 on error
1075  */
1076 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1077                           u64 *start_ret, u64 *end_ret, int bits)
1078 {
1079         struct rb_node *node;
1080         struct extent_state *state;
1081         int ret = 1;
1082
1083         spin_lock(&tree->lock);
1084         /*
1085          * this search will find all the extents that end after
1086          * our range starts.
1087          */
1088         node = tree_search(tree, start);
1089         if (!node)
1090                 goto out;
1091
1092         while (1) {
1093                 state = rb_entry(node, struct extent_state, rb_node);
1094                 if (state->end >= start && (state->state & bits)) {
1095                         *start_ret = state->start;
1096                         *end_ret = state->end;
1097                         ret = 0;
1098                         break;
1099                 }
1100                 node = rb_next(node);
1101                 if (!node)
1102                         break;
1103         }
1104 out:
1105         spin_unlock(&tree->lock);
1106         return ret;
1107 }
1108
1109 /* find the first state struct with 'bits' set after 'start', and
1110  * return it.  tree->lock must be held.  NULL will returned if
1111  * nothing was found after 'start'
1112  */
1113 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1114                                                  u64 start, int bits)
1115 {
1116         struct rb_node *node;
1117         struct extent_state *state;
1118
1119         /*
1120          * this search will find all the extents that end after
1121          * our range starts.
1122          */
1123         node = tree_search(tree, start);
1124         if (!node)
1125                 goto out;
1126
1127         while (1) {
1128                 state = rb_entry(node, struct extent_state, rb_node);
1129                 if (state->end >= start && (state->state & bits))
1130                         return state;
1131
1132                 node = rb_next(node);
1133                 if (!node)
1134                         break;
1135         }
1136 out:
1137         return NULL;
1138 }
1139
1140 /*
1141  * find a contiguous range of bytes in the file marked as delalloc, not
1142  * more than 'max_bytes'.  start and end are used to return the range,
1143  *
1144  * 1 is returned if we find something, 0 if nothing was in the tree
1145  */
1146 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1147                                         u64 *start, u64 *end, u64 max_bytes,
1148                                         struct extent_state **cached_state)
1149 {
1150         struct rb_node *node;
1151         struct extent_state *state;
1152         u64 cur_start = *start;
1153         u64 found = 0;
1154         u64 total_bytes = 0;
1155
1156         spin_lock(&tree->lock);
1157
1158         /*
1159          * this search will find all the extents that end after
1160          * our range starts.
1161          */
1162         node = tree_search(tree, cur_start);
1163         if (!node) {
1164                 if (!found)
1165                         *end = (u64)-1;
1166                 goto out;
1167         }
1168
1169         while (1) {
1170                 state = rb_entry(node, struct extent_state, rb_node);
1171                 if (found && (state->start != cur_start ||
1172                               (state->state & EXTENT_BOUNDARY))) {
1173                         goto out;
1174                 }
1175                 if (!(state->state & EXTENT_DELALLOC)) {
1176                         if (!found)
1177                                 *end = state->end;
1178                         goto out;
1179                 }
1180                 if (!found) {
1181                         *start = state->start;
1182                         *cached_state = state;
1183                         atomic_inc(&state->refs);
1184                 }
1185                 found++;
1186                 *end = state->end;
1187                 cur_start = state->end + 1;
1188                 node = rb_next(node);
1189                 if (!node)
1190                         break;
1191                 total_bytes += state->end - state->start + 1;
1192                 if (total_bytes >= max_bytes)
1193                         break;
1194         }
1195 out:
1196         spin_unlock(&tree->lock);
1197         return found;
1198 }
1199
1200 static noinline int __unlock_for_delalloc(struct inode *inode,
1201                                           struct page *locked_page,
1202                                           u64 start, u64 end)
1203 {
1204         int ret;
1205         struct page *pages[16];
1206         unsigned long index = start >> PAGE_CACHE_SHIFT;
1207         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1208         unsigned long nr_pages = end_index - index + 1;
1209         int i;
1210
1211         if (index == locked_page->index && end_index == index)
1212                 return 0;
1213
1214         while (nr_pages > 0) {
1215                 ret = find_get_pages_contig(inode->i_mapping, index,
1216                                      min_t(unsigned long, nr_pages,
1217                                      ARRAY_SIZE(pages)), pages);
1218                 for (i = 0; i < ret; i++) {
1219                         if (pages[i] != locked_page)
1220                                 unlock_page(pages[i]);
1221                         page_cache_release(pages[i]);
1222                 }
1223                 nr_pages -= ret;
1224                 index += ret;
1225                 cond_resched();
1226         }
1227         return 0;
1228 }
1229
1230 static noinline int lock_delalloc_pages(struct inode *inode,
1231                                         struct page *locked_page,
1232                                         u64 delalloc_start,
1233                                         u64 delalloc_end)
1234 {
1235         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1236         unsigned long start_index = index;
1237         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1238         unsigned long pages_locked = 0;
1239         struct page *pages[16];
1240         unsigned long nrpages;
1241         int ret;
1242         int i;
1243
1244         /* the caller is responsible for locking the start index */
1245         if (index == locked_page->index && index == end_index)
1246                 return 0;
1247
1248         /* skip the page at the start index */
1249         nrpages = end_index - index + 1;
1250         while (nrpages > 0) {
1251                 ret = find_get_pages_contig(inode->i_mapping, index,
1252                                      min_t(unsigned long,
1253                                      nrpages, ARRAY_SIZE(pages)), pages);
1254                 if (ret == 0) {
1255                         ret = -EAGAIN;
1256                         goto done;
1257                 }
1258                 /* now we have an array of pages, lock them all */
1259                 for (i = 0; i < ret; i++) {
1260                         /*
1261                          * the caller is taking responsibility for
1262                          * locked_page
1263                          */
1264                         if (pages[i] != locked_page) {
1265                                 lock_page(pages[i]);
1266                                 if (!PageDirty(pages[i]) ||
1267                                     pages[i]->mapping != inode->i_mapping) {
1268                                         ret = -EAGAIN;
1269                                         unlock_page(pages[i]);
1270                                         page_cache_release(pages[i]);
1271                                         goto done;
1272                                 }
1273                         }
1274                         page_cache_release(pages[i]);
1275                         pages_locked++;
1276                 }
1277                 nrpages -= ret;
1278                 index += ret;
1279                 cond_resched();
1280         }
1281         ret = 0;
1282 done:
1283         if (ret && pages_locked) {
1284                 __unlock_for_delalloc(inode, locked_page,
1285                               delalloc_start,
1286                               ((u64)(start_index + pages_locked - 1)) <<
1287                               PAGE_CACHE_SHIFT);
1288         }
1289         return ret;
1290 }
1291
1292 /*
1293  * find a contiguous range of bytes in the file marked as delalloc, not
1294  * more than 'max_bytes'.  start and end are used to return the range,
1295  *
1296  * 1 is returned if we find something, 0 if nothing was in the tree
1297  */
1298 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1299                                              struct extent_io_tree *tree,
1300                                              struct page *locked_page,
1301                                              u64 *start, u64 *end,
1302                                              u64 max_bytes)
1303 {
1304         u64 delalloc_start;
1305         u64 delalloc_end;
1306         u64 found;
1307         struct extent_state *cached_state = NULL;
1308         int ret;
1309         int loops = 0;
1310
1311 again:
1312         /* step one, find a bunch of delalloc bytes starting at start */
1313         delalloc_start = *start;
1314         delalloc_end = 0;
1315         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1316                                     max_bytes, &cached_state);
1317         if (!found || delalloc_end <= *start) {
1318                 *start = delalloc_start;
1319                 *end = delalloc_end;
1320                 free_extent_state(cached_state);
1321                 return found;
1322         }
1323
1324         /*
1325          * start comes from the offset of locked_page.  We have to lock
1326          * pages in order, so we can't process delalloc bytes before
1327          * locked_page
1328          */
1329         if (delalloc_start < *start)
1330                 delalloc_start = *start;
1331
1332         /*
1333          * make sure to limit the number of pages we try to lock down
1334          * if we're looping.
1335          */
1336         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1337                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1338
1339         /* step two, lock all the pages after the page that has start */
1340         ret = lock_delalloc_pages(inode, locked_page,
1341                                   delalloc_start, delalloc_end);
1342         if (ret == -EAGAIN) {
1343                 /* some of the pages are gone, lets avoid looping by
1344                  * shortening the size of the delalloc range we're searching
1345                  */
1346                 free_extent_state(cached_state);
1347                 if (!loops) {
1348                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1349                         max_bytes = PAGE_CACHE_SIZE - offset;
1350                         loops = 1;
1351                         goto again;
1352                 } else {
1353                         found = 0;
1354                         goto out_failed;
1355                 }
1356         }
1357         BUG_ON(ret);
1358
1359         /* step three, lock the state bits for the whole range */
1360         lock_extent_bits(tree, delalloc_start, delalloc_end,
1361                          0, &cached_state, GFP_NOFS);
1362
1363         /* then test to make sure it is all still delalloc */
1364         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1365                              EXTENT_DELALLOC, 1, cached_state);
1366         if (!ret) {
1367                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1368                                      &cached_state, GFP_NOFS);
1369                 __unlock_for_delalloc(inode, locked_page,
1370                               delalloc_start, delalloc_end);
1371                 cond_resched();
1372                 goto again;
1373         }
1374         free_extent_state(cached_state);
1375         *start = delalloc_start;
1376         *end = delalloc_end;
1377 out_failed:
1378         return found;
1379 }
1380
1381 int extent_clear_unlock_delalloc(struct inode *inode,
1382                                 struct extent_io_tree *tree,
1383                                 u64 start, u64 end, struct page *locked_page,
1384                                 unsigned long op)
1385 {
1386         int ret;
1387         struct page *pages[16];
1388         unsigned long index = start >> PAGE_CACHE_SHIFT;
1389         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1390         unsigned long nr_pages = end_index - index + 1;
1391         int i;
1392         int clear_bits = 0;
1393
1394         if (op & EXTENT_CLEAR_UNLOCK)
1395                 clear_bits |= EXTENT_LOCKED;
1396         if (op & EXTENT_CLEAR_DIRTY)
1397                 clear_bits |= EXTENT_DIRTY;
1398
1399         if (op & EXTENT_CLEAR_DELALLOC)
1400                 clear_bits |= EXTENT_DELALLOC;
1401
1402         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1403         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1404                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1405                     EXTENT_SET_PRIVATE2)))
1406                 return 0;
1407
1408         while (nr_pages > 0) {
1409                 ret = find_get_pages_contig(inode->i_mapping, index,
1410                                      min_t(unsigned long,
1411                                      nr_pages, ARRAY_SIZE(pages)), pages);
1412                 for (i = 0; i < ret; i++) {
1413
1414                         if (op & EXTENT_SET_PRIVATE2)
1415                                 SetPagePrivate2(pages[i]);
1416
1417                         if (pages[i] == locked_page) {
1418                                 page_cache_release(pages[i]);
1419                                 continue;
1420                         }
1421                         if (op & EXTENT_CLEAR_DIRTY)
1422                                 clear_page_dirty_for_io(pages[i]);
1423                         if (op & EXTENT_SET_WRITEBACK)
1424                                 set_page_writeback(pages[i]);
1425                         if (op & EXTENT_END_WRITEBACK)
1426                                 end_page_writeback(pages[i]);
1427                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1428                                 unlock_page(pages[i]);
1429                         page_cache_release(pages[i]);
1430                 }
1431                 nr_pages -= ret;
1432                 index += ret;
1433                 cond_resched();
1434         }
1435         return 0;
1436 }
1437
1438 /*
1439  * count the number of bytes in the tree that have a given bit(s)
1440  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1441  * cached.  The total number found is returned.
1442  */
1443 u64 count_range_bits(struct extent_io_tree *tree,
1444                      u64 *start, u64 search_end, u64 max_bytes,
1445                      unsigned long bits, int contig)
1446 {
1447         struct rb_node *node;
1448         struct extent_state *state;
1449         u64 cur_start = *start;
1450         u64 total_bytes = 0;
1451         u64 last = 0;
1452         int found = 0;
1453
1454         if (search_end <= cur_start) {
1455                 WARN_ON(1);
1456                 return 0;
1457         }
1458
1459         spin_lock(&tree->lock);
1460         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1461                 total_bytes = tree->dirty_bytes;
1462                 goto out;
1463         }
1464         /*
1465          * this search will find all the extents that end after
1466          * our range starts.
1467          */
1468         node = tree_search(tree, cur_start);
1469         if (!node)
1470                 goto out;
1471
1472         while (1) {
1473                 state = rb_entry(node, struct extent_state, rb_node);
1474                 if (state->start > search_end)
1475                         break;
1476                 if (contig && found && state->start > last + 1)
1477                         break;
1478                 if (state->end >= cur_start && (state->state & bits) == bits) {
1479                         total_bytes += min(search_end, state->end) + 1 -
1480                                        max(cur_start, state->start);
1481                         if (total_bytes >= max_bytes)
1482                                 break;
1483                         if (!found) {
1484                                 *start = state->start;
1485                                 found = 1;
1486                         }
1487                         last = state->end;
1488                 } else if (contig && found) {
1489                         break;
1490                 }
1491                 node = rb_next(node);
1492                 if (!node)
1493                         break;
1494         }
1495 out:
1496         spin_unlock(&tree->lock);
1497         return total_bytes;
1498 }
1499
1500 /*
1501  * set the private field for a given byte offset in the tree.  If there isn't
1502  * an extent_state there already, this does nothing.
1503  */
1504 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1505 {
1506         struct rb_node *node;
1507         struct extent_state *state;
1508         int ret = 0;
1509
1510         spin_lock(&tree->lock);
1511         /*
1512          * this search will find all the extents that end after
1513          * our range starts.
1514          */
1515         node = tree_search(tree, start);
1516         if (!node) {
1517                 ret = -ENOENT;
1518                 goto out;
1519         }
1520         state = rb_entry(node, struct extent_state, rb_node);
1521         if (state->start != start) {
1522                 ret = -ENOENT;
1523                 goto out;
1524         }
1525         state->private = private;
1526 out:
1527         spin_unlock(&tree->lock);
1528         return ret;
1529 }
1530
1531 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1532 {
1533         struct rb_node *node;
1534         struct extent_state *state;
1535         int ret = 0;
1536
1537         spin_lock(&tree->lock);
1538         /*
1539          * this search will find all the extents that end after
1540          * our range starts.
1541          */
1542         node = tree_search(tree, start);
1543         if (!node) {
1544                 ret = -ENOENT;
1545                 goto out;
1546         }
1547         state = rb_entry(node, struct extent_state, rb_node);
1548         if (state->start != start) {
1549                 ret = -ENOENT;
1550                 goto out;
1551         }
1552         *private = state->private;
1553 out:
1554         spin_unlock(&tree->lock);
1555         return ret;
1556 }
1557
1558 /*
1559  * searches a range in the state tree for a given mask.
1560  * If 'filled' == 1, this returns 1 only if every extent in the tree
1561  * has the bits set.  Otherwise, 1 is returned if any bit in the
1562  * range is found set.
1563  */
1564 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1565                    int bits, int filled, struct extent_state *cached)
1566 {
1567         struct extent_state *state = NULL;
1568         struct rb_node *node;
1569         int bitset = 0;
1570
1571         spin_lock(&tree->lock);
1572         if (cached && cached->tree && cached->start == start)
1573                 node = &cached->rb_node;
1574         else
1575                 node = tree_search(tree, start);
1576         while (node && start <= end) {
1577                 state = rb_entry(node, struct extent_state, rb_node);
1578
1579                 if (filled && state->start > start) {
1580                         bitset = 0;
1581                         break;
1582                 }
1583
1584                 if (state->start > end)
1585                         break;
1586
1587                 if (state->state & bits) {
1588                         bitset = 1;
1589                         if (!filled)
1590                                 break;
1591                 } else if (filled) {
1592                         bitset = 0;
1593                         break;
1594                 }
1595
1596                 if (state->end == (u64)-1)
1597                         break;
1598
1599                 start = state->end + 1;
1600                 if (start > end)
1601                         break;
1602                 node = rb_next(node);
1603                 if (!node) {
1604                         if (filled)
1605                                 bitset = 0;
1606                         break;
1607                 }
1608         }
1609         spin_unlock(&tree->lock);
1610         return bitset;
1611 }
1612
1613 /*
1614  * helper function to set a given page up to date if all the
1615  * extents in the tree for that page are up to date
1616  */
1617 static int check_page_uptodate(struct extent_io_tree *tree,
1618                                struct page *page)
1619 {
1620         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1621         u64 end = start + PAGE_CACHE_SIZE - 1;
1622         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1623                 SetPageUptodate(page);
1624         return 0;
1625 }
1626
1627 /*
1628  * helper function to unlock a page if all the extents in the tree
1629  * for that page are unlocked
1630  */
1631 static int check_page_locked(struct extent_io_tree *tree,
1632                              struct page *page)
1633 {
1634         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1635         u64 end = start + PAGE_CACHE_SIZE - 1;
1636         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1637                 unlock_page(page);
1638         return 0;
1639 }
1640
1641 /*
1642  * helper function to end page writeback if all the extents
1643  * in the tree for that page are done with writeback
1644  */
1645 static int check_page_writeback(struct extent_io_tree *tree,
1646                              struct page *page)
1647 {
1648         end_page_writeback(page);
1649         return 0;
1650 }
1651
1652 /* lots and lots of room for performance fixes in the end_bio funcs */
1653
1654 /*
1655  * after a writepage IO is done, we need to:
1656  * clear the uptodate bits on error
1657  * clear the writeback bits in the extent tree for this IO
1658  * end_page_writeback if the page has no more pending IO
1659  *
1660  * Scheduling is not allowed, so the extent state tree is expected
1661  * to have one and only one object corresponding to this IO.
1662  */
1663 static void end_bio_extent_writepage(struct bio *bio, int err)
1664 {
1665         int uptodate = err == 0;
1666         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1667         struct extent_io_tree *tree;
1668         u64 start;
1669         u64 end;
1670         int whole_page;
1671         int ret;
1672
1673         do {
1674                 struct page *page = bvec->bv_page;
1675                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1676
1677                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1678                          bvec->bv_offset;
1679                 end = start + bvec->bv_len - 1;
1680
1681                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1682                         whole_page = 1;
1683                 else
1684                         whole_page = 0;
1685
1686                 if (--bvec >= bio->bi_io_vec)
1687                         prefetchw(&bvec->bv_page->flags);
1688                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1689                         ret = tree->ops->writepage_end_io_hook(page, start,
1690                                                        end, NULL, uptodate);
1691                         if (ret)
1692                                 uptodate = 0;
1693                 }
1694
1695                 if (!uptodate && tree->ops &&
1696                     tree->ops->writepage_io_failed_hook) {
1697                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1698                                                          start, end, NULL);
1699                         if (ret == 0) {
1700                                 uptodate = (err == 0);
1701                                 continue;
1702                         }
1703                 }
1704
1705                 if (!uptodate) {
1706                         clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
1707                         ClearPageUptodate(page);
1708                         SetPageError(page);
1709                 }
1710
1711                 if (whole_page)
1712                         end_page_writeback(page);
1713                 else
1714                         check_page_writeback(tree, page);
1715         } while (bvec >= bio->bi_io_vec);
1716
1717         bio_put(bio);
1718 }
1719
1720 /*
1721  * after a readpage IO is done, we need to:
1722  * clear the uptodate bits on error
1723  * set the uptodate bits if things worked
1724  * set the page up to date if all extents in the tree are uptodate
1725  * clear the lock bit in the extent tree
1726  * unlock the page if there are no other extents locked for it
1727  *
1728  * Scheduling is not allowed, so the extent state tree is expected
1729  * to have one and only one object corresponding to this IO.
1730  */
1731 static void end_bio_extent_readpage(struct bio *bio, int err)
1732 {
1733         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1734         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
1735         struct bio_vec *bvec = bio->bi_io_vec;
1736         struct extent_io_tree *tree;
1737         u64 start;
1738         u64 end;
1739         int whole_page;
1740         int ret;
1741
1742         if (err)
1743                 uptodate = 0;
1744
1745         do {
1746                 struct page *page = bvec->bv_page;
1747                 struct extent_state *cached = NULL;
1748                 struct extent_state *state;
1749
1750                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1751
1752                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1753                         bvec->bv_offset;
1754                 end = start + bvec->bv_len - 1;
1755
1756                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1757                         whole_page = 1;
1758                 else
1759                         whole_page = 0;
1760
1761                 if (++bvec <= bvec_end)
1762                         prefetchw(&bvec->bv_page->flags);
1763
1764                 spin_lock(&tree->lock);
1765                 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
1766                 if (state && state->start == start) {
1767                         /*
1768                          * take a reference on the state, unlock will drop
1769                          * the ref
1770                          */
1771                         cache_state(state, &cached);
1772                 }
1773                 spin_unlock(&tree->lock);
1774
1775                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1776                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1777                                                               state);
1778                         if (ret)
1779                                 uptodate = 0;
1780                 }
1781                 if (!uptodate && tree->ops &&
1782                     tree->ops->readpage_io_failed_hook) {
1783                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1784                                                          start, end, NULL);
1785                         if (ret == 0) {
1786                                 uptodate =
1787                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1788                                 if (err)
1789                                         uptodate = 0;
1790                                 uncache_state(&cached);
1791                                 continue;
1792                         }
1793                 }
1794
1795                 if (uptodate) {
1796                         set_extent_uptodate(tree, start, end, &cached,
1797                                             GFP_ATOMIC);
1798                 }
1799                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
1800
1801                 if (whole_page) {
1802                         if (uptodate) {
1803                                 SetPageUptodate(page);
1804                         } else {
1805                                 ClearPageUptodate(page);
1806                                 SetPageError(page);
1807                         }
1808                         unlock_page(page);
1809                 } else {
1810                         if (uptodate) {
1811                                 check_page_uptodate(tree, page);
1812                         } else {
1813                                 ClearPageUptodate(page);
1814                                 SetPageError(page);
1815                         }
1816                         check_page_locked(tree, page);
1817                 }
1818         } while (bvec <= bvec_end);
1819
1820         bio_put(bio);
1821 }
1822
1823 /*
1824  * IO done from prepare_write is pretty simple, we just unlock
1825  * the structs in the extent tree when done, and set the uptodate bits
1826  * as appropriate.
1827  */
1828 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1829 {
1830         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1831         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1832         struct extent_io_tree *tree;
1833         u64 start;
1834         u64 end;
1835
1836         do {
1837                 struct page *page = bvec->bv_page;
1838                 struct extent_state *cached = NULL;
1839                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1840
1841                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1842                         bvec->bv_offset;
1843                 end = start + bvec->bv_len - 1;
1844
1845                 if (--bvec >= bio->bi_io_vec)
1846                         prefetchw(&bvec->bv_page->flags);
1847
1848                 if (uptodate) {
1849                         set_extent_uptodate(tree, start, end, &cached,
1850                                             GFP_ATOMIC);
1851                 } else {
1852                         ClearPageUptodate(page);
1853                         SetPageError(page);
1854                 }
1855
1856                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
1857
1858         } while (bvec >= bio->bi_io_vec);
1859
1860         bio_put(bio);
1861 }
1862
1863 struct bio *
1864 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1865                 gfp_t gfp_flags)
1866 {
1867         struct bio *bio;
1868
1869         bio = bio_alloc(gfp_flags, nr_vecs);
1870
1871         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1872                 while (!bio && (nr_vecs /= 2))
1873                         bio = bio_alloc(gfp_flags, nr_vecs);
1874         }
1875
1876         if (bio) {
1877                 bio->bi_size = 0;
1878                 bio->bi_bdev = bdev;
1879                 bio->bi_sector = first_sector;
1880         }
1881         return bio;
1882 }
1883
1884 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1885                           unsigned long bio_flags)
1886 {
1887         int ret = 0;
1888         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1889         struct page *page = bvec->bv_page;
1890         struct extent_io_tree *tree = bio->bi_private;
1891         u64 start;
1892
1893         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1894
1895         bio->bi_private = NULL;
1896
1897         bio_get(bio);
1898
1899         if (tree->ops && tree->ops->submit_bio_hook)
1900                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1901                                            mirror_num, bio_flags, start);
1902         else
1903                 submit_bio(rw, bio);
1904         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1905                 ret = -EOPNOTSUPP;
1906         bio_put(bio);
1907         return ret;
1908 }
1909
1910 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1911                               struct page *page, sector_t sector,
1912                               size_t size, unsigned long offset,
1913                               struct block_device *bdev,
1914                               struct bio **bio_ret,
1915                               unsigned long max_pages,
1916                               bio_end_io_t end_io_func,
1917                               int mirror_num,
1918                               unsigned long prev_bio_flags,
1919                               unsigned long bio_flags)
1920 {
1921         int ret = 0;
1922         struct bio *bio;
1923         int nr;
1924         int contig = 0;
1925         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1926         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1927         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1928
1929         if (bio_ret && *bio_ret) {
1930                 bio = *bio_ret;
1931                 if (old_compressed)
1932                         contig = bio->bi_sector == sector;
1933                 else
1934                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
1935                                 sector;
1936
1937                 if (prev_bio_flags != bio_flags || !contig ||
1938                     (tree->ops && tree->ops->merge_bio_hook &&
1939                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
1940                                                bio_flags)) ||
1941                     bio_add_page(bio, page, page_size, offset) < page_size) {
1942                         ret = submit_one_bio(rw, bio, mirror_num,
1943                                              prev_bio_flags);
1944                         bio = NULL;
1945                 } else {
1946                         return 0;
1947                 }
1948         }
1949         if (this_compressed)
1950                 nr = BIO_MAX_PAGES;
1951         else
1952                 nr = bio_get_nr_vecs(bdev);
1953
1954         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1955         if (!bio)
1956                 return -ENOMEM;
1957
1958         bio_add_page(bio, page, page_size, offset);
1959         bio->bi_end_io = end_io_func;
1960         bio->bi_private = tree;
1961
1962         if (bio_ret)
1963                 *bio_ret = bio;
1964         else
1965                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1966
1967         return ret;
1968 }
1969
1970 void set_page_extent_mapped(struct page *page)
1971 {
1972         if (!PagePrivate(page)) {
1973                 SetPagePrivate(page);
1974                 page_cache_get(page);
1975                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1976         }
1977 }
1978
1979 static void set_page_extent_head(struct page *page, unsigned long len)
1980 {
1981         WARN_ON(!PagePrivate(page));
1982         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1983 }
1984
1985 /*
1986  * basic readpage implementation.  Locked extent state structs are inserted
1987  * into the tree that are removed when the IO is done (by the end_io
1988  * handlers)
1989  */
1990 static int __extent_read_full_page(struct extent_io_tree *tree,
1991                                    struct page *page,
1992                                    get_extent_t *get_extent,
1993                                    struct bio **bio, int mirror_num,
1994                                    unsigned long *bio_flags)
1995 {
1996         struct inode *inode = page->mapping->host;
1997         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1998         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1999         u64 end;
2000         u64 cur = start;
2001         u64 extent_offset;
2002         u64 last_byte = i_size_read(inode);
2003         u64 block_start;
2004         u64 cur_end;
2005         sector_t sector;
2006         struct extent_map *em;
2007         struct block_device *bdev;
2008         struct btrfs_ordered_extent *ordered;
2009         int ret;
2010         int nr = 0;
2011         size_t page_offset = 0;
2012         size_t iosize;
2013         size_t disk_io_size;
2014         size_t blocksize = inode->i_sb->s_blocksize;
2015         unsigned long this_bio_flag = 0;
2016
2017         set_page_extent_mapped(page);
2018
2019         end = page_end;
2020         while (1) {
2021                 lock_extent(tree, start, end, GFP_NOFS);
2022                 ordered = btrfs_lookup_ordered_extent(inode, start);
2023                 if (!ordered)
2024                         break;
2025                 unlock_extent(tree, start, end, GFP_NOFS);
2026                 btrfs_start_ordered_extent(inode, ordered, 1);
2027                 btrfs_put_ordered_extent(ordered);
2028         }
2029
2030         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2031                 char *userpage;
2032                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2033
2034                 if (zero_offset) {
2035                         iosize = PAGE_CACHE_SIZE - zero_offset;
2036                         userpage = kmap_atomic(page, KM_USER0);
2037                         memset(userpage + zero_offset, 0, iosize);
2038                         flush_dcache_page(page);
2039                         kunmap_atomic(userpage, KM_USER0);
2040                 }
2041         }
2042         while (cur <= end) {
2043                 if (cur >= last_byte) {
2044                         char *userpage;
2045                         struct extent_state *cached = NULL;
2046
2047                         iosize = PAGE_CACHE_SIZE - page_offset;
2048                         userpage = kmap_atomic(page, KM_USER0);
2049                         memset(userpage + page_offset, 0, iosize);
2050                         flush_dcache_page(page);
2051                         kunmap_atomic(userpage, KM_USER0);
2052                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2053                                             &cached, GFP_NOFS);
2054                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2055                                              &cached, GFP_NOFS);
2056                         break;
2057                 }
2058                 em = get_extent(inode, page, page_offset, cur,
2059                                 end - cur + 1, 0);
2060                 if (IS_ERR(em) || !em) {
2061                         SetPageError(page);
2062                         unlock_extent(tree, cur, end, GFP_NOFS);
2063                         break;
2064                 }
2065                 extent_offset = cur - em->start;
2066                 BUG_ON(extent_map_end(em) <= cur);
2067                 BUG_ON(end < cur);
2068
2069                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2070                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2071                         extent_set_compress_type(&this_bio_flag,
2072                                                  em->compress_type);
2073                 }
2074
2075                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2076                 cur_end = min(extent_map_end(em) - 1, end);
2077                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2078                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2079                         disk_io_size = em->block_len;
2080                         sector = em->block_start >> 9;
2081                 } else {
2082                         sector = (em->block_start + extent_offset) >> 9;
2083                         disk_io_size = iosize;
2084                 }
2085                 bdev = em->bdev;
2086                 block_start = em->block_start;
2087                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2088                         block_start = EXTENT_MAP_HOLE;
2089                 free_extent_map(em);
2090                 em = NULL;
2091
2092                 /* we've found a hole, just zero and go on */
2093                 if (block_start == EXTENT_MAP_HOLE) {
2094                         char *userpage;
2095                         struct extent_state *cached = NULL;
2096
2097                         userpage = kmap_atomic(page, KM_USER0);
2098                         memset(userpage + page_offset, 0, iosize);
2099                         flush_dcache_page(page);
2100                         kunmap_atomic(userpage, KM_USER0);
2101
2102                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2103                                             &cached, GFP_NOFS);
2104                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2105                                              &cached, GFP_NOFS);
2106                         cur = cur + iosize;
2107                         page_offset += iosize;
2108                         continue;
2109                 }
2110                 /* the get_extent function already copied into the page */
2111                 if (test_range_bit(tree, cur, cur_end,
2112                                    EXTENT_UPTODATE, 1, NULL)) {
2113                         check_page_uptodate(tree, page);
2114                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2115                         cur = cur + iosize;
2116                         page_offset += iosize;
2117                         continue;
2118                 }
2119                 /* we have an inline extent but it didn't get marked up
2120                  * to date.  Error out
2121                  */
2122                 if (block_start == EXTENT_MAP_INLINE) {
2123                         SetPageError(page);
2124                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2125                         cur = cur + iosize;
2126                         page_offset += iosize;
2127                         continue;
2128                 }
2129
2130                 ret = 0;
2131                 if (tree->ops && tree->ops->readpage_io_hook) {
2132                         ret = tree->ops->readpage_io_hook(page, cur,
2133                                                           cur + iosize - 1);
2134                 }
2135                 if (!ret) {
2136                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2137                         pnr -= page->index;
2138                         ret = submit_extent_page(READ, tree, page,
2139                                          sector, disk_io_size, page_offset,
2140                                          bdev, bio, pnr,
2141                                          end_bio_extent_readpage, mirror_num,
2142                                          *bio_flags,
2143                                          this_bio_flag);
2144                         nr++;
2145                         *bio_flags = this_bio_flag;
2146                 }
2147                 if (ret)
2148                         SetPageError(page);
2149                 cur = cur + iosize;
2150                 page_offset += iosize;
2151         }
2152         if (!nr) {
2153                 if (!PageError(page))
2154                         SetPageUptodate(page);
2155                 unlock_page(page);
2156         }
2157         return 0;
2158 }
2159
2160 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2161                             get_extent_t *get_extent)
2162 {
2163         struct bio *bio = NULL;
2164         unsigned long bio_flags = 0;
2165         int ret;
2166
2167         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2168                                       &bio_flags);
2169         if (bio)
2170                 ret = submit_one_bio(READ, bio, 0, bio_flags);
2171         return ret;
2172 }
2173
2174 static noinline void update_nr_written(struct page *page,
2175                                       struct writeback_control *wbc,
2176                                       unsigned long nr_written)
2177 {
2178         wbc->nr_to_write -= nr_written;
2179         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2180             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2181                 page->mapping->writeback_index = page->index + nr_written;
2182 }
2183
2184 /*
2185  * the writepage semantics are similar to regular writepage.  extent
2186  * records are inserted to lock ranges in the tree, and as dirty areas
2187  * are found, they are marked writeback.  Then the lock bits are removed
2188  * and the end_io handler clears the writeback ranges
2189  */
2190 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2191                               void *data)
2192 {
2193         struct inode *inode = page->mapping->host;
2194         struct extent_page_data *epd = data;
2195         struct extent_io_tree *tree = epd->tree;
2196         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2197         u64 delalloc_start;
2198         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2199         u64 end;
2200         u64 cur = start;
2201         u64 extent_offset;
2202         u64 last_byte = i_size_read(inode);
2203         u64 block_start;
2204         u64 iosize;
2205         sector_t sector;
2206         struct extent_state *cached_state = NULL;
2207         struct extent_map *em;
2208         struct block_device *bdev;
2209         int ret;
2210         int nr = 0;
2211         size_t pg_offset = 0;
2212         size_t blocksize;
2213         loff_t i_size = i_size_read(inode);
2214         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2215         u64 nr_delalloc;
2216         u64 delalloc_end;
2217         int page_started;
2218         int compressed;
2219         int write_flags;
2220         unsigned long nr_written = 0;
2221
2222         if (wbc->sync_mode == WB_SYNC_ALL)
2223                 write_flags = WRITE_SYNC;
2224         else
2225                 write_flags = WRITE;
2226
2227         trace___extent_writepage(page, inode, wbc);
2228
2229         WARN_ON(!PageLocked(page));
2230         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2231         if (page->index > end_index ||
2232            (page->index == end_index && !pg_offset)) {
2233                 page->mapping->a_ops->invalidatepage(page, 0);
2234                 unlock_page(page);
2235                 return 0;
2236         }
2237
2238         if (page->index == end_index) {
2239                 char *userpage;
2240
2241                 userpage = kmap_atomic(page, KM_USER0);
2242                 memset(userpage + pg_offset, 0,
2243                        PAGE_CACHE_SIZE - pg_offset);
2244                 kunmap_atomic(userpage, KM_USER0);
2245                 flush_dcache_page(page);
2246         }
2247         pg_offset = 0;
2248
2249         set_page_extent_mapped(page);
2250
2251         delalloc_start = start;
2252         delalloc_end = 0;
2253         page_started = 0;
2254         if (!epd->extent_locked) {
2255                 u64 delalloc_to_write = 0;
2256                 /*
2257                  * make sure the wbc mapping index is at least updated
2258                  * to this page.
2259                  */
2260                 update_nr_written(page, wbc, 0);
2261
2262                 while (delalloc_end < page_end) {
2263                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2264                                                        page,
2265                                                        &delalloc_start,
2266                                                        &delalloc_end,
2267                                                        128 * 1024 * 1024);
2268                         if (nr_delalloc == 0) {
2269                                 delalloc_start = delalloc_end + 1;
2270                                 continue;
2271                         }
2272                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2273                                                  delalloc_end, &page_started,
2274                                                  &nr_written);
2275                         /*
2276                          * delalloc_end is already one less than the total
2277                          * length, so we don't subtract one from
2278                          * PAGE_CACHE_SIZE
2279                          */
2280                         delalloc_to_write += (delalloc_end - delalloc_start +
2281                                               PAGE_CACHE_SIZE) >>
2282                                               PAGE_CACHE_SHIFT;
2283                         delalloc_start = delalloc_end + 1;
2284                 }
2285                 if (wbc->nr_to_write < delalloc_to_write) {
2286                         int thresh = 8192;
2287
2288                         if (delalloc_to_write < thresh * 2)
2289                                 thresh = delalloc_to_write;
2290                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2291                                                  thresh);
2292                 }
2293
2294                 /* did the fill delalloc function already unlock and start
2295                  * the IO?
2296                  */
2297                 if (page_started) {
2298                         ret = 0;
2299                         /*
2300                          * we've unlocked the page, so we can't update
2301                          * the mapping's writeback index, just update
2302                          * nr_to_write.
2303                          */
2304                         wbc->nr_to_write -= nr_written;
2305                         goto done_unlocked;
2306                 }
2307         }
2308         if (tree->ops && tree->ops->writepage_start_hook) {
2309                 ret = tree->ops->writepage_start_hook(page, start,
2310                                                       page_end);
2311                 if (ret == -EAGAIN) {
2312                         redirty_page_for_writepage(wbc, page);
2313                         update_nr_written(page, wbc, nr_written);
2314                         unlock_page(page);
2315                         ret = 0;
2316                         goto done_unlocked;
2317                 }
2318         }
2319
2320         /*
2321          * we don't want to touch the inode after unlocking the page,
2322          * so we update the mapping writeback index now
2323          */
2324         update_nr_written(page, wbc, nr_written + 1);
2325
2326         end = page_end;
2327         if (last_byte <= start) {
2328                 if (tree->ops && tree->ops->writepage_end_io_hook)
2329                         tree->ops->writepage_end_io_hook(page, start,
2330                                                          page_end, NULL, 1);
2331                 goto done;
2332         }
2333
2334         blocksize = inode->i_sb->s_blocksize;
2335
2336         while (cur <= end) {
2337                 if (cur >= last_byte) {
2338                         if (tree->ops && tree->ops->writepage_end_io_hook)
2339                                 tree->ops->writepage_end_io_hook(page, cur,
2340                                                          page_end, NULL, 1);
2341                         break;
2342                 }
2343                 em = epd->get_extent(inode, page, pg_offset, cur,
2344                                      end - cur + 1, 1);
2345                 if (IS_ERR(em) || !em) {
2346                         SetPageError(page);
2347                         break;
2348                 }
2349
2350                 extent_offset = cur - em->start;
2351                 BUG_ON(extent_map_end(em) <= cur);
2352                 BUG_ON(end < cur);
2353                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2354                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2355                 sector = (em->block_start + extent_offset) >> 9;
2356                 bdev = em->bdev;
2357                 block_start = em->block_start;
2358                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2359                 free_extent_map(em);
2360                 em = NULL;
2361
2362                 /*
2363                  * compressed and inline extents are written through other
2364                  * paths in the FS
2365                  */
2366                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2367                     block_start == EXTENT_MAP_INLINE) {
2368                         /*
2369                          * end_io notification does not happen here for
2370                          * compressed extents
2371                          */
2372                         if (!compressed && tree->ops &&
2373                             tree->ops->writepage_end_io_hook)
2374                                 tree->ops->writepage_end_io_hook(page, cur,
2375                                                          cur + iosize - 1,
2376                                                          NULL, 1);
2377                         else if (compressed) {
2378                                 /* we don't want to end_page_writeback on
2379                                  * a compressed extent.  this happens
2380                                  * elsewhere
2381                                  */
2382                                 nr++;
2383                         }
2384
2385                         cur += iosize;
2386                         pg_offset += iosize;
2387                         continue;
2388                 }
2389                 /* leave this out until we have a page_mkwrite call */
2390                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2391                                    EXTENT_DIRTY, 0, NULL)) {
2392                         cur = cur + iosize;
2393                         pg_offset += iosize;
2394                         continue;
2395                 }
2396
2397                 if (tree->ops && tree->ops->writepage_io_hook) {
2398                         ret = tree->ops->writepage_io_hook(page, cur,
2399                                                 cur + iosize - 1);
2400                 } else {
2401                         ret = 0;
2402                 }
2403                 if (ret) {
2404                         SetPageError(page);
2405                 } else {
2406                         unsigned long max_nr = end_index + 1;
2407
2408                         set_range_writeback(tree, cur, cur + iosize - 1);
2409                         if (!PageWriteback(page)) {
2410                                 printk(KERN_ERR "btrfs warning page %lu not "
2411                                        "writeback, cur %llu end %llu\n",
2412                                        page->index, (unsigned long long)cur,
2413                                        (unsigned long long)end);
2414                         }
2415
2416                         ret = submit_extent_page(write_flags, tree, page,
2417                                                  sector, iosize, pg_offset,
2418                                                  bdev, &epd->bio, max_nr,
2419                                                  end_bio_extent_writepage,
2420                                                  0, 0, 0);
2421                         if (ret)
2422                                 SetPageError(page);
2423                 }
2424                 cur = cur + iosize;
2425                 pg_offset += iosize;
2426                 nr++;
2427         }
2428 done:
2429         if (nr == 0) {
2430                 /* make sure the mapping tag for page dirty gets cleared */
2431                 set_page_writeback(page);
2432                 end_page_writeback(page);
2433         }
2434         unlock_page(page);
2435
2436 done_unlocked:
2437
2438         /* drop our reference on any cached states */
2439         free_extent_state(cached_state);
2440         return 0;
2441 }
2442
2443 /**
2444  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2445  * @mapping: address space structure to write
2446  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2447  * @writepage: function called for each page
2448  * @data: data passed to writepage function
2449  *
2450  * If a page is already under I/O, write_cache_pages() skips it, even
2451  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2452  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2453  * and msync() need to guarantee that all the data which was dirty at the time
2454  * the call was made get new I/O started against them.  If wbc->sync_mode is
2455  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2456  * existing IO to complete.
2457  */
2458 static int extent_write_cache_pages(struct extent_io_tree *tree,
2459                              struct address_space *mapping,
2460                              struct writeback_control *wbc,
2461                              writepage_t writepage, void *data,
2462                              void (*flush_fn)(void *))
2463 {
2464         int ret = 0;
2465         int done = 0;
2466         int nr_to_write_done = 0;
2467         struct pagevec pvec;
2468         int nr_pages;
2469         pgoff_t index;
2470         pgoff_t end;            /* Inclusive */
2471         int scanned = 0;
2472
2473         pagevec_init(&pvec, 0);
2474         if (wbc->range_cyclic) {
2475                 index = mapping->writeback_index; /* Start from prev offset */
2476                 end = -1;
2477         } else {
2478                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2479                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2480                 scanned = 1;
2481         }
2482 retry:
2483         while (!done && !nr_to_write_done && (index <= end) &&
2484                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2485                               PAGECACHE_TAG_DIRTY, min(end - index,
2486                                   (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2487                 unsigned i;
2488
2489                 scanned = 1;
2490                 for (i = 0; i < nr_pages; i++) {
2491                         struct page *page = pvec.pages[i];
2492
2493                         /*
2494                          * At this point we hold neither mapping->tree_lock nor
2495                          * lock on the page itself: the page may be truncated or
2496                          * invalidated (changing page->mapping to NULL), or even
2497                          * swizzled back from swapper_space to tmpfs file
2498                          * mapping
2499                          */
2500                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2501                                 tree->ops->write_cache_pages_lock_hook(page);
2502                         else
2503                                 lock_page(page);
2504
2505                         if (unlikely(page->mapping != mapping)) {
2506                                 unlock_page(page);
2507                                 continue;
2508                         }
2509
2510                         if (!wbc->range_cyclic && page->index > end) {
2511                                 done = 1;
2512                                 unlock_page(page);
2513                                 continue;
2514                         }
2515
2516                         if (wbc->sync_mode != WB_SYNC_NONE) {
2517                                 if (PageWriteback(page))
2518                                         flush_fn(data);
2519                                 wait_on_page_writeback(page);
2520                         }
2521
2522                         if (PageWriteback(page) ||
2523                             !clear_page_dirty_for_io(page)) {
2524                                 unlock_page(page);
2525                                 continue;
2526                         }
2527
2528                         ret = (*writepage)(page, wbc, data);
2529
2530                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2531                                 unlock_page(page);
2532                                 ret = 0;
2533                         }
2534                         if (ret)
2535                                 done = 1;
2536
2537                         /*
2538                          * the filesystem may choose to bump up nr_to_write.
2539                          * We have to make sure to honor the new nr_to_write
2540                          * at any time
2541                          */
2542                         nr_to_write_done = wbc->nr_to_write <= 0;
2543                 }
2544                 pagevec_release(&pvec);
2545                 cond_resched();
2546         }
2547         if (!scanned && !done) {
2548                 /*
2549                  * We hit the last page and there is more work to be done: wrap
2550                  * back to the start of the file
2551                  */
2552                 scanned = 1;
2553                 index = 0;
2554                 goto retry;
2555         }
2556         return ret;
2557 }
2558
2559 static void flush_epd_write_bio(struct extent_page_data *epd)
2560 {
2561         if (epd->bio) {
2562                 if (epd->sync_io)
2563                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2564                 else
2565                         submit_one_bio(WRITE, epd->bio, 0, 0);
2566                 epd->bio = NULL;
2567         }
2568 }
2569
2570 static noinline void flush_write_bio(void *data)
2571 {
2572         struct extent_page_data *epd = data;
2573         flush_epd_write_bio(epd);
2574 }
2575
2576 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2577                           get_extent_t *get_extent,
2578                           struct writeback_control *wbc)
2579 {
2580         int ret;
2581         struct address_space *mapping = page->mapping;
2582         struct extent_page_data epd = {
2583                 .bio = NULL,
2584                 .tree = tree,
2585                 .get_extent = get_extent,
2586                 .extent_locked = 0,
2587                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2588         };
2589         struct writeback_control wbc_writepages = {
2590                 .sync_mode      = wbc->sync_mode,
2591                 .older_than_this = NULL,
2592                 .nr_to_write    = 64,
2593                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2594                 .range_end      = (loff_t)-1,
2595         };
2596
2597         ret = __extent_writepage(page, wbc, &epd);
2598
2599         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2600                                  __extent_writepage, &epd, flush_write_bio);
2601         flush_epd_write_bio(&epd);
2602         return ret;
2603 }
2604
2605 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2606                               u64 start, u64 end, get_extent_t *get_extent,
2607                               int mode)
2608 {
2609         int ret = 0;
2610         struct address_space *mapping = inode->i_mapping;
2611         struct page *page;
2612         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2613                 PAGE_CACHE_SHIFT;
2614
2615         struct extent_page_data epd = {
2616                 .bio = NULL,
2617                 .tree = tree,
2618                 .get_extent = get_extent,
2619                 .extent_locked = 1,
2620                 .sync_io = mode == WB_SYNC_ALL,
2621         };
2622         struct writeback_control wbc_writepages = {
2623                 .sync_mode      = mode,
2624                 .older_than_this = NULL,
2625                 .nr_to_write    = nr_pages * 2,
2626                 .range_start    = start,
2627                 .range_end      = end + 1,
2628         };
2629
2630         while (start <= end) {
2631                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2632                 if (clear_page_dirty_for_io(page))
2633                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2634                 else {
2635                         if (tree->ops && tree->ops->writepage_end_io_hook)
2636                                 tree->ops->writepage_end_io_hook(page, start,
2637                                                  start + PAGE_CACHE_SIZE - 1,
2638                                                  NULL, 1);
2639                         unlock_page(page);
2640                 }
2641                 page_cache_release(page);
2642                 start += PAGE_CACHE_SIZE;
2643         }
2644
2645         flush_epd_write_bio(&epd);
2646         return ret;
2647 }
2648
2649 int extent_writepages(struct extent_io_tree *tree,
2650                       struct address_space *mapping,
2651                       get_extent_t *get_extent,
2652                       struct writeback_control *wbc)
2653 {
2654         int ret = 0;
2655         struct extent_page_data epd = {
2656                 .bio = NULL,
2657                 .tree = tree,
2658                 .get_extent = get_extent,
2659                 .extent_locked = 0,
2660                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2661         };
2662
2663         ret = extent_write_cache_pages(tree, mapping, wbc,
2664                                        __extent_writepage, &epd,
2665                                        flush_write_bio);
2666         flush_epd_write_bio(&epd);
2667         return ret;
2668 }
2669
2670 int extent_readpages(struct extent_io_tree *tree,
2671                      struct address_space *mapping,
2672                      struct list_head *pages, unsigned nr_pages,
2673                      get_extent_t get_extent)
2674 {
2675         struct bio *bio = NULL;
2676         unsigned page_idx;
2677         unsigned long bio_flags = 0;
2678
2679         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2680                 struct page *page = list_entry(pages->prev, struct page, lru);
2681
2682                 prefetchw(&page->flags);
2683                 list_del(&page->lru);
2684                 if (!add_to_page_cache_lru(page, mapping,
2685                                         page->index, GFP_NOFS)) {
2686                         __extent_read_full_page(tree, page, get_extent,
2687                                                 &bio, 0, &bio_flags);
2688                 }
2689                 page_cache_release(page);
2690         }
2691         BUG_ON(!list_empty(pages));
2692         if (bio)
2693                 submit_one_bio(READ, bio, 0, bio_flags);
2694         return 0;
2695 }
2696
2697 /*
2698  * basic invalidatepage code, this waits on any locked or writeback
2699  * ranges corresponding to the page, and then deletes any extent state
2700  * records from the tree
2701  */
2702 int extent_invalidatepage(struct extent_io_tree *tree,
2703                           struct page *page, unsigned long offset)
2704 {
2705         struct extent_state *cached_state = NULL;
2706         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2707         u64 end = start + PAGE_CACHE_SIZE - 1;
2708         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2709
2710         start += (offset + blocksize - 1) & ~(blocksize - 1);
2711         if (start > end)
2712                 return 0;
2713
2714         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
2715         wait_on_page_writeback(page);
2716         clear_extent_bit(tree, start, end,
2717                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2718                          EXTENT_DO_ACCOUNTING,
2719                          1, 1, &cached_state, GFP_NOFS);
2720         return 0;
2721 }
2722
2723 /*
2724  * simple commit_write call, set_range_dirty is used to mark both
2725  * the pages and the extent records as dirty
2726  */
2727 int extent_commit_write(struct extent_io_tree *tree,
2728                         struct inode *inode, struct page *page,
2729                         unsigned from, unsigned to)
2730 {
2731         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2732
2733         set_page_extent_mapped(page);
2734         set_page_dirty(page);
2735
2736         if (pos > inode->i_size) {
2737                 i_size_write(inode, pos);
2738                 mark_inode_dirty(inode);
2739         }
2740         return 0;
2741 }
2742
2743 int extent_prepare_write(struct extent_io_tree *tree,
2744                          struct inode *inode, struct page *page,
2745                          unsigned from, unsigned to, get_extent_t *get_extent)
2746 {
2747         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2748         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2749         u64 block_start;
2750         u64 orig_block_start;
2751         u64 block_end;
2752         u64 cur_end;
2753         struct extent_map *em;
2754         unsigned blocksize = 1 << inode->i_blkbits;
2755         size_t page_offset = 0;
2756         size_t block_off_start;
2757         size_t block_off_end;
2758         int err = 0;
2759         int iocount = 0;
2760         int ret = 0;
2761         int isnew;
2762
2763         set_page_extent_mapped(page);
2764
2765         block_start = (page_start + from) & ~((u64)blocksize - 1);
2766         block_end = (page_start + to - 1) | (blocksize - 1);
2767         orig_block_start = block_start;
2768
2769         lock_extent(tree, page_start, page_end, GFP_NOFS);
2770         while (block_start <= block_end) {
2771                 em = get_extent(inode, page, page_offset, block_start,
2772                                 block_end - block_start + 1, 1);
2773                 if (IS_ERR(em) || !em)
2774                         goto err;
2775
2776                 cur_end = min(block_end, extent_map_end(em) - 1);
2777                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2778                 block_off_end = block_off_start + blocksize;
2779                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2780
2781                 if (!PageUptodate(page) && isnew &&
2782                     (block_off_end > to || block_off_start < from)) {
2783                         void *kaddr;
2784
2785                         kaddr = kmap_atomic(page, KM_USER0);
2786                         if (block_off_end > to)
2787                                 memset(kaddr + to, 0, block_off_end - to);
2788                         if (block_off_start < from)
2789                                 memset(kaddr + block_off_start, 0,
2790                                        from - block_off_start);
2791                         flush_dcache_page(page);
2792                         kunmap_atomic(kaddr, KM_USER0);
2793                 }
2794                 if ((em->block_start != EXTENT_MAP_HOLE &&
2795                      em->block_start != EXTENT_MAP_INLINE) &&
2796                     !isnew && !PageUptodate(page) &&
2797                     (block_off_end > to || block_off_start < from) &&
2798                     !test_range_bit(tree, block_start, cur_end,
2799                                     EXTENT_UPTODATE, 1, NULL)) {
2800                         u64 sector;
2801                         u64 extent_offset = block_start - em->start;
2802                         size_t iosize;
2803                         sector = (em->block_start + extent_offset) >> 9;
2804                         iosize = (cur_end - block_start + blocksize) &
2805                                 ~((u64)blocksize - 1);
2806                         /*
2807                          * we've already got the extent locked, but we
2808                          * need to split the state such that our end_bio
2809                          * handler can clear the lock.
2810                          */
2811                         set_extent_bit(tree, block_start,
2812                                        block_start + iosize - 1,
2813                                        EXTENT_LOCKED, 0, NULL, NULL, GFP_NOFS);
2814                         ret = submit_extent_page(READ, tree, page,
2815                                          sector, iosize, page_offset, em->bdev,
2816                                          NULL, 1,
2817                                          end_bio_extent_preparewrite, 0,
2818                                          0, 0);
2819                         if (ret && !err)
2820                                 err = ret;
2821                         iocount++;
2822                         block_start = block_start + iosize;
2823                 } else {
2824                         struct extent_state *cached = NULL;
2825
2826                         set_extent_uptodate(tree, block_start, cur_end, &cached,
2827                                             GFP_NOFS);
2828                         unlock_extent_cached(tree, block_start, cur_end,
2829                                              &cached, GFP_NOFS);
2830                         block_start = cur_end + 1;
2831                 }
2832                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2833                 free_extent_map(em);
2834         }
2835         if (iocount) {
2836                 wait_extent_bit(tree, orig_block_start,
2837                                 block_end, EXTENT_LOCKED);
2838         }
2839         check_page_uptodate(tree, page);
2840 err:
2841         /* FIXME, zero out newly allocated blocks on error */
2842         return err;
2843 }
2844
2845 /*
2846  * a helper for releasepage, this tests for areas of the page that
2847  * are locked or under IO and drops the related state bits if it is safe
2848  * to drop the page.
2849  */
2850 int try_release_extent_state(struct extent_map_tree *map,
2851                              struct extent_io_tree *tree, struct page *page,
2852                              gfp_t mask)
2853 {
2854         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2855         u64 end = start + PAGE_CACHE_SIZE - 1;
2856         int ret = 1;
2857
2858         if (test_range_bit(tree, start, end,
2859                            EXTENT_IOBITS, 0, NULL))
2860                 ret = 0;
2861         else {
2862                 if ((mask & GFP_NOFS) == GFP_NOFS)
2863                         mask = GFP_NOFS;
2864                 /*
2865                  * at this point we can safely clear everything except the
2866                  * locked bit and the nodatasum bit
2867                  */
2868                 ret = clear_extent_bit(tree, start, end,
2869                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2870                                  0, 0, NULL, mask);
2871
2872                 /* if clear_extent_bit failed for enomem reasons,
2873                  * we can't allow the release to continue.
2874                  */
2875                 if (ret < 0)
2876                         ret = 0;
2877                 else
2878                         ret = 1;
2879         }
2880         return ret;
2881 }
2882
2883 /*
2884  * a helper for releasepage.  As long as there are no locked extents
2885  * in the range corresponding to the page, both state records and extent
2886  * map records are removed
2887  */
2888 int try_release_extent_mapping(struct extent_map_tree *map,
2889                                struct extent_io_tree *tree, struct page *page,
2890                                gfp_t mask)
2891 {
2892         struct extent_map *em;
2893         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2894         u64 end = start + PAGE_CACHE_SIZE - 1;
2895
2896         if ((mask & __GFP_WAIT) &&
2897             page->mapping->host->i_size > 16 * 1024 * 1024) {
2898                 u64 len;
2899                 while (start <= end) {
2900                         len = end - start + 1;
2901                         write_lock(&map->lock);
2902                         em = lookup_extent_mapping(map, start, len);
2903                         if (!em || IS_ERR(em)) {
2904                                 write_unlock(&map->lock);
2905                                 break;
2906                         }
2907                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2908                             em->start != start) {
2909                                 write_unlock(&map->lock);
2910                                 free_extent_map(em);
2911                                 break;
2912                         }
2913                         if (!test_range_bit(tree, em->start,
2914                                             extent_map_end(em) - 1,
2915                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
2916                                             0, NULL)) {
2917                                 remove_extent_mapping(map, em);
2918                                 /* once for the rb tree */
2919                                 free_extent_map(em);
2920                         }
2921                         start = extent_map_end(em);
2922                         write_unlock(&map->lock);
2923
2924                         /* once for us */
2925                         free_extent_map(em);
2926                 }
2927         }
2928         return try_release_extent_state(map, tree, page, mask);
2929 }
2930
2931 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2932                 get_extent_t *get_extent)
2933 {
2934         struct inode *inode = mapping->host;
2935         struct extent_state *cached_state = NULL;
2936         u64 start = iblock << inode->i_blkbits;
2937         sector_t sector = 0;
2938         size_t blksize = (1 << inode->i_blkbits);
2939         struct extent_map *em;
2940
2941         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2942                          0, &cached_state, GFP_NOFS);
2943         em = get_extent(inode, NULL, 0, start, blksize, 0);
2944         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start,
2945                              start + blksize - 1, &cached_state, GFP_NOFS);
2946         if (!em || IS_ERR(em))
2947                 return 0;
2948
2949         if (em->block_start > EXTENT_MAP_LAST_BYTE)
2950                 goto out;
2951
2952         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2953 out:
2954         free_extent_map(em);
2955         return sector;
2956 }
2957
2958 /*
2959  * helper function for fiemap, which doesn't want to see any holes.
2960  * This maps until we find something past 'last'
2961  */
2962 static struct extent_map *get_extent_skip_holes(struct inode *inode,
2963                                                 u64 offset,
2964                                                 u64 last,
2965                                                 get_extent_t *get_extent)
2966 {
2967         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
2968         struct extent_map *em;
2969         u64 len;
2970
2971         if (offset >= last)
2972                 return NULL;
2973
2974         while(1) {
2975                 len = last - offset;
2976                 if (len == 0)
2977                         break;
2978                 len = (len + sectorsize - 1) & ~(sectorsize - 1);
2979                 em = get_extent(inode, NULL, 0, offset, len, 0);
2980                 if (!em || IS_ERR(em))
2981                         return em;
2982
2983                 /* if this isn't a hole return it */
2984                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
2985                     em->block_start != EXTENT_MAP_HOLE) {
2986                         return em;
2987                 }
2988
2989                 /* this is a hole, advance to the next extent */
2990                 offset = extent_map_end(em);
2991                 free_extent_map(em);
2992                 if (offset >= last)
2993                         break;
2994         }
2995         return NULL;
2996 }
2997
2998 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2999                 __u64 start, __u64 len, get_extent_t *get_extent)
3000 {
3001         int ret = 0;
3002         u64 off = start;
3003         u64 max = start + len;
3004         u32 flags = 0;
3005         u32 found_type;
3006         u64 last;
3007         u64 last_for_get_extent = 0;
3008         u64 disko = 0;
3009         u64 isize = i_size_read(inode);
3010         struct btrfs_key found_key;
3011         struct extent_map *em = NULL;
3012         struct extent_state *cached_state = NULL;
3013         struct btrfs_path *path;
3014         struct btrfs_file_extent_item *item;
3015         int end = 0;
3016         u64 em_start = 0;
3017         u64 em_len = 0;
3018         u64 em_end = 0;
3019         unsigned long emflags;
3020
3021         if (len == 0)
3022                 return -EINVAL;
3023
3024         path = btrfs_alloc_path();
3025         if (!path)
3026                 return -ENOMEM;
3027         path->leave_spinning = 1;
3028
3029         /*
3030          * lookup the last file extent.  We're not using i_size here
3031          * because there might be preallocation past i_size
3032          */
3033         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3034                                        path, inode->i_ino, -1, 0);
3035         if (ret < 0) {
3036                 btrfs_free_path(path);
3037                 return ret;
3038         }
3039         WARN_ON(!ret);
3040         path->slots[0]--;
3041         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3042                               struct btrfs_file_extent_item);
3043         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3044         found_type = btrfs_key_type(&found_key);
3045
3046         /* No extents, but there might be delalloc bits */
3047         if (found_key.objectid != inode->i_ino ||
3048             found_type != BTRFS_EXTENT_DATA_KEY) {
3049                 /* have to trust i_size as the end */
3050                 last = (u64)-1;
3051                 last_for_get_extent = isize;
3052         } else {
3053                 /*
3054                  * remember the start of the last extent.  There are a
3055                  * bunch of different factors that go into the length of the
3056                  * extent, so its much less complex to remember where it started
3057                  */
3058                 last = found_key.offset;
3059                 last_for_get_extent = last + 1;
3060         }
3061         btrfs_free_path(path);
3062
3063         /*
3064          * we might have some extents allocated but more delalloc past those
3065          * extents.  so, we trust isize unless the start of the last extent is
3066          * beyond isize
3067          */
3068         if (last < isize) {
3069                 last = (u64)-1;
3070                 last_for_get_extent = isize;
3071         }
3072
3073         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3074                          &cached_state, GFP_NOFS);
3075
3076         em = get_extent_skip_holes(inode, off, last_for_get_extent,
3077                                    get_extent);
3078         if (!em)
3079                 goto out;
3080         if (IS_ERR(em)) {
3081                 ret = PTR_ERR(em);
3082                 goto out;
3083         }
3084
3085         while (!end) {
3086                 u64 offset_in_extent;
3087
3088                 /* break if the extent we found is outside the range */
3089                 if (em->start >= max || extent_map_end(em) < off)
3090                         break;
3091
3092                 /*
3093                  * get_extent may return an extent that starts before our
3094                  * requested range.  We have to make sure the ranges
3095                  * we return to fiemap always move forward and don't
3096                  * overlap, so adjust the offsets here
3097                  */
3098                 em_start = max(em->start, off);
3099
3100                 /*
3101                  * record the offset from the start of the extent
3102                  * for adjusting the disk offset below
3103                  */
3104                 offset_in_extent = em_start - em->start;
3105                 em_end = extent_map_end(em);
3106                 em_len = em_end - em_start;
3107                 emflags = em->flags;
3108                 disko = 0;
3109                 flags = 0;
3110
3111                 /*
3112                  * bump off for our next call to get_extent
3113                  */
3114                 off = extent_map_end(em);
3115                 if (off >= max)
3116                         end = 1;
3117
3118                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3119                         end = 1;
3120                         flags |= FIEMAP_EXTENT_LAST;
3121                 } else if (em->block_start == EXTENT_MAP_INLINE) {
3122                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
3123                                   FIEMAP_EXTENT_NOT_ALIGNED);
3124                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3125                         flags |= (FIEMAP_EXTENT_DELALLOC |
3126                                   FIEMAP_EXTENT_UNKNOWN);
3127                 } else {
3128                         disko = em->block_start + offset_in_extent;
3129                 }
3130                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3131                         flags |= FIEMAP_EXTENT_ENCODED;
3132
3133                 free_extent_map(em);
3134                 em = NULL;
3135                 if ((em_start >= last) || em_len == (u64)-1 ||
3136                    (last == (u64)-1 && isize <= em_end)) {
3137                         flags |= FIEMAP_EXTENT_LAST;
3138                         end = 1;
3139                 }
3140
3141                 /* now scan forward to see if this is really the last extent. */
3142                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
3143                                            get_extent);
3144                 if (IS_ERR(em)) {
3145                         ret = PTR_ERR(em);
3146                         goto out;
3147                 }
3148                 if (!em) {
3149                         flags |= FIEMAP_EXTENT_LAST;
3150                         end = 1;
3151                 }
3152                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3153                                               em_len, flags);
3154                 if (ret)
3155                         goto out_free;
3156         }
3157 out_free:
3158         free_extent_map(em);
3159 out:
3160         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3161                              &cached_state, GFP_NOFS);
3162         return ret;
3163 }
3164
3165 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
3166                                               unsigned long i)
3167 {
3168         struct page *p;
3169         struct address_space *mapping;
3170
3171         if (i == 0)
3172                 return eb->first_page;
3173         i += eb->start >> PAGE_CACHE_SHIFT;
3174         mapping = eb->first_page->mapping;
3175         if (!mapping)
3176                 return NULL;
3177
3178         /*
3179          * extent_buffer_page is only called after pinning the page
3180          * by increasing the reference count.  So we know the page must
3181          * be in the radix tree.
3182          */
3183         rcu_read_lock();
3184         p = radix_tree_lookup(&mapping->page_tree, i);
3185         rcu_read_unlock();
3186
3187         return p;
3188 }
3189
3190 static inline unsigned long num_extent_pages(u64 start, u64 len)
3191 {
3192         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3193                 (start >> PAGE_CACHE_SHIFT);
3194 }
3195
3196 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3197                                                    u64 start,
3198                                                    unsigned long len,
3199                                                    gfp_t mask)
3200 {
3201         struct extent_buffer *eb = NULL;
3202 #if LEAK_DEBUG
3203         unsigned long flags;
3204 #endif
3205
3206         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3207         if (eb == NULL)
3208                 return NULL;
3209         eb->start = start;
3210         eb->len = len;
3211         spin_lock_init(&eb->lock);
3212         init_waitqueue_head(&eb->lock_wq);
3213
3214 #if LEAK_DEBUG
3215         spin_lock_irqsave(&leak_lock, flags);
3216         list_add(&eb->leak_list, &buffers);
3217         spin_unlock_irqrestore(&leak_lock, flags);
3218 #endif
3219         atomic_set(&eb->refs, 1);
3220
3221         return eb;
3222 }
3223
3224 static void __free_extent_buffer(struct extent_buffer *eb)
3225 {
3226 #if LEAK_DEBUG
3227         unsigned long flags;
3228         spin_lock_irqsave(&leak_lock, flags);
3229         list_del(&eb->leak_list);
3230         spin_unlock_irqrestore(&leak_lock, flags);
3231 #endif
3232         kmem_cache_free(extent_buffer_cache, eb);
3233 }
3234
3235 /*
3236  * Helper for releasing extent buffer page.
3237  */
3238 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3239                                                 unsigned long start_idx)
3240 {
3241         unsigned long index;
3242         struct page *page;
3243
3244         if (!eb->first_page)
3245                 return;
3246
3247         index = num_extent_pages(eb->start, eb->len);
3248         if (start_idx >= index)
3249                 return;
3250
3251         do {
3252                 index--;
3253                 page = extent_buffer_page(eb, index);
3254                 if (page)
3255                         page_cache_release(page);
3256         } while (index != start_idx);
3257 }
3258
3259 /*
3260  * Helper for releasing the extent buffer.
3261  */
3262 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3263 {
3264         btrfs_release_extent_buffer_page(eb, 0);
3265         __free_extent_buffer(eb);
3266 }
3267
3268 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3269                                           u64 start, unsigned long len,
3270                                           struct page *page0,
3271                                           gfp_t mask)
3272 {
3273         unsigned long num_pages = num_extent_pages(start, len);
3274         unsigned long i;
3275         unsigned long index = start >> PAGE_CACHE_SHIFT;
3276         struct extent_buffer *eb;
3277         struct extent_buffer *exists = NULL;
3278         struct page *p;
3279         struct address_space *mapping = tree->mapping;
3280         int uptodate = 1;
3281         int ret;
3282
3283         rcu_read_lock();
3284         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3285         if (eb && atomic_inc_not_zero(&eb->refs)) {
3286                 rcu_read_unlock();
3287                 mark_page_accessed(eb->first_page);
3288                 return eb;
3289         }
3290         rcu_read_unlock();
3291
3292         eb = __alloc_extent_buffer(tree, start, len, mask);
3293         if (!eb)
3294                 return NULL;
3295
3296         if (page0) {
3297                 eb->first_page = page0;
3298                 i = 1;
3299                 index++;
3300                 page_cache_get(page0);
3301                 mark_page_accessed(page0);
3302                 set_page_extent_mapped(page0);
3303                 set_page_extent_head(page0, len);
3304                 uptodate = PageUptodate(page0);
3305         } else {
3306                 i = 0;
3307         }
3308         for (; i < num_pages; i++, index++) {
3309                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
3310                 if (!p) {
3311                         WARN_ON(1);
3312                         goto free_eb;
3313                 }
3314                 set_page_extent_mapped(p);
3315                 mark_page_accessed(p);
3316                 if (i == 0) {
3317                         eb->first_page = p;
3318                         set_page_extent_head(p, len);
3319                 } else {
3320                         set_page_private(p, EXTENT_PAGE_PRIVATE);
3321                 }
3322                 if (!PageUptodate(p))
3323                         uptodate = 0;
3324
3325                 /*
3326                  * see below about how we avoid a nasty race with release page
3327                  * and why we unlock later
3328                  */
3329                 if (i != 0)
3330                         unlock_page(p);
3331         }
3332         if (uptodate)
3333                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3334
3335         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
3336         if (ret)
3337                 goto free_eb;
3338
3339         spin_lock(&tree->buffer_lock);
3340         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
3341         if (ret == -EEXIST) {
3342                 exists = radix_tree_lookup(&tree->buffer,
3343                                                 start >> PAGE_CACHE_SHIFT);
3344                 /* add one reference for the caller */
3345                 atomic_inc(&exists->refs);
3346                 spin_unlock(&tree->buffer_lock);
3347                 radix_tree_preload_end();
3348                 goto free_eb;
3349         }
3350         /* add one reference for the tree */
3351         atomic_inc(&eb->refs);
3352         spin_unlock(&tree->buffer_lock);
3353         radix_tree_preload_end();
3354
3355         /*
3356          * there is a race where release page may have
3357          * tried to find this extent buffer in the radix
3358          * but failed.  It will tell the VM it is safe to
3359          * reclaim the, and it will clear the page private bit.
3360          * We must make sure to set the page private bit properly
3361          * after the extent buffer is in the radix tree so
3362          * it doesn't get lost
3363          */
3364         set_page_extent_mapped(eb->first_page);
3365         set_page_extent_head(eb->first_page, eb->len);
3366         if (!page0)
3367                 unlock_page(eb->first_page);
3368         return eb;
3369
3370 free_eb:
3371         if (eb->first_page && !page0)
3372                 unlock_page(eb->first_page);
3373
3374         if (!atomic_dec_and_test(&eb->refs))
3375                 return exists;
3376         btrfs_release_extent_buffer(eb);
3377         return exists;
3378 }
3379
3380 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3381                                          u64 start, unsigned long len,
3382                                           gfp_t mask)
3383 {
3384         struct extent_buffer *eb;
3385
3386         rcu_read_lock();
3387         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3388         if (eb && atomic_inc_not_zero(&eb->refs)) {
3389                 rcu_read_unlock();
3390                 mark_page_accessed(eb->first_page);
3391                 return eb;
3392         }
3393         rcu_read_unlock();
3394
3395         return NULL;
3396 }
3397
3398 void free_extent_buffer(struct extent_buffer *eb)
3399 {
3400         if (!eb)
3401                 return;
3402
3403         if (!atomic_dec_and_test(&eb->refs))
3404                 return;
3405
3406         WARN_ON(1);
3407 }
3408
3409 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3410                               struct extent_buffer *eb)
3411 {
3412         unsigned long i;
3413         unsigned long num_pages;
3414         struct page *page;
3415
3416         num_pages = num_extent_pages(eb->start, eb->len);
3417
3418         for (i = 0; i < num_pages; i++) {
3419                 page = extent_buffer_page(eb, i);
3420                 if (!PageDirty(page))
3421                         continue;
3422
3423                 lock_page(page);
3424                 WARN_ON(!PagePrivate(page));
3425
3426                 set_page_extent_mapped(page);
3427                 if (i == 0)
3428                         set_page_extent_head(page, eb->len);
3429
3430                 clear_page_dirty_for_io(page);
3431                 spin_lock_irq(&page->mapping->tree_lock);
3432                 if (!PageDirty(page)) {
3433                         radix_tree_tag_clear(&page->mapping->page_tree,
3434                                                 page_index(page),
3435                                                 PAGECACHE_TAG_DIRTY);
3436                 }
3437                 spin_unlock_irq(&page->mapping->tree_lock);
3438                 unlock_page(page);
3439         }
3440         return 0;
3441 }
3442
3443 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
3444                                     struct extent_buffer *eb)
3445 {
3446         return wait_on_extent_writeback(tree, eb->start,
3447                                         eb->start + eb->len - 1);
3448 }
3449
3450 int set_extent_buffer_dirty(struct extent_io_tree *tree,
3451                              struct extent_buffer *eb)
3452 {
3453         unsigned long i;
3454         unsigned long num_pages;
3455         int was_dirty = 0;
3456
3457         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3458         num_pages = num_extent_pages(eb->start, eb->len);
3459         for (i = 0; i < num_pages; i++)
3460                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3461         return was_dirty;
3462 }
3463
3464 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3465                                 struct extent_buffer *eb,
3466                                 struct extent_state **cached_state)
3467 {
3468         unsigned long i;
3469         struct page *page;
3470         unsigned long num_pages;
3471
3472         num_pages = num_extent_pages(eb->start, eb->len);
3473         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3474
3475         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3476                               cached_state, GFP_NOFS);
3477         for (i = 0; i < num_pages; i++) {
3478                 page = extent_buffer_page(eb, i);
3479                 if (page)
3480                         ClearPageUptodate(page);
3481         }
3482         return 0;
3483 }
3484
3485 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3486                                 struct extent_buffer *eb)
3487 {
3488         unsigned long i;
3489         struct page *page;
3490         unsigned long num_pages;
3491
3492         num_pages = num_extent_pages(eb->start, eb->len);
3493
3494         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3495                             NULL, GFP_NOFS);
3496         for (i = 0; i < num_pages; i++) {
3497                 page = extent_buffer_page(eb, i);
3498                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3499                     ((i == num_pages - 1) &&
3500                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3501                         check_page_uptodate(tree, page);
3502                         continue;
3503                 }
3504                 SetPageUptodate(page);
3505         }
3506         return 0;
3507 }
3508
3509 int extent_range_uptodate(struct extent_io_tree *tree,
3510                           u64 start, u64 end)
3511 {
3512         struct page *page;
3513         int ret;
3514         int pg_uptodate = 1;
3515         int uptodate;
3516         unsigned long index;
3517
3518         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL);
3519         if (ret)
3520                 return 1;
3521         while (start <= end) {
3522                 index = start >> PAGE_CACHE_SHIFT;
3523                 page = find_get_page(tree->mapping, index);
3524                 uptodate = PageUptodate(page);
3525                 page_cache_release(page);
3526                 if (!uptodate) {
3527                         pg_uptodate = 0;
3528                         break;
3529                 }
3530                 start += PAGE_CACHE_SIZE;
3531         }
3532         return pg_uptodate;
3533 }
3534
3535 int extent_buffer_uptodate(struct extent_io_tree *tree,
3536                            struct extent_buffer *eb,
3537                            struct extent_state *cached_state)
3538 {
3539         int ret = 0;
3540         unsigned long num_pages;
3541         unsigned long i;
3542         struct page *page;
3543         int pg_uptodate = 1;
3544
3545         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3546                 return 1;
3547
3548         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3549                            EXTENT_UPTODATE, 1, cached_state);
3550         if (ret)
3551                 return ret;
3552
3553         num_pages = num_extent_pages(eb->start, eb->len);
3554         for (i = 0; i < num_pages; i++) {
3555                 page = extent_buffer_page(eb, i);
3556                 if (!PageUptodate(page)) {
3557                         pg_uptodate = 0;
3558                         break;
3559                 }
3560         }
3561         return pg_uptodate;
3562 }
3563
3564 int read_extent_buffer_pages(struct extent_io_tree *tree,
3565                              struct extent_buffer *eb,
3566                              u64 start, int wait,
3567                              get_extent_t *get_extent, int mirror_num)
3568 {
3569         unsigned long i;
3570         unsigned long start_i;
3571         struct page *page;
3572         int err;
3573         int ret = 0;
3574         int locked_pages = 0;
3575         int all_uptodate = 1;
3576         int inc_all_pages = 0;
3577         unsigned long num_pages;
3578         struct bio *bio = NULL;
3579         unsigned long bio_flags = 0;
3580
3581         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3582                 return 0;
3583
3584         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3585                            EXTENT_UPTODATE, 1, NULL)) {
3586                 return 0;
3587         }
3588
3589         if (start) {
3590                 WARN_ON(start < eb->start);
3591                 start_i = (start >> PAGE_CACHE_SHIFT) -
3592                         (eb->start >> PAGE_CACHE_SHIFT);
3593         } else {
3594                 start_i = 0;
3595         }
3596
3597         num_pages = num_extent_pages(eb->start, eb->len);
3598         for (i = start_i; i < num_pages; i++) {
3599                 page = extent_buffer_page(eb, i);
3600                 if (!wait) {
3601                         if (!trylock_page(page))
3602                                 goto unlock_exit;
3603                 } else {
3604                         lock_page(page);
3605                 }
3606                 locked_pages++;
3607                 if (!PageUptodate(page))
3608                         all_uptodate = 0;
3609         }
3610         if (all_uptodate) {
3611                 if (start_i == 0)
3612                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3613                 goto unlock_exit;
3614         }
3615
3616         for (i = start_i; i < num_pages; i++) {
3617                 page = extent_buffer_page(eb, i);
3618
3619                 WARN_ON(!PagePrivate(page));
3620
3621                 set_page_extent_mapped(page);
3622                 if (i == 0)
3623                         set_page_extent_head(page, eb->len);
3624
3625                 if (inc_all_pages)
3626                         page_cache_get(page);
3627                 if (!PageUptodate(page)) {
3628                         if (start_i == 0)
3629                                 inc_all_pages = 1;
3630                         ClearPageError(page);
3631                         err = __extent_read_full_page(tree, page,
3632                                                       get_extent, &bio,
3633                                                       mirror_num, &bio_flags);
3634                         if (err)
3635                                 ret = err;
3636                 } else {
3637                         unlock_page(page);
3638                 }
3639         }
3640
3641         if (bio)
3642                 submit_one_bio(READ, bio, mirror_num, bio_flags);
3643
3644         if (ret || !wait)
3645                 return ret;
3646
3647         for (i = start_i; i < num_pages; i++) {
3648                 page = extent_buffer_page(eb, i);
3649                 wait_on_page_locked(page);
3650                 if (!PageUptodate(page))
3651                         ret = -EIO;
3652         }
3653
3654         if (!ret)
3655                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3656         return ret;
3657
3658 unlock_exit:
3659         i = start_i;
3660         while (locked_pages > 0) {
3661                 page = extent_buffer_page(eb, i);
3662                 i++;
3663                 unlock_page(page);
3664                 locked_pages--;
3665         }
3666         return ret;
3667 }
3668
3669 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3670                         unsigned long start,
3671                         unsigned long len)
3672 {
3673         size_t cur;
3674         size_t offset;
3675         struct page *page;
3676         char *kaddr;
3677         char *dst = (char *)dstv;
3678         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3679         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3680
3681         WARN_ON(start > eb->len);
3682         WARN_ON(start + len > eb->start + eb->len);
3683
3684         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3685
3686         while (len > 0) {
3687                 page = extent_buffer_page(eb, i);
3688
3689                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3690                 kaddr = kmap_atomic(page, KM_USER1);
3691                 memcpy(dst, kaddr + offset, cur);
3692                 kunmap_atomic(kaddr, KM_USER1);
3693
3694                 dst += cur;
3695                 len -= cur;
3696                 offset = 0;
3697                 i++;
3698         }
3699 }
3700
3701 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3702                                unsigned long min_len, char **token, char **map,
3703                                unsigned long *map_start,
3704                                unsigned long *map_len, int km)
3705 {
3706         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3707         char *kaddr;
3708         struct page *p;
3709         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3710         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3711         unsigned long end_i = (start_offset + start + min_len - 1) >>
3712                 PAGE_CACHE_SHIFT;
3713
3714         if (i != end_i)
3715                 return -EINVAL;
3716
3717         if (i == 0) {
3718                 offset = start_offset;
3719                 *map_start = 0;
3720         } else {
3721                 offset = 0;
3722                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3723         }
3724
3725         if (start + min_len > eb->len) {
3726                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3727                        "wanted %lu %lu\n", (unsigned long long)eb->start,
3728                        eb->len, start, min_len);
3729                 WARN_ON(1);
3730                 return -EINVAL;
3731         }
3732
3733         p = extent_buffer_page(eb, i);
3734         kaddr = kmap_atomic(p, km);
3735         *token = kaddr;
3736         *map = kaddr + offset;
3737         *map_len = PAGE_CACHE_SIZE - offset;
3738         return 0;
3739 }
3740
3741 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3742                       unsigned long min_len,
3743                       char **token, char **map,
3744                       unsigned long *map_start,
3745                       unsigned long *map_len, int km)
3746 {
3747         int err;
3748         int save = 0;
3749         if (eb->map_token) {
3750                 unmap_extent_buffer(eb, eb->map_token, km);
3751                 eb->map_token = NULL;
3752                 save = 1;
3753         }
3754         err = map_private_extent_buffer(eb, start, min_len, token, map,
3755                                        map_start, map_len, km);
3756         if (!err && save) {
3757                 eb->map_token = *token;
3758                 eb->kaddr = *map;
3759                 eb->map_start = *map_start;
3760                 eb->map_len = *map_len;
3761         }
3762         return err;
3763 }
3764
3765 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3766 {
3767         kunmap_atomic(token, km);
3768 }
3769
3770 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3771                           unsigned long start,
3772                           unsigned long len)
3773 {
3774         size_t cur;
3775         size_t offset;
3776         struct page *page;
3777         char *kaddr;
3778         char *ptr = (char *)ptrv;
3779         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3780         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3781         int ret = 0;
3782
3783         WARN_ON(start > eb->len);
3784         WARN_ON(start + len > eb->start + eb->len);
3785
3786         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3787
3788         while (len > 0) {
3789                 page = extent_buffer_page(eb, i);
3790
3791                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3792
3793                 kaddr = kmap_atomic(page, KM_USER0);
3794                 ret = memcmp(ptr, kaddr + offset, cur);
3795                 kunmap_atomic(kaddr, KM_USER0);
3796                 if (ret)
3797                         break;
3798
3799                 ptr += cur;
3800                 len -= cur;
3801                 offset = 0;
3802                 i++;
3803         }
3804         return ret;
3805 }
3806
3807 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3808                          unsigned long start, unsigned long len)
3809 {
3810         size_t cur;
3811         size_t offset;
3812         struct page *page;
3813         char *kaddr;
3814         char *src = (char *)srcv;
3815         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3816         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3817
3818         WARN_ON(start > eb->len);
3819         WARN_ON(start + len > eb->start + eb->len);
3820
3821         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3822
3823         while (len > 0) {
3824                 page = extent_buffer_page(eb, i);
3825                 WARN_ON(!PageUptodate(page));
3826
3827                 cur = min(len, PAGE_CACHE_SIZE - offset);
3828                 kaddr = kmap_atomic(page, KM_USER1);
3829                 memcpy(kaddr + offset, src, cur);
3830                 kunmap_atomic(kaddr, KM_USER1);
3831
3832                 src += cur;
3833                 len -= cur;
3834                 offset = 0;
3835                 i++;
3836         }
3837 }
3838
3839 void memset_extent_buffer(struct extent_buffer *eb, char c,
3840                           unsigned long start, unsigned long len)
3841 {
3842         size_t cur;
3843         size_t offset;
3844         struct page *page;
3845         char *kaddr;
3846         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3847         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3848
3849         WARN_ON(start > eb->len);
3850         WARN_ON(start + len > eb->start + eb->len);
3851
3852         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3853
3854         while (len > 0) {
3855                 page = extent_buffer_page(eb, i);
3856                 WARN_ON(!PageUptodate(page));
3857
3858                 cur = min(len, PAGE_CACHE_SIZE - offset);
3859                 kaddr = kmap_atomic(page, KM_USER0);
3860                 memset(kaddr + offset, c, cur);
3861                 kunmap_atomic(kaddr, KM_USER0);
3862
3863                 len -= cur;
3864                 offset = 0;
3865                 i++;
3866         }
3867 }
3868
3869 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3870                         unsigned long dst_offset, unsigned long src_offset,
3871                         unsigned long len)
3872 {
3873         u64 dst_len = dst->len;
3874         size_t cur;
3875         size_t offset;
3876         struct page *page;
3877         char *kaddr;
3878         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3879         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3880
3881         WARN_ON(src->len != dst_len);
3882
3883         offset = (start_offset + dst_offset) &
3884                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3885
3886         while (len > 0) {
3887                 page = extent_buffer_page(dst, i);
3888                 WARN_ON(!PageUptodate(page));
3889
3890                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3891
3892                 kaddr = kmap_atomic(page, KM_USER0);
3893                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3894                 kunmap_atomic(kaddr, KM_USER0);
3895
3896                 src_offset += cur;
3897                 len -= cur;
3898                 offset = 0;
3899                 i++;
3900         }
3901 }
3902
3903 static void move_pages(struct page *dst_page, struct page *src_page,
3904                        unsigned long dst_off, unsigned long src_off,
3905                        unsigned long len)
3906 {
3907         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3908         if (dst_page == src_page) {
3909                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3910         } else {
3911                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3912                 char *p = dst_kaddr + dst_off + len;
3913                 char *s = src_kaddr + src_off + len;
3914
3915                 while (len--)
3916                         *--p = *--s;
3917
3918                 kunmap_atomic(src_kaddr, KM_USER1);
3919         }
3920         kunmap_atomic(dst_kaddr, KM_USER0);
3921 }
3922
3923 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
3924 {
3925         unsigned long distance = (src > dst) ? src - dst : dst - src;
3926         return distance < len;
3927 }
3928
3929 static void copy_pages(struct page *dst_page, struct page *src_page,
3930                        unsigned long dst_off, unsigned long src_off,
3931                        unsigned long len)
3932 {
3933         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3934         char *src_kaddr;
3935
3936         if (dst_page != src_page) {
3937                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3938         } else {
3939                 src_kaddr = dst_kaddr;
3940                 BUG_ON(areas_overlap(src_off, dst_off, len));
3941         }
3942
3943         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3944         kunmap_atomic(dst_kaddr, KM_USER0);
3945         if (dst_page != src_page)
3946                 kunmap_atomic(src_kaddr, KM_USER1);
3947 }
3948
3949 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3950                            unsigned long src_offset, unsigned long len)
3951 {
3952         size_t cur;
3953         size_t dst_off_in_page;
3954         size_t src_off_in_page;
3955         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3956         unsigned long dst_i;
3957         unsigned long src_i;
3958
3959         if (src_offset + len > dst->len) {
3960                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3961                        "len %lu dst len %lu\n", src_offset, len, dst->len);
3962                 BUG_ON(1);
3963         }
3964         if (dst_offset + len > dst->len) {
3965                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3966                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
3967                 BUG_ON(1);
3968         }
3969
3970         while (len > 0) {
3971                 dst_off_in_page = (start_offset + dst_offset) &
3972                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3973                 src_off_in_page = (start_offset + src_offset) &
3974                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3975
3976                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3977                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3978
3979                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3980                                                src_off_in_page));
3981                 cur = min_t(unsigned long, cur,
3982                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3983
3984                 copy_pages(extent_buffer_page(dst, dst_i),
3985                            extent_buffer_page(dst, src_i),
3986                            dst_off_in_page, src_off_in_page, cur);
3987
3988                 src_offset += cur;
3989                 dst_offset += cur;
3990                 len -= cur;
3991         }
3992 }
3993
3994 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3995                            unsigned long src_offset, unsigned long len)
3996 {
3997         size_t cur;
3998         size_t dst_off_in_page;
3999         size_t src_off_in_page;
4000         unsigned long dst_end = dst_offset + len - 1;
4001         unsigned long src_end = src_offset + len - 1;
4002         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4003         unsigned long dst_i;
4004         unsigned long src_i;
4005
4006         if (src_offset + len > dst->len) {
4007                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4008                        "len %lu len %lu\n", src_offset, len, dst->len);
4009                 BUG_ON(1);
4010         }
4011         if (dst_offset + len > dst->len) {
4012                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4013                        "len %lu len %lu\n", dst_offset, len, dst->len);
4014                 BUG_ON(1);
4015         }
4016         if (!areas_overlap(src_offset, dst_offset, len)) {
4017                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4018                 return;
4019         }
4020         while (len > 0) {
4021                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4022                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4023
4024                 dst_off_in_page = (start_offset + dst_end) &
4025                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4026                 src_off_in_page = (start_offset + src_end) &
4027                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4028
4029                 cur = min_t(unsigned long, len, src_off_in_page + 1);
4030                 cur = min(cur, dst_off_in_page + 1);
4031                 move_pages(extent_buffer_page(dst, dst_i),
4032                            extent_buffer_page(dst, src_i),
4033                            dst_off_in_page - cur + 1,
4034                            src_off_in_page - cur + 1, cur);
4035
4036                 dst_end -= cur;
4037                 src_end -= cur;
4038                 len -= cur;
4039         }
4040 }
4041
4042 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4043 {
4044         struct extent_buffer *eb =
4045                         container_of(head, struct extent_buffer, rcu_head);
4046
4047         btrfs_release_extent_buffer(eb);
4048 }
4049
4050 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
4051 {
4052         u64 start = page_offset(page);
4053         struct extent_buffer *eb;
4054         int ret = 1;
4055
4056         spin_lock(&tree->buffer_lock);
4057         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4058         if (!eb) {
4059                 spin_unlock(&tree->buffer_lock);
4060                 return ret;
4061         }
4062
4063         if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
4064                 ret = 0;
4065                 goto out;
4066         }
4067
4068         /*
4069          * set @eb->refs to 0 if it is already 1, and then release the @eb.
4070          * Or go back.
4071          */
4072         if (atomic_cmpxchg(&eb->refs, 1, 0) != 1) {
4073                 ret = 0;
4074                 goto out;
4075         }
4076
4077         radix_tree_delete(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4078 out:
4079         spin_unlock(&tree->buffer_lock);
4080
4081         /* at this point we can safely release the extent buffer */
4082         if (atomic_read(&eb->refs) == 0)
4083                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4084         return ret;
4085 }