git.oblomov.eu Git - linux-2.6/blob - fs/btrfs/extent_io.c

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