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