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/cleancache.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         if (!PageUptodate(page)) {
2020                 if (cleancache_get_page(page) == 0) {
2021                         BUG_ON(blocksize != PAGE_SIZE);
2022                         goto out;
2023                 }
2024         }
2025
2026         end = page_end;
2027         while (1) {
2028                 lock_extent(tree, start, end, GFP_NOFS);
2029                 ordered = btrfs_lookup_ordered_extent(inode, start);
2030                 if (!ordered)
2031                         break;
2032                 unlock_extent(tree, start, end, GFP_NOFS);
2033                 btrfs_start_ordered_extent(inode, ordered, 1);
2034                 btrfs_put_ordered_extent(ordered);
2035         }
2036
2037         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2038                 char *userpage;
2039                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2040
2041                 if (zero_offset) {
2042                         iosize = PAGE_CACHE_SIZE - zero_offset;
2043                         userpage = kmap_atomic(page, KM_USER0);
2044                         memset(userpage + zero_offset, 0, iosize);
2045                         flush_dcache_page(page);
2046                         kunmap_atomic(userpage, KM_USER0);
2047                 }
2048         }
2049         while (cur <= end) {
2050                 if (cur >= last_byte) {
2051                         char *userpage;
2052                         struct extent_state *cached = NULL;
2053
2054                         iosize = PAGE_CACHE_SIZE - page_offset;
2055                         userpage = kmap_atomic(page, KM_USER0);
2056                         memset(userpage + page_offset, 0, iosize);
2057                         flush_dcache_page(page);
2058                         kunmap_atomic(userpage, KM_USER0);
2059                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2060                                             &cached, GFP_NOFS);
2061                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2062                                              &cached, GFP_NOFS);
2063                         break;
2064                 }
2065                 em = get_extent(inode, page, page_offset, cur,
2066                                 end - cur + 1, 0);
2067                 if (IS_ERR(em) || !em) {
2068                         SetPageError(page);
2069                         unlock_extent(tree, cur, end, GFP_NOFS);
2070                         break;
2071                 }
2072                 extent_offset = cur - em->start;
2073                 BUG_ON(extent_map_end(em) <= cur);
2074                 BUG_ON(end < cur);
2075
2076                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2077                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2078                         extent_set_compress_type(&this_bio_flag,
2079                                                  em->compress_type);
2080                 }
2081
2082                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2083                 cur_end = min(extent_map_end(em) - 1, end);
2084                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2085                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2086                         disk_io_size = em->block_len;
2087                         sector = em->block_start >> 9;
2088                 } else {
2089                         sector = (em->block_start + extent_offset) >> 9;
2090                         disk_io_size = iosize;
2091                 }
2092                 bdev = em->bdev;
2093                 block_start = em->block_start;
2094                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2095                         block_start = EXTENT_MAP_HOLE;
2096                 free_extent_map(em);
2097                 em = NULL;
2098
2099                 /* we've found a hole, just zero and go on */
2100                 if (block_start == EXTENT_MAP_HOLE) {
2101                         char *userpage;
2102                         struct extent_state *cached = NULL;
2103
2104                         userpage = kmap_atomic(page, KM_USER0);
2105                         memset(userpage + page_offset, 0, iosize);
2106                         flush_dcache_page(page);
2107                         kunmap_atomic(userpage, KM_USER0);
2108
2109                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2110                                             &cached, GFP_NOFS);
2111                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2112                                              &cached, GFP_NOFS);
2113                         cur = cur + iosize;
2114                         page_offset += iosize;
2115                         continue;
2116                 }
2117                 /* the get_extent function already copied into the page */
2118                 if (test_range_bit(tree, cur, cur_end,
2119                                    EXTENT_UPTODATE, 1, NULL)) {
2120                         check_page_uptodate(tree, page);
2121                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2122                         cur = cur + iosize;
2123                         page_offset += iosize;
2124                         continue;
2125                 }
2126                 /* we have an inline extent but it didn't get marked up
2127                  * to date.  Error out
2128                  */
2129                 if (block_start == EXTENT_MAP_INLINE) {
2130                         SetPageError(page);
2131                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2132                         cur = cur + iosize;
2133                         page_offset += iosize;
2134                         continue;
2135                 }
2136
2137                 ret = 0;
2138                 if (tree->ops && tree->ops->readpage_io_hook) {
2139                         ret = tree->ops->readpage_io_hook(page, cur,
2140                                                           cur + iosize - 1);
2141                 }
2142                 if (!ret) {
2143                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2144                         pnr -= page->index;
2145                         ret = submit_extent_page(READ, tree, page,
2146                                          sector, disk_io_size, page_offset,
2147                                          bdev, bio, pnr,
2148                                          end_bio_extent_readpage, mirror_num,
2149                                          *bio_flags,
2150                                          this_bio_flag);
2151                         nr++;
2152                         *bio_flags = this_bio_flag;
2153                 }
2154                 if (ret)
2155                         SetPageError(page);
2156                 cur = cur + iosize;
2157                 page_offset += iosize;
2158         }
2159 out:
2160         if (!nr) {
2161                 if (!PageError(page))
2162                         SetPageUptodate(page);
2163                 unlock_page(page);
2164         }
2165         return 0;
2166 }
2167
2168 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2169                             get_extent_t *get_extent)
2170 {
2171         struct bio *bio = NULL;
2172         unsigned long bio_flags = 0;
2173         int ret;
2174
2175         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2176                                       &bio_flags);
2177         if (bio)
2178                 ret = submit_one_bio(READ, bio, 0, bio_flags);
2179         return ret;
2180 }
2181
2182 static noinline void update_nr_written(struct page *page,
2183                                       struct writeback_control *wbc,
2184                                       unsigned long nr_written)
2185 {
2186         wbc->nr_to_write -= nr_written;
2187         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2188             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2189                 page->mapping->writeback_index = page->index + nr_written;
2190 }
2191
2192 /*
2193  * the writepage semantics are similar to regular writepage.  extent
2194  * records are inserted to lock ranges in the tree, and as dirty areas
2195  * are found, they are marked writeback.  Then the lock bits are removed
2196  * and the end_io handler clears the writeback ranges
2197  */
2198 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2199                               void *data)
2200 {
2201         struct inode *inode = page->mapping->host;
2202         struct extent_page_data *epd = data;
2203         struct extent_io_tree *tree = epd->tree;
2204         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2205         u64 delalloc_start;
2206         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2207         u64 end;
2208         u64 cur = start;
2209         u64 extent_offset;
2210         u64 last_byte = i_size_read(inode);
2211         u64 block_start;
2212         u64 iosize;
2213         sector_t sector;
2214         struct extent_state *cached_state = NULL;
2215         struct extent_map *em;
2216         struct block_device *bdev;
2217         int ret;
2218         int nr = 0;
2219         size_t pg_offset = 0;
2220         size_t blocksize;
2221         loff_t i_size = i_size_read(inode);
2222         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2223         u64 nr_delalloc;
2224         u64 delalloc_end;
2225         int page_started;
2226         int compressed;
2227         int write_flags;
2228         unsigned long nr_written = 0;
2229
2230         if (wbc->sync_mode == WB_SYNC_ALL)
2231                 write_flags = WRITE_SYNC;
2232         else
2233                 write_flags = WRITE;
2234
2235         trace___extent_writepage(page, inode, wbc);
2236
2237         WARN_ON(!PageLocked(page));
2238         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2239         if (page->index > end_index ||
2240            (page->index == end_index && !pg_offset)) {
2241                 page->mapping->a_ops->invalidatepage(page, 0);
2242                 unlock_page(page);
2243                 return 0;
2244         }
2245
2246         if (page->index == end_index) {
2247                 char *userpage;
2248
2249                 userpage = kmap_atomic(page, KM_USER0);
2250                 memset(userpage + pg_offset, 0,
2251                        PAGE_CACHE_SIZE - pg_offset);
2252                 kunmap_atomic(userpage, KM_USER0);
2253                 flush_dcache_page(page);
2254         }
2255         pg_offset = 0;
2256
2257         set_page_extent_mapped(page);
2258
2259         delalloc_start = start;
2260         delalloc_end = 0;
2261         page_started = 0;
2262         if (!epd->extent_locked) {
2263                 u64 delalloc_to_write = 0;
2264                 /*
2265                  * make sure the wbc mapping index is at least updated
2266                  * to this page.
2267                  */
2268                 update_nr_written(page, wbc, 0);
2269
2270                 while (delalloc_end < page_end) {
2271                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2272                                                        page,
2273                                                        &delalloc_start,
2274                                                        &delalloc_end,
2275                                                        128 * 1024 * 1024);
2276                         if (nr_delalloc == 0) {
2277                                 delalloc_start = delalloc_end + 1;
2278                                 continue;
2279                         }
2280                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2281                                                  delalloc_end, &page_started,
2282                                                  &nr_written);
2283                         /*
2284                          * delalloc_end is already one less than the total
2285                          * length, so we don't subtract one from
2286                          * PAGE_CACHE_SIZE
2287                          */
2288                         delalloc_to_write += (delalloc_end - delalloc_start +
2289                                               PAGE_CACHE_SIZE) >>
2290                                               PAGE_CACHE_SHIFT;
2291                         delalloc_start = delalloc_end + 1;
2292                 }
2293                 if (wbc->nr_to_write < delalloc_to_write) {
2294                         int thresh = 8192;
2295
2296                         if (delalloc_to_write < thresh * 2)
2297                                 thresh = delalloc_to_write;
2298                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2299                                                  thresh);
2300                 }
2301
2302                 /* did the fill delalloc function already unlock and start
2303                  * the IO?
2304                  */
2305                 if (page_started) {
2306                         ret = 0;
2307                         /*
2308                          * we've unlocked the page, so we can't update
2309                          * the mapping's writeback index, just update
2310                          * nr_to_write.
2311                          */
2312                         wbc->nr_to_write -= nr_written;
2313                         goto done_unlocked;
2314                 }
2315         }
2316         if (tree->ops && tree->ops->writepage_start_hook) {
2317                 ret = tree->ops->writepage_start_hook(page, start,
2318                                                       page_end);
2319                 if (ret == -EAGAIN) {
2320                         redirty_page_for_writepage(wbc, page);
2321                         update_nr_written(page, wbc, nr_written);
2322                         unlock_page(page);
2323                         ret = 0;
2324                         goto done_unlocked;
2325                 }
2326         }
2327
2328         /*
2329          * we don't want to touch the inode after unlocking the page,
2330          * so we update the mapping writeback index now
2331          */
2332         update_nr_written(page, wbc, nr_written + 1);
2333
2334         end = page_end;
2335         if (last_byte <= start) {
2336                 if (tree->ops && tree->ops->writepage_end_io_hook)
2337                         tree->ops->writepage_end_io_hook(page, start,
2338                                                          page_end, NULL, 1);
2339                 goto done;
2340         }
2341
2342         blocksize = inode->i_sb->s_blocksize;
2343
2344         while (cur <= end) {
2345                 if (cur >= last_byte) {
2346                         if (tree->ops && tree->ops->writepage_end_io_hook)
2347                                 tree->ops->writepage_end_io_hook(page, cur,
2348                                                          page_end, NULL, 1);
2349                         break;
2350                 }
2351                 em = epd->get_extent(inode, page, pg_offset, cur,
2352                                      end - cur + 1, 1);
2353                 if (IS_ERR(em) || !em) {
2354                         SetPageError(page);
2355                         break;
2356                 }
2357
2358                 extent_offset = cur - em->start;
2359                 BUG_ON(extent_map_end(em) <= cur);
2360                 BUG_ON(end < cur);
2361                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2362                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2363                 sector = (em->block_start + extent_offset) >> 9;
2364                 bdev = em->bdev;
2365                 block_start = em->block_start;
2366                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2367                 free_extent_map(em);
2368                 em = NULL;
2369
2370                 /*
2371                  * compressed and inline extents are written through other
2372                  * paths in the FS
2373                  */
2374                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2375                     block_start == EXTENT_MAP_INLINE) {
2376                         /*
2377                          * end_io notification does not happen here for
2378                          * compressed extents
2379                          */
2380                         if (!compressed && tree->ops &&
2381                             tree->ops->writepage_end_io_hook)
2382                                 tree->ops->writepage_end_io_hook(page, cur,
2383                                                          cur + iosize - 1,
2384                                                          NULL, 1);
2385                         else if (compressed) {
2386                                 /* we don't want to end_page_writeback on
2387                                  * a compressed extent.  this happens
2388                                  * elsewhere
2389                                  */
2390                                 nr++;
2391                         }
2392
2393                         cur += iosize;
2394                         pg_offset += iosize;
2395                         continue;
2396                 }
2397                 /* leave this out until we have a page_mkwrite call */
2398                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2399                                    EXTENT_DIRTY, 0, NULL)) {
2400                         cur = cur + iosize;
2401                         pg_offset += iosize;
2402                         continue;
2403                 }
2404
2405                 if (tree->ops && tree->ops->writepage_io_hook) {
2406                         ret = tree->ops->writepage_io_hook(page, cur,
2407                                                 cur + iosize - 1);
2408                 } else {
2409                         ret = 0;
2410                 }
2411                 if (ret) {
2412                         SetPageError(page);
2413                 } else {
2414                         unsigned long max_nr = end_index + 1;
2415
2416                         set_range_writeback(tree, cur, cur + iosize - 1);
2417                         if (!PageWriteback(page)) {
2418                                 printk(KERN_ERR "btrfs warning page %lu not "
2419                                        "writeback, cur %llu end %llu\n",
2420                                        page->index, (unsigned long long)cur,
2421                                        (unsigned long long)end);
2422                         }
2423
2424                         ret = submit_extent_page(write_flags, tree, page,
2425                                                  sector, iosize, pg_offset,
2426                                                  bdev, &epd->bio, max_nr,
2427                                                  end_bio_extent_writepage,
2428                                                  0, 0, 0);
2429                         if (ret)
2430                                 SetPageError(page);
2431                 }
2432                 cur = cur + iosize;
2433                 pg_offset += iosize;
2434                 nr++;
2435         }
2436 done:
2437         if (nr == 0) {
2438                 /* make sure the mapping tag for page dirty gets cleared */
2439                 set_page_writeback(page);
2440                 end_page_writeback(page);
2441         }
2442         unlock_page(page);
2443
2444 done_unlocked:
2445
2446         /* drop our reference on any cached states */
2447         free_extent_state(cached_state);
2448         return 0;
2449 }
2450
2451 /**
2452  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2453  * @mapping: address space structure to write
2454  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2455  * @writepage: function called for each page
2456  * @data: data passed to writepage function
2457  *
2458  * If a page is already under I/O, write_cache_pages() skips it, even
2459  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2460  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2461  * and msync() need to guarantee that all the data which was dirty at the time
2462  * the call was made get new I/O started against them.  If wbc->sync_mode is
2463  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2464  * existing IO to complete.
2465  */
2466 static int extent_write_cache_pages(struct extent_io_tree *tree,
2467                              struct address_space *mapping,
2468                              struct writeback_control *wbc,
2469                              writepage_t writepage, void *data,
2470                              void (*flush_fn)(void *))
2471 {
2472         int ret = 0;
2473         int done = 0;
2474         int nr_to_write_done = 0;
2475         struct pagevec pvec;
2476         int nr_pages;
2477         pgoff_t index;
2478         pgoff_t end;            /* Inclusive */
2479         int scanned = 0;
2480
2481         pagevec_init(&pvec, 0);
2482         if (wbc->range_cyclic) {
2483                 index = mapping->writeback_index; /* Start from prev offset */
2484                 end = -1;
2485         } else {
2486                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2487                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2488                 scanned = 1;
2489         }
2490 retry:
2491         while (!done && !nr_to_write_done && (index <= end) &&
2492                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2493                               PAGECACHE_TAG_DIRTY, min(end - index,
2494                                   (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2495                 unsigned i;
2496
2497                 scanned = 1;
2498                 for (i = 0; i < nr_pages; i++) {
2499                         struct page *page = pvec.pages[i];
2500
2501                         /*
2502                          * At this point we hold neither mapping->tree_lock nor
2503                          * lock on the page itself: the page may be truncated or
2504                          * invalidated (changing page->mapping to NULL), or even
2505                          * swizzled back from swapper_space to tmpfs file
2506                          * mapping
2507                          */
2508                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2509                                 tree->ops->write_cache_pages_lock_hook(page);
2510                         else
2511                                 lock_page(page);
2512
2513                         if (unlikely(page->mapping != mapping)) {
2514                                 unlock_page(page);
2515                                 continue;
2516                         }
2517
2518                         if (!wbc->range_cyclic && page->index > end) {
2519                                 done = 1;
2520                                 unlock_page(page);
2521                                 continue;
2522                         }
2523
2524                         if (wbc->sync_mode != WB_SYNC_NONE) {
2525                                 if (PageWriteback(page))
2526                                         flush_fn(data);
2527                                 wait_on_page_writeback(page);
2528                         }
2529
2530                         if (PageWriteback(page) ||
2531                             !clear_page_dirty_for_io(page)) {
2532                                 unlock_page(page);
2533                                 continue;
2534                         }
2535
2536                         ret = (*writepage)(page, wbc, data);
2537
2538                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2539                                 unlock_page(page);
2540                                 ret = 0;
2541                         }
2542                         if (ret)
2543                                 done = 1;
2544
2545                         /*
2546                          * the filesystem may choose to bump up nr_to_write.
2547                          * We have to make sure to honor the new nr_to_write
2548                          * at any time
2549                          */
2550                         nr_to_write_done = wbc->nr_to_write <= 0;
2551                 }
2552                 pagevec_release(&pvec);
2553                 cond_resched();
2554         }
2555         if (!scanned && !done) {
2556                 /*
2557                  * We hit the last page and there is more work to be done: wrap
2558                  * back to the start of the file
2559                  */
2560                 scanned = 1;
2561                 index = 0;
2562                 goto retry;
2563         }
2564         return ret;
2565 }
2566
2567 static void flush_epd_write_bio(struct extent_page_data *epd)
2568 {
2569         if (epd->bio) {
2570                 if (epd->sync_io)
2571                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2572                 else
2573                         submit_one_bio(WRITE, epd->bio, 0, 0);
2574                 epd->bio = NULL;
2575         }
2576 }
2577
2578 static noinline void flush_write_bio(void *data)
2579 {
2580         struct extent_page_data *epd = data;
2581         flush_epd_write_bio(epd);
2582 }
2583
2584 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2585                           get_extent_t *get_extent,
2586                           struct writeback_control *wbc)
2587 {
2588         int ret;
2589         struct address_space *mapping = page->mapping;
2590         struct extent_page_data epd = {
2591                 .bio = NULL,
2592                 .tree = tree,
2593                 .get_extent = get_extent,
2594                 .extent_locked = 0,
2595                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2596         };
2597         struct writeback_control wbc_writepages = {
2598                 .sync_mode      = wbc->sync_mode,
2599                 .older_than_this = NULL,
2600                 .nr_to_write    = 64,
2601                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2602                 .range_end      = (loff_t)-1,
2603         };
2604
2605         ret = __extent_writepage(page, wbc, &epd);
2606
2607         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2608                                  __extent_writepage, &epd, flush_write_bio);
2609         flush_epd_write_bio(&epd);
2610         return ret;
2611 }
2612
2613 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2614                               u64 start, u64 end, get_extent_t *get_extent,
2615                               int mode)
2616 {
2617         int ret = 0;
2618         struct address_space *mapping = inode->i_mapping;
2619         struct page *page;
2620         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2621                 PAGE_CACHE_SHIFT;
2622
2623         struct extent_page_data epd = {
2624                 .bio = NULL,
2625                 .tree = tree,
2626                 .get_extent = get_extent,
2627                 .extent_locked = 1,
2628                 .sync_io = mode == WB_SYNC_ALL,
2629         };
2630         struct writeback_control wbc_writepages = {
2631                 .sync_mode      = mode,
2632                 .older_than_this = NULL,
2633                 .nr_to_write    = nr_pages * 2,
2634                 .range_start    = start,
2635                 .range_end      = end + 1,
2636         };
2637
2638         while (start <= end) {
2639                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2640                 if (clear_page_dirty_for_io(page))
2641                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2642                 else {
2643                         if (tree->ops && tree->ops->writepage_end_io_hook)
2644                                 tree->ops->writepage_end_io_hook(page, start,
2645                                                  start + PAGE_CACHE_SIZE - 1,
2646                                                  NULL, 1);
2647                         unlock_page(page);
2648                 }
2649                 page_cache_release(page);
2650                 start += PAGE_CACHE_SIZE;
2651         }
2652
2653         flush_epd_write_bio(&epd);
2654         return ret;
2655 }
2656
2657 int extent_writepages(struct extent_io_tree *tree,
2658                       struct address_space *mapping,
2659                       get_extent_t *get_extent,
2660                       struct writeback_control *wbc)
2661 {
2662         int ret = 0;
2663         struct extent_page_data epd = {
2664                 .bio = NULL,
2665                 .tree = tree,
2666                 .get_extent = get_extent,
2667                 .extent_locked = 0,
2668                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2669         };
2670
2671         ret = extent_write_cache_pages(tree, mapping, wbc,
2672                                        __extent_writepage, &epd,
2673                                        flush_write_bio);
2674         flush_epd_write_bio(&epd);
2675         return ret;
2676 }
2677
2678 int extent_readpages(struct extent_io_tree *tree,
2679                      struct address_space *mapping,
2680                      struct list_head *pages, unsigned nr_pages,
2681                      get_extent_t get_extent)
2682 {
2683         struct bio *bio = NULL;
2684         unsigned page_idx;
2685         unsigned long bio_flags = 0;
2686
2687         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2688                 struct page *page = list_entry(pages->prev, struct page, lru);
2689
2690                 prefetchw(&page->flags);
2691                 list_del(&page->lru);
2692                 if (!add_to_page_cache_lru(page, mapping,
2693                                         page->index, GFP_NOFS)) {
2694                         __extent_read_full_page(tree, page, get_extent,
2695                                                 &bio, 0, &bio_flags);
2696                 }
2697                 page_cache_release(page);
2698         }
2699         BUG_ON(!list_empty(pages));
2700         if (bio)
2701                 submit_one_bio(READ, bio, 0, bio_flags);
2702         return 0;
2703 }
2704
2705 /*
2706  * basic invalidatepage code, this waits on any locked or writeback
2707  * ranges corresponding to the page, and then deletes any extent state
2708  * records from the tree
2709  */
2710 int extent_invalidatepage(struct extent_io_tree *tree,
2711                           struct page *page, unsigned long offset)
2712 {
2713         struct extent_state *cached_state = NULL;
2714         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2715         u64 end = start + PAGE_CACHE_SIZE - 1;
2716         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2717
2718         start += (offset + blocksize - 1) & ~(blocksize - 1);
2719         if (start > end)
2720                 return 0;
2721
2722         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
2723         wait_on_page_writeback(page);
2724         clear_extent_bit(tree, start, end,
2725                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2726                          EXTENT_DO_ACCOUNTING,
2727                          1, 1, &cached_state, GFP_NOFS);
2728         return 0;
2729 }
2730
2731 /*
2732  * simple commit_write call, set_range_dirty is used to mark both
2733  * the pages and the extent records as dirty
2734  */
2735 int extent_commit_write(struct extent_io_tree *tree,
2736                         struct inode *inode, struct page *page,
2737                         unsigned from, unsigned to)
2738 {
2739         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2740
2741         set_page_extent_mapped(page);
2742         set_page_dirty(page);
2743
2744         if (pos > inode->i_size) {
2745                 i_size_write(inode, pos);
2746                 mark_inode_dirty(inode);
2747         }
2748         return 0;
2749 }
2750
2751 int extent_prepare_write(struct extent_io_tree *tree,
2752                          struct inode *inode, struct page *page,
2753                          unsigned from, unsigned to, get_extent_t *get_extent)
2754 {
2755         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2756         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2757         u64 block_start;
2758         u64 orig_block_start;
2759         u64 block_end;
2760         u64 cur_end;
2761         struct extent_map *em;
2762         unsigned blocksize = 1 << inode->i_blkbits;
2763         size_t page_offset = 0;
2764         size_t block_off_start;
2765         size_t block_off_end;
2766         int err = 0;
2767         int iocount = 0;
2768         int ret = 0;
2769         int isnew;
2770
2771         set_page_extent_mapped(page);
2772
2773         block_start = (page_start + from) & ~((u64)blocksize - 1);
2774         block_end = (page_start + to - 1) | (blocksize - 1);
2775         orig_block_start = block_start;
2776
2777         lock_extent(tree, page_start, page_end, GFP_NOFS);
2778         while (block_start <= block_end) {
2779                 em = get_extent(inode, page, page_offset, block_start,
2780                                 block_end - block_start + 1, 1);
2781                 if (IS_ERR(em) || !em)
2782                         goto err;
2783
2784                 cur_end = min(block_end, extent_map_end(em) - 1);
2785                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2786                 block_off_end = block_off_start + blocksize;
2787                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2788
2789                 if (!PageUptodate(page) && isnew &&
2790                     (block_off_end > to || block_off_start < from)) {
2791                         void *kaddr;
2792
2793                         kaddr = kmap_atomic(page, KM_USER0);
2794                         if (block_off_end > to)
2795                                 memset(kaddr + to, 0, block_off_end - to);
2796                         if (block_off_start < from)
2797                                 memset(kaddr + block_off_start, 0,
2798                                        from - block_off_start);
2799                         flush_dcache_page(page);
2800                         kunmap_atomic(kaddr, KM_USER0);
2801                 }
2802                 if ((em->block_start != EXTENT_MAP_HOLE &&
2803                      em->block_start != EXTENT_MAP_INLINE) &&
2804                     !isnew && !PageUptodate(page) &&
2805                     (block_off_end > to || block_off_start < from) &&
2806                     !test_range_bit(tree, block_start, cur_end,
2807                                     EXTENT_UPTODATE, 1, NULL)) {
2808                         u64 sector;
2809                         u64 extent_offset = block_start - em->start;
2810                         size_t iosize;
2811                         sector = (em->block_start + extent_offset) >> 9;
2812                         iosize = (cur_end - block_start + blocksize) &
2813                                 ~((u64)blocksize - 1);
2814                         /*
2815                          * we've already got the extent locked, but we
2816                          * need to split the state such that our end_bio
2817                          * handler can clear the lock.
2818                          */
2819                         set_extent_bit(tree, block_start,
2820                                        block_start + iosize - 1,
2821                                        EXTENT_LOCKED, 0, NULL, NULL, GFP_NOFS);
2822                         ret = submit_extent_page(READ, tree, page,
2823                                          sector, iosize, page_offset, em->bdev,
2824                                          NULL, 1,
2825                                          end_bio_extent_preparewrite, 0,
2826                                          0, 0);
2827                         if (ret && !err)
2828                                 err = ret;
2829                         iocount++;
2830                         block_start = block_start + iosize;
2831                 } else {
2832                         struct extent_state *cached = NULL;
2833
2834                         set_extent_uptodate(tree, block_start, cur_end, &cached,
2835                                             GFP_NOFS);
2836                         unlock_extent_cached(tree, block_start, cur_end,
2837                                              &cached, GFP_NOFS);
2838                         block_start = cur_end + 1;
2839                 }
2840                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2841                 free_extent_map(em);
2842         }
2843         if (iocount) {
2844                 wait_extent_bit(tree, orig_block_start,
2845                                 block_end, EXTENT_LOCKED);
2846         }
2847         check_page_uptodate(tree, page);
2848 err:
2849         /* FIXME, zero out newly allocated blocks on error */
2850         return err;
2851 }
2852
2853 /*
2854  * a helper for releasepage, this tests for areas of the page that
2855  * are locked or under IO and drops the related state bits if it is safe
2856  * to drop the page.
2857  */
2858 int try_release_extent_state(struct extent_map_tree *map,
2859                              struct extent_io_tree *tree, struct page *page,
2860                              gfp_t mask)
2861 {
2862         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2863         u64 end = start + PAGE_CACHE_SIZE - 1;
2864         int ret = 1;
2865
2866         if (test_range_bit(tree, start, end,
2867                            EXTENT_IOBITS, 0, NULL))
2868                 ret = 0;
2869         else {
2870                 if ((mask & GFP_NOFS) == GFP_NOFS)
2871                         mask = GFP_NOFS;
2872                 /*
2873                  * at this point we can safely clear everything except the
2874                  * locked bit and the nodatasum bit
2875                  */
2876                 ret = clear_extent_bit(tree, start, end,
2877                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2878                                  0, 0, NULL, mask);
2879
2880                 /* if clear_extent_bit failed for enomem reasons,
2881                  * we can't allow the release to continue.
2882                  */
2883                 if (ret < 0)
2884                         ret = 0;
2885                 else
2886                         ret = 1;
2887         }
2888         return ret;
2889 }
2890
2891 /*
2892  * a helper for releasepage.  As long as there are no locked extents
2893  * in the range corresponding to the page, both state records and extent
2894  * map records are removed
2895  */
2896 int try_release_extent_mapping(struct extent_map_tree *map,
2897                                struct extent_io_tree *tree, struct page *page,
2898                                gfp_t mask)
2899 {
2900         struct extent_map *em;
2901         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2902         u64 end = start + PAGE_CACHE_SIZE - 1;
2903
2904         if ((mask & __GFP_WAIT) &&
2905             page->mapping->host->i_size > 16 * 1024 * 1024) {
2906                 u64 len;
2907                 while (start <= end) {
2908                         len = end - start + 1;
2909                         write_lock(&map->lock);
2910                         em = lookup_extent_mapping(map, start, len);
2911                         if (!em || IS_ERR(em)) {
2912                                 write_unlock(&map->lock);
2913                                 break;
2914                         }
2915                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2916                             em->start != start) {
2917                                 write_unlock(&map->lock);
2918                                 free_extent_map(em);
2919                                 break;
2920                         }
2921                         if (!test_range_bit(tree, em->start,
2922                                             extent_map_end(em) - 1,
2923                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
2924                                             0, NULL)) {
2925                                 remove_extent_mapping(map, em);
2926                                 /* once for the rb tree */
2927                                 free_extent_map(em);
2928                         }
2929                         start = extent_map_end(em);
2930                         write_unlock(&map->lock);
2931
2932                         /* once for us */
2933                         free_extent_map(em);
2934                 }
2935         }
2936         return try_release_extent_state(map, tree, page, mask);
2937 }
2938
2939 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2940                 get_extent_t *get_extent)
2941 {
2942         struct inode *inode = mapping->host;
2943         struct extent_state *cached_state = NULL;
2944         u64 start = iblock << inode->i_blkbits;
2945         sector_t sector = 0;
2946         size_t blksize = (1 << inode->i_blkbits);
2947         struct extent_map *em;
2948
2949         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2950                          0, &cached_state, GFP_NOFS);
2951         em = get_extent(inode, NULL, 0, start, blksize, 0);
2952         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start,
2953                              start + blksize - 1, &cached_state, GFP_NOFS);
2954         if (!em || IS_ERR(em))
2955                 return 0;
2956
2957         if (em->block_start > EXTENT_MAP_LAST_BYTE)
2958                 goto out;
2959
2960         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2961 out:
2962         free_extent_map(em);
2963         return sector;
2964 }
2965
2966 /*
2967  * helper function for fiemap, which doesn't want to see any holes.
2968  * This maps until we find something past 'last'
2969  */
2970 static struct extent_map *get_extent_skip_holes(struct inode *inode,
2971                                                 u64 offset,
2972                                                 u64 last,
2973                                                 get_extent_t *get_extent)
2974 {
2975         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
2976         struct extent_map *em;
2977         u64 len;
2978
2979         if (offset >= last)
2980                 return NULL;
2981
2982         while(1) {
2983                 len = last - offset;
2984                 if (len == 0)
2985                         break;
2986                 len = (len + sectorsize - 1) & ~(sectorsize - 1);
2987                 em = get_extent(inode, NULL, 0, offset, len, 0);
2988                 if (!em || IS_ERR(em))
2989                         return em;
2990
2991                 /* if this isn't a hole return it */
2992                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
2993                     em->block_start != EXTENT_MAP_HOLE) {
2994                         return em;
2995                 }
2996
2997                 /* this is a hole, advance to the next extent */
2998                 offset = extent_map_end(em);
2999                 free_extent_map(em);
3000                 if (offset >= last)
3001                         break;
3002         }
3003         return NULL;
3004 }
3005
3006 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
3007                 __u64 start, __u64 len, get_extent_t *get_extent)
3008 {
3009         int ret = 0;
3010         u64 off = start;
3011         u64 max = start + len;
3012         u32 flags = 0;
3013         u32 found_type;
3014         u64 last;
3015         u64 last_for_get_extent = 0;
3016         u64 disko = 0;
3017         u64 isize = i_size_read(inode);
3018         struct btrfs_key found_key;
3019         struct extent_map *em = NULL;
3020         struct extent_state *cached_state = NULL;
3021         struct btrfs_path *path;
3022         struct btrfs_file_extent_item *item;
3023         int end = 0;
3024         u64 em_start = 0;
3025         u64 em_len = 0;
3026         u64 em_end = 0;
3027         unsigned long emflags;
3028
3029         if (len == 0)
3030                 return -EINVAL;
3031
3032         path = btrfs_alloc_path();
3033         if (!path)
3034                 return -ENOMEM;
3035         path->leave_spinning = 1;
3036
3037         /*
3038          * lookup the last file extent.  We're not using i_size here
3039          * because there might be preallocation past i_size
3040          */
3041         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3042                                        path, inode->i_ino, -1, 0);
3043         if (ret < 0) {
3044                 btrfs_free_path(path);
3045                 return ret;
3046         }
3047         WARN_ON(!ret);
3048         path->slots[0]--;
3049         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3050                               struct btrfs_file_extent_item);
3051         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3052         found_type = btrfs_key_type(&found_key);
3053
3054         /* No extents, but there might be delalloc bits */
3055         if (found_key.objectid != inode->i_ino ||
3056             found_type != BTRFS_EXTENT_DATA_KEY) {
3057                 /* have to trust i_size as the end */
3058                 last = (u64)-1;
3059                 last_for_get_extent = isize;
3060         } else {
3061                 /*
3062                  * remember the start of the last extent.  There are a
3063                  * bunch of different factors that go into the length of the
3064                  * extent, so its much less complex to remember where it started
3065                  */
3066                 last = found_key.offset;
3067                 last_for_get_extent = last + 1;
3068         }
3069         btrfs_free_path(path);
3070
3071         /*
3072          * we might have some extents allocated but more delalloc past those
3073          * extents.  so, we trust isize unless the start of the last extent is
3074          * beyond isize
3075          */
3076         if (last < isize) {
3077                 last = (u64)-1;
3078                 last_for_get_extent = isize;
3079         }
3080
3081         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3082                          &cached_state, GFP_NOFS);
3083
3084         em = get_extent_skip_holes(inode, off, last_for_get_extent,
3085                                    get_extent);
3086         if (!em)
3087                 goto out;
3088         if (IS_ERR(em)) {
3089                 ret = PTR_ERR(em);
3090                 goto out;
3091         }
3092
3093         while (!end) {
3094                 u64 offset_in_extent;
3095
3096                 /* break if the extent we found is outside the range */
3097                 if (em->start >= max || extent_map_end(em) < off)
3098                         break;
3099
3100                 /*
3101                  * get_extent may return an extent that starts before our
3102                  * requested range.  We have to make sure the ranges
3103                  * we return to fiemap always move forward and don't
3104                  * overlap, so adjust the offsets here
3105                  */
3106                 em_start = max(em->start, off);
3107
3108                 /*
3109                  * record the offset from the start of the extent
3110                  * for adjusting the disk offset below
3111                  */
3112                 offset_in_extent = em_start - em->start;
3113                 em_end = extent_map_end(em);
3114                 em_len = em_end - em_start;
3115                 emflags = em->flags;
3116                 disko = 0;
3117                 flags = 0;
3118
3119                 /*
3120                  * bump off for our next call to get_extent
3121                  */
3122                 off = extent_map_end(em);
3123                 if (off >= max)
3124                         end = 1;
3125
3126                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3127                         end = 1;
3128                         flags |= FIEMAP_EXTENT_LAST;
3129                 } else if (em->block_start == EXTENT_MAP_INLINE) {
3130                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
3131                                   FIEMAP_EXTENT_NOT_ALIGNED);
3132                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3133                         flags |= (FIEMAP_EXTENT_DELALLOC |
3134                                   FIEMAP_EXTENT_UNKNOWN);
3135                 } else {
3136                         disko = em->block_start + offset_in_extent;
3137                 }
3138                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3139                         flags |= FIEMAP_EXTENT_ENCODED;
3140
3141                 free_extent_map(em);
3142                 em = NULL;
3143                 if ((em_start >= last) || em_len == (u64)-1 ||
3144                    (last == (u64)-1 && isize <= em_end)) {
3145                         flags |= FIEMAP_EXTENT_LAST;
3146                         end = 1;
3147                 }
3148
3149                 /* now scan forward to see if this is really the last extent. */
3150                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
3151                                            get_extent);
3152                 if (IS_ERR(em)) {
3153                         ret = PTR_ERR(em);
3154                         goto out;
3155                 }
3156                 if (!em) {
3157                         flags |= FIEMAP_EXTENT_LAST;
3158                         end = 1;
3159                 }
3160                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3161                                               em_len, flags);
3162                 if (ret)
3163                         goto out_free;
3164         }
3165 out_free:
3166         free_extent_map(em);
3167 out:
3168         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3169                              &cached_state, GFP_NOFS);
3170         return ret;
3171 }
3172
3173 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
3174                                               unsigned long i)
3175 {
3176         struct page *p;
3177         struct address_space *mapping;
3178
3179         if (i == 0)
3180                 return eb->first_page;
3181         i += eb->start >> PAGE_CACHE_SHIFT;
3182         mapping = eb->first_page->mapping;
3183         if (!mapping)
3184                 return NULL;
3185
3186         /*
3187          * extent_buffer_page is only called after pinning the page
3188          * by increasing the reference count.  So we know the page must
3189          * be in the radix tree.
3190          */
3191         rcu_read_lock();
3192         p = radix_tree_lookup(&mapping->page_tree, i);
3193         rcu_read_unlock();
3194
3195         return p;
3196 }
3197
3198 static inline unsigned long num_extent_pages(u64 start, u64 len)
3199 {
3200         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3201                 (start >> PAGE_CACHE_SHIFT);
3202 }
3203
3204 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3205                                                    u64 start,
3206                                                    unsigned long len,
3207                                                    gfp_t mask)
3208 {
3209         struct extent_buffer *eb = NULL;
3210 #if LEAK_DEBUG
3211         unsigned long flags;
3212 #endif
3213
3214         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3215         if (eb == NULL)
3216                 return NULL;
3217         eb->start = start;
3218         eb->len = len;
3219         spin_lock_init(&eb->lock);
3220         init_waitqueue_head(&eb->lock_wq);
3221
3222 #if LEAK_DEBUG
3223         spin_lock_irqsave(&leak_lock, flags);
3224         list_add(&eb->leak_list, &buffers);
3225         spin_unlock_irqrestore(&leak_lock, flags);
3226 #endif
3227         atomic_set(&eb->refs, 1);
3228
3229         return eb;
3230 }
3231
3232 static void __free_extent_buffer(struct extent_buffer *eb)
3233 {
3234 #if LEAK_DEBUG
3235         unsigned long flags;
3236         spin_lock_irqsave(&leak_lock, flags);
3237         list_del(&eb->leak_list);
3238         spin_unlock_irqrestore(&leak_lock, flags);
3239 #endif
3240         kmem_cache_free(extent_buffer_cache, eb);
3241 }
3242
3243 /*
3244  * Helper for releasing extent buffer page.
3245  */
3246 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3247                                                 unsigned long start_idx)
3248 {
3249         unsigned long index;
3250         struct page *page;
3251
3252         if (!eb->first_page)
3253                 return;
3254
3255         index = num_extent_pages(eb->start, eb->len);
3256         if (start_idx >= index)
3257                 return;
3258
3259         do {
3260                 index--;
3261                 page = extent_buffer_page(eb, index);
3262                 if (page)
3263                         page_cache_release(page);
3264         } while (index != start_idx);
3265 }
3266
3267 /*
3268  * Helper for releasing the extent buffer.
3269  */
3270 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3271 {
3272         btrfs_release_extent_buffer_page(eb, 0);
3273         __free_extent_buffer(eb);
3274 }
3275
3276 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3277                                           u64 start, unsigned long len,
3278                                           struct page *page0,
3279                                           gfp_t mask)
3280 {
3281         unsigned long num_pages = num_extent_pages(start, len);
3282         unsigned long i;
3283         unsigned long index = start >> PAGE_CACHE_SHIFT;
3284         struct extent_buffer *eb;
3285         struct extent_buffer *exists = NULL;
3286         struct page *p;
3287         struct address_space *mapping = tree->mapping;
3288         int uptodate = 1;
3289         int ret;
3290
3291         rcu_read_lock();
3292         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3293         if (eb && atomic_inc_not_zero(&eb->refs)) {
3294                 rcu_read_unlock();
3295                 mark_page_accessed(eb->first_page);
3296                 return eb;
3297         }
3298         rcu_read_unlock();
3299
3300         eb = __alloc_extent_buffer(tree, start, len, mask);
3301         if (!eb)
3302                 return NULL;
3303
3304         if (page0) {
3305                 eb->first_page = page0;
3306                 i = 1;
3307                 index++;
3308                 page_cache_get(page0);
3309                 mark_page_accessed(page0);
3310                 set_page_extent_mapped(page0);
3311                 set_page_extent_head(page0, len);
3312                 uptodate = PageUptodate(page0);
3313         } else {
3314                 i = 0;
3315         }
3316         for (; i < num_pages; i++, index++) {
3317                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
3318                 if (!p) {
3319                         WARN_ON(1);
3320                         goto free_eb;
3321                 }
3322                 set_page_extent_mapped(p);
3323                 mark_page_accessed(p);
3324                 if (i == 0) {
3325                         eb->first_page = p;
3326                         set_page_extent_head(p, len);
3327                 } else {
3328                         set_page_private(p, EXTENT_PAGE_PRIVATE);
3329                 }
3330                 if (!PageUptodate(p))
3331                         uptodate = 0;
3332
3333                 /*
3334                  * see below about how we avoid a nasty race with release page
3335                  * and why we unlock later
3336                  */
3337                 if (i != 0)
3338                         unlock_page(p);
3339         }
3340         if (uptodate)
3341                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3342
3343         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
3344         if (ret)
3345                 goto free_eb;
3346
3347         spin_lock(&tree->buffer_lock);
3348         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
3349         if (ret == -EEXIST) {
3350                 exists = radix_tree_lookup(&tree->buffer,
3351                                                 start >> PAGE_CACHE_SHIFT);
3352                 /* add one reference for the caller */
3353                 atomic_inc(&exists->refs);
3354                 spin_unlock(&tree->buffer_lock);
3355                 radix_tree_preload_end();
3356                 goto free_eb;
3357         }
3358         /* add one reference for the tree */
3359         atomic_inc(&eb->refs);
3360         spin_unlock(&tree->buffer_lock);
3361         radix_tree_preload_end();
3362
3363         /*
3364          * there is a race where release page may have
3365          * tried to find this extent buffer in the radix
3366          * but failed.  It will tell the VM it is safe to
3367          * reclaim the, and it will clear the page private bit.
3368          * We must make sure to set the page private bit properly
3369          * after the extent buffer is in the radix tree so
3370          * it doesn't get lost
3371          */
3372         set_page_extent_mapped(eb->first_page);
3373         set_page_extent_head(eb->first_page, eb->len);
3374         if (!page0)
3375                 unlock_page(eb->first_page);
3376         return eb;
3377
3378 free_eb:
3379         if (eb->first_page && !page0)
3380                 unlock_page(eb->first_page);
3381
3382         if (!atomic_dec_and_test(&eb->refs))
3383                 return exists;
3384         btrfs_release_extent_buffer(eb);
3385         return exists;
3386 }
3387
3388 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3389                                          u64 start, unsigned long len,
3390                                           gfp_t mask)
3391 {
3392         struct extent_buffer *eb;
3393
3394         rcu_read_lock();
3395         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3396         if (eb && atomic_inc_not_zero(&eb->refs)) {
3397                 rcu_read_unlock();
3398                 mark_page_accessed(eb->first_page);
3399                 return eb;
3400         }
3401         rcu_read_unlock();
3402
3403         return NULL;
3404 }
3405
3406 void free_extent_buffer(struct extent_buffer *eb)
3407 {
3408         if (!eb)
3409                 return;
3410
3411         if (!atomic_dec_and_test(&eb->refs))
3412                 return;
3413
3414         WARN_ON(1);
3415 }
3416
3417 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3418                               struct extent_buffer *eb)
3419 {
3420         unsigned long i;
3421         unsigned long num_pages;
3422         struct page *page;
3423
3424         num_pages = num_extent_pages(eb->start, eb->len);
3425
3426         for (i = 0; i < num_pages; i++) {
3427                 page = extent_buffer_page(eb, i);
3428                 if (!PageDirty(page))
3429                         continue;
3430
3431                 lock_page(page);
3432                 WARN_ON(!PagePrivate(page));
3433
3434                 set_page_extent_mapped(page);
3435                 if (i == 0)
3436                         set_page_extent_head(page, eb->len);
3437
3438                 clear_page_dirty_for_io(page);
3439                 spin_lock_irq(&page->mapping->tree_lock);
3440                 if (!PageDirty(page)) {
3441                         radix_tree_tag_clear(&page->mapping->page_tree,
3442                                                 page_index(page),
3443                                                 PAGECACHE_TAG_DIRTY);
3444                 }
3445                 spin_unlock_irq(&page->mapping->tree_lock);
3446                 unlock_page(page);
3447         }
3448         return 0;
3449 }
3450
3451 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
3452                                     struct extent_buffer *eb)
3453 {
3454         return wait_on_extent_writeback(tree, eb->start,
3455                                         eb->start + eb->len - 1);
3456 }
3457
3458 int set_extent_buffer_dirty(struct extent_io_tree *tree,
3459                              struct extent_buffer *eb)
3460 {
3461         unsigned long i;
3462         unsigned long num_pages;
3463         int was_dirty = 0;
3464
3465         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3466         num_pages = num_extent_pages(eb->start, eb->len);
3467         for (i = 0; i < num_pages; i++)
3468                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3469         return was_dirty;
3470 }
3471
3472 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3473                                 struct extent_buffer *eb,
3474                                 struct extent_state **cached_state)
3475 {
3476         unsigned long i;
3477         struct page *page;
3478         unsigned long num_pages;
3479
3480         num_pages = num_extent_pages(eb->start, eb->len);
3481         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3482
3483         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3484                               cached_state, GFP_NOFS);
3485         for (i = 0; i < num_pages; i++) {
3486                 page = extent_buffer_page(eb, i);
3487                 if (page)
3488                         ClearPageUptodate(page);
3489         }
3490         return 0;
3491 }
3492
3493 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3494                                 struct extent_buffer *eb)
3495 {
3496         unsigned long i;
3497         struct page *page;
3498         unsigned long num_pages;
3499
3500         num_pages = num_extent_pages(eb->start, eb->len);
3501
3502         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3503                             NULL, GFP_NOFS);
3504         for (i = 0; i < num_pages; i++) {
3505                 page = extent_buffer_page(eb, i);
3506                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3507                     ((i == num_pages - 1) &&
3508                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3509                         check_page_uptodate(tree, page);
3510                         continue;
3511                 }
3512                 SetPageUptodate(page);
3513         }
3514         return 0;
3515 }
3516
3517 int extent_range_uptodate(struct extent_io_tree *tree,
3518                           u64 start, u64 end)
3519 {
3520         struct page *page;
3521         int ret;
3522         int pg_uptodate = 1;
3523         int uptodate;
3524         unsigned long index;
3525
3526         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL);
3527         if (ret)
3528                 return 1;
3529         while (start <= end) {
3530                 index = start >> PAGE_CACHE_SHIFT;
3531                 page = find_get_page(tree->mapping, index);
3532                 uptodate = PageUptodate(page);
3533                 page_cache_release(page);
3534                 if (!uptodate) {
3535                         pg_uptodate = 0;
3536                         break;
3537                 }
3538                 start += PAGE_CACHE_SIZE;
3539         }
3540         return pg_uptodate;
3541 }
3542
3543 int extent_buffer_uptodate(struct extent_io_tree *tree,
3544                            struct extent_buffer *eb,
3545                            struct extent_state *cached_state)
3546 {
3547         int ret = 0;
3548         unsigned long num_pages;
3549         unsigned long i;
3550         struct page *page;
3551         int pg_uptodate = 1;
3552
3553         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3554                 return 1;
3555
3556         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3557                            EXTENT_UPTODATE, 1, cached_state);
3558         if (ret)
3559                 return ret;
3560
3561         num_pages = num_extent_pages(eb->start, eb->len);
3562         for (i = 0; i < num_pages; i++) {
3563                 page = extent_buffer_page(eb, i);
3564                 if (!PageUptodate(page)) {
3565                         pg_uptodate = 0;
3566                         break;
3567                 }
3568         }
3569         return pg_uptodate;
3570 }
3571
3572 int read_extent_buffer_pages(struct extent_io_tree *tree,
3573                              struct extent_buffer *eb,
3574                              u64 start, int wait,
3575                              get_extent_t *get_extent, int mirror_num)
3576 {
3577         unsigned long i;
3578         unsigned long start_i;
3579         struct page *page;
3580         int err;
3581         int ret = 0;
3582         int locked_pages = 0;
3583         int all_uptodate = 1;
3584         int inc_all_pages = 0;
3585         unsigned long num_pages;
3586         struct bio *bio = NULL;
3587         unsigned long bio_flags = 0;
3588
3589         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3590                 return 0;
3591
3592         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3593                            EXTENT_UPTODATE, 1, NULL)) {
3594                 return 0;
3595         }
3596
3597         if (start) {
3598                 WARN_ON(start < eb->start);
3599                 start_i = (start >> PAGE_CACHE_SHIFT) -
3600                         (eb->start >> PAGE_CACHE_SHIFT);
3601         } else {
3602                 start_i = 0;
3603         }
3604
3605         num_pages = num_extent_pages(eb->start, eb->len);
3606         for (i = start_i; i < num_pages; i++) {
3607                 page = extent_buffer_page(eb, i);
3608                 if (!wait) {
3609                         if (!trylock_page(page))
3610                                 goto unlock_exit;
3611                 } else {
3612                         lock_page(page);
3613                 }
3614                 locked_pages++;
3615                 if (!PageUptodate(page))
3616                         all_uptodate = 0;
3617         }
3618         if (all_uptodate) {
3619                 if (start_i == 0)
3620                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3621                 goto unlock_exit;
3622         }
3623
3624         for (i = start_i; i < num_pages; i++) {
3625                 page = extent_buffer_page(eb, i);
3626
3627                 WARN_ON(!PagePrivate(page));
3628
3629                 set_page_extent_mapped(page);
3630                 if (i == 0)
3631                         set_page_extent_head(page, eb->len);
3632
3633                 if (inc_all_pages)
3634                         page_cache_get(page);
3635                 if (!PageUptodate(page)) {
3636                         if (start_i == 0)
3637                                 inc_all_pages = 1;
3638                         ClearPageError(page);
3639                         err = __extent_read_full_page(tree, page,
3640                                                       get_extent, &bio,
3641                                                       mirror_num, &bio_flags);
3642                         if (err)
3643                                 ret = err;
3644                 } else {
3645                         unlock_page(page);
3646                 }
3647         }
3648
3649         if (bio)
3650                 submit_one_bio(READ, bio, mirror_num, bio_flags);
3651
3652         if (ret || !wait)
3653                 return ret;
3654
3655         for (i = start_i; i < num_pages; i++) {
3656                 page = extent_buffer_page(eb, i);
3657                 wait_on_page_locked(page);
3658                 if (!PageUptodate(page))
3659                         ret = -EIO;
3660         }
3661
3662         if (!ret)
3663                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3664         return ret;
3665
3666 unlock_exit:
3667         i = start_i;
3668         while (locked_pages > 0) {
3669                 page = extent_buffer_page(eb, i);
3670                 i++;
3671                 unlock_page(page);
3672                 locked_pages--;
3673         }
3674         return ret;
3675 }
3676
3677 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3678                         unsigned long start,
3679                         unsigned long len)
3680 {
3681         size_t cur;
3682         size_t offset;
3683         struct page *page;
3684         char *kaddr;
3685         char *dst = (char *)dstv;
3686         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3687         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3688
3689         WARN_ON(start > eb->len);
3690         WARN_ON(start + len > eb->start + eb->len);
3691
3692         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3693
3694         while (len > 0) {
3695                 page = extent_buffer_page(eb, i);
3696
3697                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3698                 kaddr = kmap_atomic(page, KM_USER1);
3699                 memcpy(dst, kaddr + offset, cur);
3700                 kunmap_atomic(kaddr, KM_USER1);
3701
3702                 dst += cur;
3703                 len -= cur;
3704                 offset = 0;
3705                 i++;
3706         }
3707 }
3708
3709 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3710                                unsigned long min_len, char **token, char **map,
3711                                unsigned long *map_start,
3712                                unsigned long *map_len, int km)
3713 {
3714         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3715         char *kaddr;
3716         struct page *p;
3717         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3718         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3719         unsigned long end_i = (start_offset + start + min_len - 1) >>
3720                 PAGE_CACHE_SHIFT;
3721
3722         if (i != end_i)
3723                 return -EINVAL;
3724
3725         if (i == 0) {
3726                 offset = start_offset;
3727                 *map_start = 0;
3728         } else {
3729                 offset = 0;
3730                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3731         }
3732
3733         if (start + min_len > eb->len) {
3734                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3735                        "wanted %lu %lu\n", (unsigned long long)eb->start,
3736                        eb->len, start, min_len);
3737                 WARN_ON(1);
3738                 return -EINVAL;
3739         }
3740
3741         p = extent_buffer_page(eb, i);
3742         kaddr = kmap_atomic(p, km);
3743         *token = kaddr;
3744         *map = kaddr + offset;
3745         *map_len = PAGE_CACHE_SIZE - offset;
3746         return 0;
3747 }
3748
3749 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3750                       unsigned long min_len,
3751                       char **token, char **map,
3752                       unsigned long *map_start,
3753                       unsigned long *map_len, int km)
3754 {
3755         int err;
3756         int save = 0;
3757         if (eb->map_token) {
3758                 unmap_extent_buffer(eb, eb->map_token, km);
3759                 eb->map_token = NULL;
3760                 save = 1;
3761         }
3762         err = map_private_extent_buffer(eb, start, min_len, token, map,
3763                                        map_start, map_len, km);
3764         if (!err && save) {
3765                 eb->map_token = *token;
3766                 eb->kaddr = *map;
3767                 eb->map_start = *map_start;
3768                 eb->map_len = *map_len;
3769         }
3770         return err;
3771 }
3772
3773 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3774 {
3775         kunmap_atomic(token, km);
3776 }
3777
3778 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3779                           unsigned long start,
3780                           unsigned long len)
3781 {
3782         size_t cur;
3783         size_t offset;
3784         struct page *page;
3785         char *kaddr;
3786         char *ptr = (char *)ptrv;
3787         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3788         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3789         int ret = 0;
3790
3791         WARN_ON(start > eb->len);
3792         WARN_ON(start + len > eb->start + eb->len);
3793
3794         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3795
3796         while (len > 0) {
3797                 page = extent_buffer_page(eb, i);
3798
3799                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3800
3801                 kaddr = kmap_atomic(page, KM_USER0);
3802                 ret = memcmp(ptr, kaddr + offset, cur);
3803                 kunmap_atomic(kaddr, KM_USER0);
3804                 if (ret)
3805                         break;
3806
3807                 ptr += cur;
3808                 len -= cur;
3809                 offset = 0;
3810                 i++;
3811         }
3812         return ret;
3813 }
3814
3815 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3816                          unsigned long start, unsigned long len)
3817 {
3818         size_t cur;
3819         size_t offset;
3820         struct page *page;
3821         char *kaddr;
3822         char *src = (char *)srcv;
3823         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3824         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3825
3826         WARN_ON(start > eb->len);
3827         WARN_ON(start + len > eb->start + eb->len);
3828
3829         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3830
3831         while (len > 0) {
3832                 page = extent_buffer_page(eb, i);
3833                 WARN_ON(!PageUptodate(page));
3834
3835                 cur = min(len, PAGE_CACHE_SIZE - offset);
3836                 kaddr = kmap_atomic(page, KM_USER1);
3837                 memcpy(kaddr + offset, src, cur);
3838                 kunmap_atomic(kaddr, KM_USER1);
3839
3840                 src += cur;
3841                 len -= cur;
3842                 offset = 0;
3843                 i++;
3844         }
3845 }
3846
3847 void memset_extent_buffer(struct extent_buffer *eb, char c,
3848                           unsigned long start, unsigned long len)
3849 {
3850         size_t cur;
3851         size_t offset;
3852         struct page *page;
3853         char *kaddr;
3854         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3855         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3856
3857         WARN_ON(start > eb->len);
3858         WARN_ON(start + len > eb->start + eb->len);
3859
3860         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3861
3862         while (len > 0) {
3863                 page = extent_buffer_page(eb, i);
3864                 WARN_ON(!PageUptodate(page));
3865
3866                 cur = min(len, PAGE_CACHE_SIZE - offset);
3867                 kaddr = kmap_atomic(page, KM_USER0);
3868                 memset(kaddr + offset, c, cur);
3869                 kunmap_atomic(kaddr, KM_USER0);
3870
3871                 len -= cur;
3872                 offset = 0;
3873                 i++;
3874         }
3875 }
3876
3877 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3878                         unsigned long dst_offset, unsigned long src_offset,
3879                         unsigned long len)
3880 {
3881         u64 dst_len = dst->len;
3882         size_t cur;
3883         size_t offset;
3884         struct page *page;
3885         char *kaddr;
3886         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3887         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3888
3889         WARN_ON(src->len != dst_len);
3890
3891         offset = (start_offset + dst_offset) &
3892                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3893
3894         while (len > 0) {
3895                 page = extent_buffer_page(dst, i);
3896                 WARN_ON(!PageUptodate(page));
3897
3898                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3899
3900                 kaddr = kmap_atomic(page, KM_USER0);
3901                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3902                 kunmap_atomic(kaddr, KM_USER0);
3903
3904                 src_offset += cur;
3905                 len -= cur;
3906                 offset = 0;
3907                 i++;
3908         }
3909 }
3910
3911 static void move_pages(struct page *dst_page, struct page *src_page,
3912                        unsigned long dst_off, unsigned long src_off,
3913                        unsigned long len)
3914 {
3915         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3916         if (dst_page == src_page) {
3917                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3918         } else {
3919                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3920                 char *p = dst_kaddr + dst_off + len;
3921                 char *s = src_kaddr + src_off + len;
3922
3923                 while (len--)
3924                         *--p = *--s;
3925
3926                 kunmap_atomic(src_kaddr, KM_USER1);
3927         }
3928         kunmap_atomic(dst_kaddr, KM_USER0);
3929 }
3930
3931 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
3932 {
3933         unsigned long distance = (src > dst) ? src - dst : dst - src;
3934         return distance < len;
3935 }
3936
3937 static void copy_pages(struct page *dst_page, struct page *src_page,
3938                        unsigned long dst_off, unsigned long src_off,
3939                        unsigned long len)
3940 {
3941         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3942         char *src_kaddr;
3943
3944         if (dst_page != src_page) {
3945                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3946         } else {
3947                 src_kaddr = dst_kaddr;
3948                 BUG_ON(areas_overlap(src_off, dst_off, len));
3949         }
3950
3951         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3952         kunmap_atomic(dst_kaddr, KM_USER0);
3953         if (dst_page != src_page)
3954                 kunmap_atomic(src_kaddr, KM_USER1);
3955 }
3956
3957 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3958                            unsigned long src_offset, unsigned long len)
3959 {
3960         size_t cur;
3961         size_t dst_off_in_page;
3962         size_t src_off_in_page;
3963         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3964         unsigned long dst_i;
3965         unsigned long src_i;
3966
3967         if (src_offset + len > dst->len) {
3968                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3969                        "len %lu dst len %lu\n", src_offset, len, dst->len);
3970                 BUG_ON(1);
3971         }
3972         if (dst_offset + len > dst->len) {
3973                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3974                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
3975                 BUG_ON(1);
3976         }
3977
3978         while (len > 0) {
3979                 dst_off_in_page = (start_offset + dst_offset) &
3980                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3981                 src_off_in_page = (start_offset + src_offset) &
3982                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3983
3984                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3985                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3986
3987                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3988                                                src_off_in_page));
3989                 cur = min_t(unsigned long, cur,
3990                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3991
3992                 copy_pages(extent_buffer_page(dst, dst_i),
3993                            extent_buffer_page(dst, src_i),
3994                            dst_off_in_page, src_off_in_page, cur);
3995
3996                 src_offset += cur;
3997                 dst_offset += cur;
3998                 len -= cur;
3999         }
4000 }
4001
4002 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4003                            unsigned long src_offset, unsigned long len)
4004 {
4005         size_t cur;
4006         size_t dst_off_in_page;
4007         size_t src_off_in_page;
4008         unsigned long dst_end = dst_offset + len - 1;
4009         unsigned long src_end = src_offset + len - 1;
4010         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4011         unsigned long dst_i;
4012         unsigned long src_i;
4013
4014         if (src_offset + len > dst->len) {
4015                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4016                        "len %lu len %lu\n", src_offset, len, dst->len);
4017                 BUG_ON(1);
4018         }
4019         if (dst_offset + len > dst->len) {
4020                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4021                        "len %lu len %lu\n", dst_offset, len, dst->len);
4022                 BUG_ON(1);
4023         }
4024         if (!areas_overlap(src_offset, dst_offset, len)) {
4025                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4026                 return;
4027         }
4028         while (len > 0) {
4029                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4030                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4031
4032                 dst_off_in_page = (start_offset + dst_end) &
4033                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4034                 src_off_in_page = (start_offset + src_end) &
4035                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4036
4037                 cur = min_t(unsigned long, len, src_off_in_page + 1);
4038                 cur = min(cur, dst_off_in_page + 1);
4039                 move_pages(extent_buffer_page(dst, dst_i),
4040                            extent_buffer_page(dst, src_i),
4041                            dst_off_in_page - cur + 1,
4042                            src_off_in_page - cur + 1, cur);
4043
4044                 dst_end -= cur;
4045                 src_end -= cur;
4046                 len -= cur;
4047         }
4048 }
4049
4050 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4051 {
4052         struct extent_buffer *eb =
4053                         container_of(head, struct extent_buffer, rcu_head);
4054
4055         btrfs_release_extent_buffer(eb);
4056 }
4057
4058 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
4059 {
4060         u64 start = page_offset(page);
4061         struct extent_buffer *eb;
4062         int ret = 1;
4063
4064         spin_lock(&tree->buffer_lock);
4065         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4066         if (!eb) {
4067                 spin_unlock(&tree->buffer_lock);
4068                 return ret;
4069         }
4070
4071         if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
4072                 ret = 0;
4073                 goto out;
4074         }
4075
4076         /*
4077          * set @eb->refs to 0 if it is already 1, and then release the @eb.
4078          * Or go back.
4079          */
4080         if (atomic_cmpxchg(&eb->refs, 1, 0) != 1) {
4081                 ret = 0;
4082                 goto out;
4083         }
4084
4085         radix_tree_delete(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4086 out:
4087         spin_unlock(&tree->buffer_lock);
4088
4089         /* at this point we can safely release the extent buffer */
4090         if (atomic_read(&eb->refs) == 0)
4091                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4092         return ret;
4093 }