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         state = rb_entry(node, struct extent_state, rb_node);
 698 hit_next:
 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                 struct rb_node *next_node;
 710                 set = state->state & bits;
 711                 if (set && exclusive) {
 712                         *failed_start = state->start;
 713                         err = -EEXIST;
 714                         goto out;
 715                 }
 716                 set_state_bits(tree, state, bits);
 717                 merge_state(tree, state);
 718                 if (last_end == (u64)-1)
 719                         goto out;
 720
 721                 start = last_end + 1;
 722                 if (start < end && prealloc && !need_resched()) {
 723                         next_node = rb_next(node);
 724                         if (next_node) {
 725                                 state = rb_entry(next_node, struct extent_state,
 726                                                  rb_node);
 727                                 if (state->start == start)
 728                                         goto hit_next;
 729                         }
 730                 }
 731                 goto search_again;
 732         }
 733
 734         /*
 735          *     | ---- desired range ---- |
 736          * | state |
 737          *   or
 738          * | ------------- state -------------- |
 739          *
 740          * We need to split the extent we found, and may flip bits on
 741          * second half.
 742          *
 743          * If the extent we found extends past our
 744          * range, we just split and search again.  It'll get split
 745          * again the next time though.
 746          *
 747          * If the extent we found is inside our range, we set the
 748          * desired bit on it.
 749          */
 750         if (state->start < start) {
 751                 set = state->state & bits;
 752                 if (exclusive && set) {
 753                         *failed_start = start;
 754                         err = -EEXIST;
 755                         goto out;
 756                 }
 757                 err = split_state(tree, state, prealloc, start);
 758                 BUG_ON(err == -EEXIST);
 759                 prealloc = NULL;
 760                 if (err)
 761                         goto out;
 762                 if (state->end <= end) {
 763                         set_state_bits(tree, state, bits);
 764                         merge_state(tree, state);
 765                         if (last_end == (u64)-1)
 766                                 goto out;
 767                         start = last_end + 1;
 768                 } else {
 769                         start = state->start;
 770                 }
 771                 goto search_again;
 772         }
 773         /*
 774          * | ---- desired range ---- |
 775          *     | state | or               | state |
 776          *
 777          * There's a hole, we need to insert something in it and
 778          * ignore the extent we found.
 779          */
 780         if (state->start > start) {
 781                 u64 this_end;
 782                 if (end < last_start)
 783                         this_end = end;
 784                 else
 785                         this_end = last_start - 1;
 786                 err = insert_state(tree, prealloc, start, this_end,
 787                                    bits);
 788                 prealloc = NULL;
 789                 BUG_ON(err == -EEXIST);
 790                 if (err)
 791                         goto out;
 792                 start = this_end + 1;
 793                 goto search_again;
 794         }
 795         /*
 796          * | ---- desired range ---- |
 797          *                        | state |
 798          * We need to split the extent, and set the bit
 799          * on the first half
 800          */
 801         if (state->start <= end && state->end > end) {
 802                 set = state->state & bits;
 803                 if (exclusive && set) {
 804                         *failed_start = start;
 805                         err = -EEXIST;
 806                         goto out;
 807                 }
 808                 err = split_state(tree, state, prealloc, end + 1);
 809                 BUG_ON(err == -EEXIST);
 810
 811                 set_state_bits(tree, prealloc, bits);
 812                 merge_state(tree, prealloc);
 813                 prealloc = NULL;
 814                 goto out;
 815         }
 816
 817         goto search_again;
 818
 819 out:
 820         spin_unlock(&tree->lock);
 821         if (prealloc)
 822                 free_extent_state(prealloc);
 823
 824         return err;
 825
 826 search_again:
 827         if (start > end)
 828                 goto out;
 829         spin_unlock(&tree->lock);
 830         if (mask & __GFP_WAIT)
 831                 cond_resched();
 832         goto again;
 833 }
 834
 835 /* wrappers around set/clear extent bit */
 836 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 837                      gfp_t mask)
 838 {
 839         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 840                               mask);
 841 }
 842
 843 int set_extent_ordered(struct extent_io_tree *tree, u64 start, u64 end,
 844                        gfp_t mask)
 845 {
 846         return set_extent_bit(tree, start, end, EXTENT_ORDERED, 0, NULL, mask);
 847 }
 848
 849 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 850                     int bits, gfp_t mask)
 851 {
 852         return set_extent_bit(tree, start, end, bits, 0, NULL,
 853                               mask);
 854 }
 855
 856 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 857                       int bits, gfp_t mask)
 858 {
 859         return clear_extent_bit(tree, start, end, bits, 0, 0, mask);
 860 }
 861
 862 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
 863                      gfp_t mask)
 864 {
 865         return set_extent_bit(tree, start, end,
 866                               EXTENT_DELALLOC | EXTENT_DIRTY | EXTENT_UPTODATE,
 867                               0, NULL, mask);
 868 }
 869
 870 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 871                        gfp_t mask)
 872 {
 873         return clear_extent_bit(tree, start, end,
 874                                 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, mask);
 875 }
 876
 877 int clear_extent_ordered(struct extent_io_tree *tree, u64 start, u64 end,
 878                          gfp_t mask)
 879 {
 880         return clear_extent_bit(tree, start, end, EXTENT_ORDERED, 1, 0, mask);
 881 }
 882
 883 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 884                      gfp_t mask)
 885 {
 886         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 887                               mask);
 888 }
 889
 890 static int clear_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 891                        gfp_t mask)
 892 {
 893         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0, mask);
 894 }
 895
 896 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 897                         gfp_t mask)
 898 {
 899         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, NULL,
 900                               mask);
 901 }
 902
 903 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
 904                                  u64 end, gfp_t mask)
 905 {
 906         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0, mask);
 907 }
 908
 909 static int set_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end,
 910                          gfp_t mask)
 911 {
 912         return set_extent_bit(tree, start, end, EXTENT_WRITEBACK,
 913                               0, NULL, mask);
 914 }
 915
 916 static int clear_extent_writeback(struct extent_io_tree *tree, u64 start,
 917                                   u64 end, gfp_t mask)
 918 {
 919         return clear_extent_bit(tree, start, end, EXTENT_WRITEBACK, 1, 0, mask);
 920 }
 921
 922 int wait_on_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 923 {
 924         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 925 }
 926
 927 /*
 928  * either insert or lock state struct between start and end use mask to tell
 929  * us if waiting is desired.
 930  */
 931 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
 932 {
 933         int err;
 934         u64 failed_start;
 935         while (1) {
 936                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED, 1,
 937                                      &failed_start, mask);
 938                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
 939                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
 940                         start = failed_start;
 941                 } else {
 942                         break;
 943                 }
 944                 WARN_ON(start > end);
 945         }
 946         return err;
 947 }
 948
 949 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
 950                     gfp_t mask)
 951 {
 952         int err;
 953         u64 failed_start;
 954
 955         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, 1,
 956                              &failed_start, mask);
 957         if (err == -EEXIST) {
 958                 if (failed_start > start)
 959                         clear_extent_bit(tree, start, failed_start - 1,
 960                                          EXTENT_LOCKED, 1, 0, mask);
 961                 return 0;
 962         }
 963         return 1;
 964 }
 965
 966 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end,
 967                   gfp_t mask)
 968 {
 969         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, mask);
 970 }
 971
 972 /*
 973  * helper function to set pages and extents in the tree dirty
 974  */
 975 int set_range_dirty(struct extent_io_tree *tree, u64 start, u64 end)
 976 {
 977         unsigned long index = start >> PAGE_CACHE_SHIFT;
 978         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 979         struct page *page;
 980
 981         while (index <= end_index) {
 982                 page = find_get_page(tree->mapping, index);
 983                 BUG_ON(!page);
 984                 __set_page_dirty_nobuffers(page);
 985                 page_cache_release(page);
 986                 index++;
 987         }
 988         set_extent_dirty(tree, start, end, GFP_NOFS);
 989         return 0;
 990 }
 991
 992 /*
 993  * helper function to set both pages and extents in the tree writeback
 994  */
 995 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 996 {
 997         unsigned long index = start >> PAGE_CACHE_SHIFT;
 998         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 999         struct page *page;
1000
1001         while (index <= end_index) {
1002                 page = find_get_page(tree->mapping, index);
1003                 BUG_ON(!page);
1004                 set_page_writeback(page);
1005                 page_cache_release(page);
1006                 index++;
1007         }
1008         set_extent_writeback(tree, start, end, GFP_NOFS);
1009         return 0;
1010 }
1011
1012 /*
1013  * find the first offset in the io tree with 'bits' set. zero is
1014  * returned if we find something, and *start_ret and *end_ret are
1015  * set to reflect the state struct that was found.
1016  *
1017  * If nothing was found, 1 is returned, < 0 on error
1018  */
1019 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1020                           u64 *start_ret, u64 *end_ret, int bits)
1021 {
1022         struct rb_node *node;
1023         struct extent_state *state;
1024         int ret = 1;
1025
1026         spin_lock(&tree->lock);
1027         /*
1028          * this search will find all the extents that end after
1029          * our range starts.
1030          */
1031         node = tree_search(tree, start);
1032         if (!node)
1033                 goto out;
1034
1035         while (1) {
1036                 state = rb_entry(node, struct extent_state, rb_node);
1037                 if (state->end >= start && (state->state & bits)) {
1038                         *start_ret = state->start;
1039                         *end_ret = state->end;
1040                         ret = 0;
1041                         break;
1042                 }
1043                 node = rb_next(node);
1044                 if (!node)
1045                         break;
1046         }
1047 out:
1048         spin_unlock(&tree->lock);
1049         return ret;
1050 }
1051
1052 /* find the first state struct with 'bits' set after 'start', and
1053  * return it.  tree->lock must be held.  NULL will returned if
1054  * nothing was found after 'start'
1055  */
1056 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1057                                                  u64 start, int bits)
1058 {
1059         struct rb_node *node;
1060         struct extent_state *state;
1061
1062         /*
1063          * this search will find all the extents that end after
1064          * our range starts.
1065          */
1066         node = tree_search(tree, start);
1067         if (!node)
1068                 goto out;
1069
1070         while (1) {
1071                 state = rb_entry(node, struct extent_state, rb_node);
1072                 if (state->end >= start && (state->state & bits))
1073                         return state;
1074
1075                 node = rb_next(node);
1076                 if (!node)
1077                         break;
1078         }
1079 out:
1080         return NULL;
1081 }
1082
1083 /*
1084  * find a contiguous range of bytes in the file marked as delalloc, not
1085  * more than 'max_bytes'.  start and end are used to return the range,
1086  *
1087  * 1 is returned if we find something, 0 if nothing was in the tree
1088  */
1089 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1090                                         u64 *start, u64 *end, u64 max_bytes)
1091 {
1092         struct rb_node *node;
1093         struct extent_state *state;
1094         u64 cur_start = *start;
1095         u64 found = 0;
1096         u64 total_bytes = 0;
1097
1098         spin_lock(&tree->lock);
1099
1100         /*
1101          * this search will find all the extents that end after
1102          * our range starts.
1103          */
1104         node = tree_search(tree, cur_start);
1105         if (!node) {
1106                 if (!found)
1107                         *end = (u64)-1;
1108                 goto out;
1109         }
1110
1111         while (1) {
1112                 state = rb_entry(node, struct extent_state, rb_node);
1113                 if (found && (state->start != cur_start ||
1114                               (state->state & EXTENT_BOUNDARY))) {
1115                         goto out;
1116                 }
1117                 if (!(state->state & EXTENT_DELALLOC)) {
1118                         if (!found)
1119                                 *end = state->end;
1120                         goto out;
1121                 }
1122                 if (!found)
1123                         *start = state->start;
1124                 found++;
1125                 *end = state->end;
1126                 cur_start = state->end + 1;
1127                 node = rb_next(node);
1128                 if (!node)
1129                         break;
1130                 total_bytes += state->end - state->start + 1;
1131                 if (total_bytes >= max_bytes)
1132                         break;
1133         }
1134 out:
1135         spin_unlock(&tree->lock);
1136         return found;
1137 }
1138
1139 static noinline int __unlock_for_delalloc(struct inode *inode,
1140                                           struct page *locked_page,
1141                                           u64 start, u64 end)
1142 {
1143         int ret;
1144         struct page *pages[16];
1145         unsigned long index = start >> PAGE_CACHE_SHIFT;
1146         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1147         unsigned long nr_pages = end_index - index + 1;
1148         int i;
1149
1150         if (index == locked_page->index && end_index == index)
1151                 return 0;
1152
1153         while (nr_pages > 0) {
1154                 ret = find_get_pages_contig(inode->i_mapping, index,
1155                                      min_t(unsigned long, nr_pages,
1156                                      ARRAY_SIZE(pages)), pages);
1157                 for (i = 0; i < ret; i++) {
1158                         if (pages[i] != locked_page)
1159                                 unlock_page(pages[i]);
1160                         page_cache_release(pages[i]);
1161                 }
1162                 nr_pages -= ret;
1163                 index += ret;
1164                 cond_resched();
1165         }
1166         return 0;
1167 }
1168
1169 static noinline int lock_delalloc_pages(struct inode *inode,
1170                                         struct page *locked_page,
1171                                         u64 delalloc_start,
1172                                         u64 delalloc_end)
1173 {
1174         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1175         unsigned long start_index = index;
1176         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1177         unsigned long pages_locked = 0;
1178         struct page *pages[16];
1179         unsigned long nrpages;
1180         int ret;
1181         int i;
1182
1183         /* the caller is responsible for locking the start index */
1184         if (index == locked_page->index && index == end_index)
1185                 return 0;
1186
1187         /* skip the page at the start index */
1188         nrpages = end_index - index + 1;
1189         while (nrpages > 0) {
1190                 ret = find_get_pages_contig(inode->i_mapping, index,
1191                                      min_t(unsigned long,
1192                                      nrpages, ARRAY_SIZE(pages)), pages);
1193                 if (ret == 0) {
1194                         ret = -EAGAIN;
1195                         goto done;
1196                 }
1197                 /* now we have an array of pages, lock them all */
1198                 for (i = 0; i < ret; i++) {
1199                         /*
1200                          * the caller is taking responsibility for
1201                          * locked_page
1202                          */
1203                         if (pages[i] != locked_page) {
1204                                 lock_page(pages[i]);
1205                                 if (!PageDirty(pages[i]) ||
1206                                     pages[i]->mapping != inode->i_mapping) {
1207                                         ret = -EAGAIN;
1208                                         unlock_page(pages[i]);
1209                                         page_cache_release(pages[i]);
1210                                         goto done;
1211                                 }
1212                         }
1213                         page_cache_release(pages[i]);
1214                         pages_locked++;
1215                 }
1216                 nrpages -= ret;
1217                 index += ret;
1218                 cond_resched();
1219         }
1220         ret = 0;
1221 done:
1222         if (ret && pages_locked) {
1223                 __unlock_for_delalloc(inode, locked_page,
1224                               delalloc_start,
1225                               ((u64)(start_index + pages_locked - 1)) <<
1226                               PAGE_CACHE_SHIFT);
1227         }
1228         return ret;
1229 }
1230
1231 /*
1232  * find a contiguous range of bytes in the file marked as delalloc, not
1233  * more than 'max_bytes'.  start and end are used to return the range,
1234  *
1235  * 1 is returned if we find something, 0 if nothing was in the tree
1236  */
1237 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1238                                              struct extent_io_tree *tree,
1239                                              struct page *locked_page,
1240                                              u64 *start, u64 *end,
1241                                              u64 max_bytes)
1242 {
1243         u64 delalloc_start;
1244         u64 delalloc_end;
1245         u64 found;
1246         int ret;
1247         int loops = 0;
1248
1249 again:
1250         /* step one, find a bunch of delalloc bytes starting at start */
1251         delalloc_start = *start;
1252         delalloc_end = 0;
1253         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1254                                     max_bytes);
1255         if (!found || delalloc_end <= *start) {
1256                 *start = delalloc_start;
1257                 *end = delalloc_end;
1258                 return found;
1259         }
1260
1261         /*
1262          * start comes from the offset of locked_page.  We have to lock
1263          * pages in order, so we can't process delalloc bytes before
1264          * locked_page
1265          */
1266         if (delalloc_start < *start)
1267                 delalloc_start = *start;
1268
1269         /*
1270          * make sure to limit the number of pages we try to lock down
1271          * if we're looping.
1272          */
1273         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1274                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1275
1276         /* step two, lock all the pages after the page that has start */
1277         ret = lock_delalloc_pages(inode, locked_page,
1278                                   delalloc_start, delalloc_end);
1279         if (ret == -EAGAIN) {
1280                 /* some of the pages are gone, lets avoid looping by
1281                  * shortening the size of the delalloc range we're searching
1282                  */
1283                 if (!loops) {
1284                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1285                         max_bytes = PAGE_CACHE_SIZE - offset;
1286                         loops = 1;
1287                         goto again;
1288                 } else {
1289                         found = 0;
1290                         goto out_failed;
1291                 }
1292         }
1293         BUG_ON(ret);
1294
1295         /* step three, lock the state bits for the whole range */
1296         lock_extent(tree, delalloc_start, delalloc_end, GFP_NOFS);
1297
1298         /* then test to make sure it is all still delalloc */
1299         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1300                              EXTENT_DELALLOC, 1);
1301         if (!ret) {
1302                 unlock_extent(tree, delalloc_start, delalloc_end, GFP_NOFS);
1303                 __unlock_for_delalloc(inode, locked_page,
1304                               delalloc_start, delalloc_end);
1305                 cond_resched();
1306                 goto again;
1307         }
1308         *start = delalloc_start;
1309         *end = delalloc_end;
1310 out_failed:
1311         return found;
1312 }
1313
1314 int extent_clear_unlock_delalloc(struct inode *inode,
1315                                 struct extent_io_tree *tree,
1316                                 u64 start, u64 end, struct page *locked_page,
1317                                 int unlock_pages,
1318                                 int clear_unlock,
1319                                 int clear_delalloc, int clear_dirty,
1320                                 int set_writeback,
1321                                 int end_writeback)
1322 {
1323         int ret;
1324         struct page *pages[16];
1325         unsigned long index = start >> PAGE_CACHE_SHIFT;
1326         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1327         unsigned long nr_pages = end_index - index + 1;
1328         int i;
1329         int clear_bits = 0;
1330
1331         if (clear_unlock)
1332                 clear_bits |= EXTENT_LOCKED;
1333         if (clear_dirty)
1334                 clear_bits |= EXTENT_DIRTY;
1335
1336         if (clear_delalloc)
1337                 clear_bits |= EXTENT_DELALLOC;
1338
1339         clear_extent_bit(tree, start, end, clear_bits, 1, 0, GFP_NOFS);
1340         if (!(unlock_pages || clear_dirty || set_writeback || end_writeback))
1341                 return 0;
1342
1343         while (nr_pages > 0) {
1344                 ret = find_get_pages_contig(inode->i_mapping, index,
1345                                      min_t(unsigned long,
1346                                      nr_pages, ARRAY_SIZE(pages)), pages);
1347                 for (i = 0; i < ret; i++) {
1348                         if (pages[i] == locked_page) {
1349                                 page_cache_release(pages[i]);
1350                                 continue;
1351                         }
1352                         if (clear_dirty)
1353                                 clear_page_dirty_for_io(pages[i]);
1354                         if (set_writeback)
1355                                 set_page_writeback(pages[i]);
1356                         if (end_writeback)
1357                                 end_page_writeback(pages[i]);
1358                         if (unlock_pages)
1359                                 unlock_page(pages[i]);
1360                         page_cache_release(pages[i]);
1361                 }
1362                 nr_pages -= ret;
1363                 index += ret;
1364                 cond_resched();
1365         }
1366         return 0;
1367 }
1368
1369 /*
1370  * count the number of bytes in the tree that have a given bit(s)
1371  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1372  * cached.  The total number found is returned.
1373  */
1374 u64 count_range_bits(struct extent_io_tree *tree,
1375                      u64 *start, u64 search_end, u64 max_bytes,
1376                      unsigned long bits)
1377 {
1378         struct rb_node *node;
1379         struct extent_state *state;
1380         u64 cur_start = *start;
1381         u64 total_bytes = 0;
1382         int found = 0;
1383
1384         if (search_end <= cur_start) {
1385                 WARN_ON(1);
1386                 return 0;
1387         }
1388
1389         spin_lock(&tree->lock);
1390         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1391                 total_bytes = tree->dirty_bytes;
1392                 goto out;
1393         }
1394         /*
1395          * this search will find all the extents that end after
1396          * our range starts.
1397          */
1398         node = tree_search(tree, cur_start);
1399         if (!node)
1400                 goto out;
1401
1402         while (1) {
1403                 state = rb_entry(node, struct extent_state, rb_node);
1404                 if (state->start > search_end)
1405                         break;
1406                 if (state->end >= cur_start && (state->state & bits)) {
1407                         total_bytes += min(search_end, state->end) + 1 -
1408                                        max(cur_start, state->start);
1409                         if (total_bytes >= max_bytes)
1410                                 break;
1411                         if (!found) {
1412                                 *start = state->start;
1413                                 found = 1;
1414                         }
1415                 }
1416                 node = rb_next(node);
1417                 if (!node)
1418                         break;
1419         }
1420 out:
1421         spin_unlock(&tree->lock);
1422         return total_bytes;
1423 }
1424
1425 /*
1426  * set the private field for a given byte offset in the tree.  If there isn't
1427  * an extent_state there already, this does nothing.
1428  */
1429 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1430 {
1431         struct rb_node *node;
1432         struct extent_state *state;
1433         int ret = 0;
1434
1435         spin_lock(&tree->lock);
1436         /*
1437          * this search will find all the extents that end after
1438          * our range starts.
1439          */
1440         node = tree_search(tree, start);
1441         if (!node) {
1442                 ret = -ENOENT;
1443                 goto out;
1444         }
1445         state = rb_entry(node, struct extent_state, rb_node);
1446         if (state->start != start) {
1447                 ret = -ENOENT;
1448                 goto out;
1449         }
1450         state->private = private;
1451 out:
1452         spin_unlock(&tree->lock);
1453         return ret;
1454 }
1455
1456 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1457 {
1458         struct rb_node *node;
1459         struct extent_state *state;
1460         int ret = 0;
1461
1462         spin_lock(&tree->lock);
1463         /*
1464          * this search will find all the extents that end after
1465          * our range starts.
1466          */
1467         node = tree_search(tree, start);
1468         if (!node) {
1469                 ret = -ENOENT;
1470                 goto out;
1471         }
1472         state = rb_entry(node, struct extent_state, rb_node);
1473         if (state->start != start) {
1474                 ret = -ENOENT;
1475                 goto out;
1476         }
1477         *private = state->private;
1478 out:
1479         spin_unlock(&tree->lock);
1480         return ret;
1481 }
1482
1483 /*
1484  * searches a range in the state tree for a given mask.
1485  * If 'filled' == 1, this returns 1 only if every extent in the tree
1486  * has the bits set.  Otherwise, 1 is returned if any bit in the
1487  * range is found set.
1488  */
1489 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1490                    int bits, int filled)
1491 {
1492         struct extent_state *state = NULL;
1493         struct rb_node *node;
1494         int bitset = 0;
1495
1496         spin_lock(&tree->lock);
1497         node = tree_search(tree, start);
1498         while (node && start <= end) {
1499                 state = rb_entry(node, struct extent_state, rb_node);
1500
1501                 if (filled && state->start > start) {
1502                         bitset = 0;
1503                         break;
1504                 }
1505
1506                 if (state->start > end)
1507                         break;
1508
1509                 if (state->state & bits) {
1510                         bitset = 1;
1511                         if (!filled)
1512                                 break;
1513                 } else if (filled) {
1514                         bitset = 0;
1515                         break;
1516                 }
1517                 start = state->end + 1;
1518                 if (start > end)
1519                         break;
1520                 node = rb_next(node);
1521                 if (!node) {
1522                         if (filled)
1523                                 bitset = 0;
1524                         break;
1525                 }
1526         }
1527         spin_unlock(&tree->lock);
1528         return bitset;
1529 }
1530
1531 /*
1532  * helper function to set a given page up to date if all the
1533  * extents in the tree for that page are up to date
1534  */
1535 static int check_page_uptodate(struct extent_io_tree *tree,
1536                                struct page *page)
1537 {
1538         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1539         u64 end = start + PAGE_CACHE_SIZE - 1;
1540         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1))
1541                 SetPageUptodate(page);
1542         return 0;
1543 }
1544
1545 /*
1546  * helper function to unlock a page if all the extents in the tree
1547  * for that page are unlocked
1548  */
1549 static int check_page_locked(struct extent_io_tree *tree,
1550                              struct page *page)
1551 {
1552         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1553         u64 end = start + PAGE_CACHE_SIZE - 1;
1554         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0))
1555                 unlock_page(page);
1556         return 0;
1557 }
1558
1559 /*
1560  * helper function to end page writeback if all the extents
1561  * in the tree for that page are done with writeback
1562  */
1563 static int check_page_writeback(struct extent_io_tree *tree,
1564                              struct page *page)
1565 {
1566         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1567         u64 end = start + PAGE_CACHE_SIZE - 1;
1568         if (!test_range_bit(tree, start, end, EXTENT_WRITEBACK, 0))
1569                 end_page_writeback(page);
1570         return 0;
1571 }
1572
1573 /* lots and lots of room for performance fixes in the end_bio funcs */
1574
1575 /*
1576  * after a writepage IO is done, we need to:
1577  * clear the uptodate bits on error
1578  * clear the writeback bits in the extent tree for this IO
1579  * end_page_writeback if the page has no more pending IO
1580  *
1581  * Scheduling is not allowed, so the extent state tree is expected
1582  * to have one and only one object corresponding to this IO.
1583  */
1584 static void end_bio_extent_writepage(struct bio *bio, int err)
1585 {
1586         int uptodate = err == 0;
1587         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1588         struct extent_io_tree *tree;
1589         u64 start;
1590         u64 end;
1591         int whole_page;
1592         int ret;
1593
1594         do {
1595                 struct page *page = bvec->bv_page;
1596                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1597
1598                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1599                          bvec->bv_offset;
1600                 end = start + bvec->bv_len - 1;
1601
1602                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1603                         whole_page = 1;
1604                 else
1605                         whole_page = 0;
1606
1607                 if (--bvec >= bio->bi_io_vec)
1608                         prefetchw(&bvec->bv_page->flags);
1609                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1610                         ret = tree->ops->writepage_end_io_hook(page, start,
1611                                                        end, NULL, uptodate);
1612                         if (ret)
1613                                 uptodate = 0;
1614                 }
1615
1616                 if (!uptodate && tree->ops &&
1617                     tree->ops->writepage_io_failed_hook) {
1618                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1619                                                          start, end, NULL);
1620                         if (ret == 0) {
1621                                 uptodate = (err == 0);
1622                                 continue;
1623                         }
1624                 }
1625
1626                 if (!uptodate) {
1627                         clear_extent_uptodate(tree, start, end, GFP_ATOMIC);
1628                         ClearPageUptodate(page);
1629                         SetPageError(page);
1630                 }
1631
1632                 clear_extent_writeback(tree, start, end, GFP_ATOMIC);
1633
1634                 if (whole_page)
1635                         end_page_writeback(page);
1636                 else
1637                         check_page_writeback(tree, page);
1638         } while (bvec >= bio->bi_io_vec);
1639
1640         bio_put(bio);
1641 }
1642
1643 /*
1644  * after a readpage IO is done, we need to:
1645  * clear the uptodate bits on error
1646  * set the uptodate bits if things worked
1647  * set the page up to date if all extents in the tree are uptodate
1648  * clear the lock bit in the extent tree
1649  * unlock the page if there are no other extents locked for it
1650  *
1651  * Scheduling is not allowed, so the extent state tree is expected
1652  * to have one and only one object corresponding to this IO.
1653  */
1654 static void end_bio_extent_readpage(struct bio *bio, int err)
1655 {
1656         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1657         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1658         struct extent_io_tree *tree;
1659         u64 start;
1660         u64 end;
1661         int whole_page;
1662         int ret;
1663
1664         if (err)
1665                 uptodate = 0;
1666
1667         do {
1668                 struct page *page = bvec->bv_page;
1669                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1670
1671                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1672                         bvec->bv_offset;
1673                 end = start + bvec->bv_len - 1;
1674
1675                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1676                         whole_page = 1;
1677                 else
1678                         whole_page = 0;
1679
1680                 if (--bvec >= bio->bi_io_vec)
1681                         prefetchw(&bvec->bv_page->flags);
1682
1683                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1684                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1685                                                               NULL);
1686                         if (ret)
1687                                 uptodate = 0;
1688                 }
1689                 if (!uptodate && tree->ops &&
1690                     tree->ops->readpage_io_failed_hook) {
1691                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1692                                                          start, end, NULL);
1693                         if (ret == 0) {
1694                                 uptodate =
1695                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1696                                 if (err)
1697                                         uptodate = 0;
1698                                 continue;
1699                         }
1700                 }
1701
1702                 if (uptodate) {
1703                         set_extent_uptodate(tree, start, end,
1704                                             GFP_ATOMIC);
1705                 }
1706                 unlock_extent(tree, start, end, GFP_ATOMIC);
1707
1708                 if (whole_page) {
1709                         if (uptodate) {
1710                                 SetPageUptodate(page);
1711                         } else {
1712                                 ClearPageUptodate(page);
1713                                 SetPageError(page);
1714                         }
1715                         unlock_page(page);
1716                 } else {
1717                         if (uptodate) {
1718                                 check_page_uptodate(tree, page);
1719                         } else {
1720                                 ClearPageUptodate(page);
1721                                 SetPageError(page);
1722                         }
1723                         check_page_locked(tree, page);
1724                 }
1725         } while (bvec >= bio->bi_io_vec);
1726
1727         bio_put(bio);
1728 }
1729
1730 /*
1731  * IO done from prepare_write is pretty simple, we just unlock
1732  * the structs in the extent tree when done, and set the uptodate bits
1733  * as appropriate.
1734  */
1735 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1736 {
1737         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1738         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1739         struct extent_io_tree *tree;
1740         u64 start;
1741         u64 end;
1742
1743         do {
1744                 struct page *page = bvec->bv_page;
1745                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1746
1747                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1748                         bvec->bv_offset;
1749                 end = start + bvec->bv_len - 1;
1750
1751                 if (--bvec >= bio->bi_io_vec)
1752                         prefetchw(&bvec->bv_page->flags);
1753
1754                 if (uptodate) {
1755                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1756                 } else {
1757                         ClearPageUptodate(page);
1758                         SetPageError(page);
1759                 }
1760
1761                 unlock_extent(tree, start, end, GFP_ATOMIC);
1762
1763         } while (bvec >= bio->bi_io_vec);
1764
1765         bio_put(bio);
1766 }
1767
1768 static struct bio *
1769 extent_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1770                  gfp_t gfp_flags)
1771 {
1772         struct bio *bio;
1773
1774         bio = bio_alloc(gfp_flags, nr_vecs);
1775
1776         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1777                 while (!bio && (nr_vecs /= 2))
1778                         bio = bio_alloc(gfp_flags, nr_vecs);
1779         }
1780
1781         if (bio) {
1782                 bio->bi_size = 0;
1783                 bio->bi_bdev = bdev;
1784                 bio->bi_sector = first_sector;
1785         }
1786         return bio;
1787 }
1788
1789 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1790                           unsigned long bio_flags)
1791 {
1792         int ret = 0;
1793         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1794         struct page *page = bvec->bv_page;
1795         struct extent_io_tree *tree = bio->bi_private;
1796         u64 start;
1797         u64 end;
1798
1799         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1800         end = start + bvec->bv_len - 1;
1801
1802         bio->bi_private = NULL;
1803
1804         bio_get(bio);
1805
1806         if (tree->ops && tree->ops->submit_bio_hook)
1807                 tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1808                                            mirror_num, bio_flags);
1809         else
1810                 submit_bio(rw, bio);
1811         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1812                 ret = -EOPNOTSUPP;
1813         bio_put(bio);
1814         return ret;
1815 }
1816
1817 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1818                               struct page *page, sector_t sector,
1819                               size_t size, unsigned long offset,
1820                               struct block_device *bdev,
1821                               struct bio **bio_ret,
1822                               unsigned long max_pages,
1823                               bio_end_io_t end_io_func,
1824                               int mirror_num,
1825                               unsigned long prev_bio_flags,
1826                               unsigned long bio_flags)
1827 {
1828         int ret = 0;
1829         struct bio *bio;
1830         int nr;
1831         int contig = 0;
1832         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1833         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1834         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1835
1836         if (bio_ret && *bio_ret) {
1837                 bio = *bio_ret;
1838                 if (old_compressed)
1839                         contig = bio->bi_sector == sector;
1840                 else
1841                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
1842                                 sector;
1843
1844                 if (prev_bio_flags != bio_flags || !contig ||
1845                     (tree->ops && tree->ops->merge_bio_hook &&
1846                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
1847                                                bio_flags)) ||
1848                     bio_add_page(bio, page, page_size, offset) < page_size) {
1849                         ret = submit_one_bio(rw, bio, mirror_num,
1850                                              prev_bio_flags);
1851                         bio = NULL;
1852                 } else {
1853                         return 0;
1854                 }
1855         }
1856         if (this_compressed)
1857                 nr = BIO_MAX_PAGES;
1858         else
1859                 nr = bio_get_nr_vecs(bdev);
1860
1861         bio = extent_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1862
1863         bio_add_page(bio, page, page_size, offset);
1864         bio->bi_end_io = end_io_func;
1865         bio->bi_private = tree;
1866
1867         if (bio_ret)
1868                 *bio_ret = bio;
1869         else
1870                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1871
1872         return ret;
1873 }
1874
1875 void set_page_extent_mapped(struct page *page)
1876 {
1877         if (!PagePrivate(page)) {
1878                 SetPagePrivate(page);
1879                 page_cache_get(page);
1880                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1881         }
1882 }
1883
1884 static void set_page_extent_head(struct page *page, unsigned long len)
1885 {
1886         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1887 }
1888
1889 /*
1890  * basic readpage implementation.  Locked extent state structs are inserted
1891  * into the tree that are removed when the IO is done (by the end_io
1892  * handlers)
1893  */
1894 static int __extent_read_full_page(struct extent_io_tree *tree,
1895                                    struct page *page,
1896                                    get_extent_t *get_extent,
1897                                    struct bio **bio, int mirror_num,
1898                                    unsigned long *bio_flags)
1899 {
1900         struct inode *inode = page->mapping->host;
1901         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1902         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1903         u64 end;
1904         u64 cur = start;
1905         u64 extent_offset;
1906         u64 last_byte = i_size_read(inode);
1907         u64 block_start;
1908         u64 cur_end;
1909         sector_t sector;
1910         struct extent_map *em;
1911         struct block_device *bdev;
1912         int ret;
1913         int nr = 0;
1914         size_t page_offset = 0;
1915         size_t iosize;
1916         size_t disk_io_size;
1917         size_t blocksize = inode->i_sb->s_blocksize;
1918         unsigned long this_bio_flag = 0;
1919
1920         set_page_extent_mapped(page);
1921
1922         end = page_end;
1923         lock_extent(tree, start, end, GFP_NOFS);
1924
1925         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
1926                 char *userpage;
1927                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
1928
1929                 if (zero_offset) {
1930                         iosize = PAGE_CACHE_SIZE - zero_offset;
1931                         userpage = kmap_atomic(page, KM_USER0);
1932                         memset(userpage + zero_offset, 0, iosize);
1933                         flush_dcache_page(page);
1934                         kunmap_atomic(userpage, KM_USER0);
1935                 }
1936         }
1937         while (cur <= end) {
1938                 if (cur >= last_byte) {
1939                         char *userpage;
1940                         iosize = PAGE_CACHE_SIZE - page_offset;
1941                         userpage = kmap_atomic(page, KM_USER0);
1942                         memset(userpage + page_offset, 0, iosize);
1943                         flush_dcache_page(page);
1944                         kunmap_atomic(userpage, KM_USER0);
1945                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1946                                             GFP_NOFS);
1947                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1948                         break;
1949                 }
1950                 em = get_extent(inode, page, page_offset, cur,
1951                                 end - cur + 1, 0);
1952                 if (IS_ERR(em) || !em) {
1953                         SetPageError(page);
1954                         unlock_extent(tree, cur, end, GFP_NOFS);
1955                         break;
1956                 }
1957                 extent_offset = cur - em->start;
1958                 BUG_ON(extent_map_end(em) <= cur);
1959                 BUG_ON(end < cur);
1960
1961                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
1962                         this_bio_flag = EXTENT_BIO_COMPRESSED;
1963
1964                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1965                 cur_end = min(extent_map_end(em) - 1, end);
1966                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1967                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
1968                         disk_io_size = em->block_len;
1969                         sector = em->block_start >> 9;
1970                 } else {
1971                         sector = (em->block_start + extent_offset) >> 9;
1972                         disk_io_size = iosize;
1973                 }
1974                 bdev = em->bdev;
1975                 block_start = em->block_start;
1976                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
1977                         block_start = EXTENT_MAP_HOLE;
1978                 free_extent_map(em);
1979                 em = NULL;
1980
1981                 /* we've found a hole, just zero and go on */
1982                 if (block_start == EXTENT_MAP_HOLE) {
1983                         char *userpage;
1984                         userpage = kmap_atomic(page, KM_USER0);
1985                         memset(userpage + page_offset, 0, iosize);
1986                         flush_dcache_page(page);
1987                         kunmap_atomic(userpage, KM_USER0);
1988
1989                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1990                                             GFP_NOFS);
1991                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1992                         cur = cur + iosize;
1993                         page_offset += iosize;
1994                         continue;
1995                 }
1996                 /* the get_extent function already copied into the page */
1997                 if (test_range_bit(tree, cur, cur_end, EXTENT_UPTODATE, 1)) {
1998                         check_page_uptodate(tree, page);
1999                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2000                         cur = cur + iosize;
2001                         page_offset += iosize;
2002                         continue;
2003                 }
2004                 /* we have an inline extent but it didn't get marked up
2005                  * to date.  Error out
2006                  */
2007                 if (block_start == EXTENT_MAP_INLINE) {
2008                         SetPageError(page);
2009                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2010                         cur = cur + iosize;
2011                         page_offset += iosize;
2012                         continue;
2013                 }
2014
2015                 ret = 0;
2016                 if (tree->ops && tree->ops->readpage_io_hook) {
2017                         ret = tree->ops->readpage_io_hook(page, cur,
2018                                                           cur + iosize - 1);
2019                 }
2020                 if (!ret) {
2021                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2022                         pnr -= page->index;
2023                         ret = submit_extent_page(READ, tree, page,
2024                                          sector, disk_io_size, page_offset,
2025                                          bdev, bio, pnr,
2026                                          end_bio_extent_readpage, mirror_num,
2027                                          *bio_flags,
2028                                          this_bio_flag);
2029                         nr++;
2030                         *bio_flags = this_bio_flag;
2031                 }
2032                 if (ret)
2033                         SetPageError(page);
2034                 cur = cur + iosize;
2035                 page_offset += iosize;
2036         }
2037         if (!nr) {
2038                 if (!PageError(page))
2039                         SetPageUptodate(page);
2040                 unlock_page(page);
2041         }
2042         return 0;
2043 }
2044
2045 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2046                             get_extent_t *get_extent)
2047 {
2048         struct bio *bio = NULL;
2049         unsigned long bio_flags = 0;
2050         int ret;
2051
2052         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2053                                       &bio_flags);
2054         if (bio)
2055                 submit_one_bio(READ, bio, 0, bio_flags);
2056         return ret;
2057 }
2058
2059 static noinline void update_nr_written(struct page *page,
2060                                       struct writeback_control *wbc,
2061                                       unsigned long nr_written)
2062 {
2063         wbc->nr_to_write -= nr_written;
2064         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2065             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2066                 page->mapping->writeback_index = page->index + nr_written;
2067 }
2068
2069 /*
2070  * the writepage semantics are similar to regular writepage.  extent
2071  * records are inserted to lock ranges in the tree, and as dirty areas
2072  * are found, they are marked writeback.  Then the lock bits are removed
2073  * and the end_io handler clears the writeback ranges
2074  */
2075 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2076                               void *data)
2077 {
2078         struct inode *inode = page->mapping->host;
2079         struct extent_page_data *epd = data;
2080         struct extent_io_tree *tree = epd->tree;
2081         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2082         u64 delalloc_start;
2083         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2084         u64 end;
2085         u64 cur = start;
2086         u64 extent_offset;
2087         u64 last_byte = i_size_read(inode);
2088         u64 block_start;
2089         u64 iosize;
2090         u64 unlock_start;
2091         sector_t sector;
2092         struct extent_map *em;
2093         struct block_device *bdev;
2094         int ret;
2095         int nr = 0;
2096         size_t pg_offset = 0;
2097         size_t blocksize;
2098         loff_t i_size = i_size_read(inode);
2099         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2100         u64 nr_delalloc;
2101         u64 delalloc_end;
2102         int page_started;
2103         int compressed;
2104         int write_flags;
2105         unsigned long nr_written = 0;
2106
2107         if (wbc->sync_mode == WB_SYNC_ALL)
2108                 write_flags = WRITE_SYNC_PLUG;
2109         else
2110                 write_flags = WRITE;
2111
2112         WARN_ON(!PageLocked(page));
2113         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2114         if (page->index > end_index ||
2115            (page->index == end_index && !pg_offset)) {
2116                 page->mapping->a_ops->invalidatepage(page, 0);
2117                 unlock_page(page);
2118                 return 0;
2119         }
2120
2121         if (page->index == end_index) {
2122                 char *userpage;
2123
2124                 userpage = kmap_atomic(page, KM_USER0);
2125                 memset(userpage + pg_offset, 0,
2126                        PAGE_CACHE_SIZE - pg_offset);
2127                 kunmap_atomic(userpage, KM_USER0);
2128                 flush_dcache_page(page);
2129         }
2130         pg_offset = 0;
2131
2132         set_page_extent_mapped(page);
2133
2134         delalloc_start = start;
2135         delalloc_end = 0;
2136         page_started = 0;
2137         if (!epd->extent_locked) {
2138                 u64 delalloc_to_write;
2139                 /*
2140                  * make sure the wbc mapping index is at least updated
2141                  * to this page.
2142                  */
2143                 update_nr_written(page, wbc, 0);
2144
2145                 while (delalloc_end < page_end) {
2146                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2147                                                        page,
2148                                                        &delalloc_start,
2149                                                        &delalloc_end,
2150                                                        128 * 1024 * 1024);
2151                         if (nr_delalloc == 0) {
2152                                 delalloc_start = delalloc_end + 1;
2153                                 continue;
2154                         }
2155                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2156                                                  delalloc_end, &page_started,
2157                                                  &nr_written);
2158                         delalloc_to_write = (delalloc_end -
2159                                         max_t(u64, page_offset(page),
2160                                               delalloc_start) + 1) >>
2161                                         PAGE_CACHE_SHIFT;
2162                         if (wbc->nr_to_write < delalloc_to_write) {
2163                                 wbc->nr_to_write = min_t(long, 8192,
2164                                                  delalloc_to_write);
2165                         }
2166                         delalloc_start = delalloc_end + 1;
2167                 }
2168
2169                 /* did the fill delalloc function already unlock and start
2170                  * the IO?
2171                  */
2172                 if (page_started) {
2173                         ret = 0;
2174                         /*
2175                          * we've unlocked the page, so we can't update
2176                          * the mapping's writeback index, just update
2177                          * nr_to_write.
2178                          */
2179                         wbc->nr_to_write -= nr_written;
2180                         goto done_unlocked;
2181                 }
2182         }
2183         lock_extent(tree, start, page_end, GFP_NOFS);
2184
2185         unlock_start = start;
2186
2187         if (tree->ops && tree->ops->writepage_start_hook) {
2188                 ret = tree->ops->writepage_start_hook(page, start,
2189                                                       page_end);
2190                 if (ret == -EAGAIN) {
2191                         unlock_extent(tree, start, page_end, GFP_NOFS);
2192                         redirty_page_for_writepage(wbc, page);
2193                         update_nr_written(page, wbc, nr_written);
2194                         unlock_page(page);
2195                         ret = 0;
2196                         goto done_unlocked;
2197                 }
2198         }
2199
2200         /*
2201          * we don't want to touch the inode after unlocking the page,
2202          * so we update the mapping writeback index now
2203          */
2204         update_nr_written(page, wbc, nr_written + 1);
2205
2206         end = page_end;
2207         if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0))
2208                 printk(KERN_ERR "btrfs delalloc bits after lock_extent\n");
2209
2210         if (last_byte <= start) {
2211                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
2212                 unlock_extent(tree, start, page_end, GFP_NOFS);
2213                 if (tree->ops && tree->ops->writepage_end_io_hook)
2214                         tree->ops->writepage_end_io_hook(page, start,
2215                                                          page_end, NULL, 1);
2216                 unlock_start = page_end + 1;
2217                 goto done;
2218         }
2219
2220         set_extent_uptodate(tree, start, page_end, GFP_NOFS);
2221         blocksize = inode->i_sb->s_blocksize;
2222
2223         while (cur <= end) {
2224                 if (cur >= last_byte) {
2225                         clear_extent_dirty(tree, cur, page_end, GFP_NOFS);
2226                         unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
2227                         if (tree->ops && tree->ops->writepage_end_io_hook)
2228                                 tree->ops->writepage_end_io_hook(page, cur,
2229                                                          page_end, NULL, 1);
2230                         unlock_start = page_end + 1;
2231                         break;
2232                 }
2233                 em = epd->get_extent(inode, page, pg_offset, cur,
2234                                      end - cur + 1, 1);
2235                 if (IS_ERR(em) || !em) {
2236                         SetPageError(page);
2237                         break;
2238                 }
2239
2240                 extent_offset = cur - em->start;
2241                 BUG_ON(extent_map_end(em) <= cur);
2242                 BUG_ON(end < cur);
2243                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2244                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2245                 sector = (em->block_start + extent_offset) >> 9;
2246                 bdev = em->bdev;
2247                 block_start = em->block_start;
2248                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2249                 free_extent_map(em);
2250                 em = NULL;
2251
2252                 /*
2253                  * compressed and inline extents are written through other
2254                  * paths in the FS
2255                  */
2256                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2257                     block_start == EXTENT_MAP_INLINE) {
2258                         clear_extent_dirty(tree, cur,
2259                                            cur + iosize - 1, GFP_NOFS);
2260
2261                         unlock_extent(tree, unlock_start, cur + iosize - 1,
2262                                       GFP_NOFS);
2263
2264                         /*
2265                          * end_io notification does not happen here for
2266                          * compressed extents
2267                          */
2268                         if (!compressed && tree->ops &&
2269                             tree->ops->writepage_end_io_hook)
2270                                 tree->ops->writepage_end_io_hook(page, cur,
2271                                                          cur + iosize - 1,
2272                                                          NULL, 1);
2273                         else if (compressed) {
2274                                 /* we don't want to end_page_writeback on
2275                                  * a compressed extent.  this happens
2276                                  * elsewhere
2277                                  */
2278                                 nr++;
2279                         }
2280
2281                         cur += iosize;
2282                         pg_offset += iosize;
2283                         unlock_start = cur;
2284                         continue;
2285                 }
2286                 /* leave this out until we have a page_mkwrite call */
2287                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2288                                    EXTENT_DIRTY, 0)) {
2289                         cur = cur + iosize;
2290                         pg_offset += iosize;
2291                         continue;
2292                 }
2293
2294                 clear_extent_dirty(tree, cur, cur + iosize - 1, GFP_NOFS);
2295                 if (tree->ops && tree->ops->writepage_io_hook) {
2296                         ret = tree->ops->writepage_io_hook(page, cur,
2297                                                 cur + iosize - 1);
2298                 } else {
2299                         ret = 0;
2300                 }
2301                 if (ret) {
2302                         SetPageError(page);
2303                 } else {
2304                         unsigned long max_nr = end_index + 1;
2305
2306                         set_range_writeback(tree, cur, cur + iosize - 1);
2307                         if (!PageWriteback(page)) {
2308                                 printk(KERN_ERR "btrfs warning page %lu not "
2309                                        "writeback, cur %llu end %llu\n",
2310                                        page->index, (unsigned long long)cur,
2311                                        (unsigned long long)end);
2312                         }
2313
2314                         ret = submit_extent_page(write_flags, tree, page,
2315                                                  sector, iosize, pg_offset,
2316                                                  bdev, &epd->bio, max_nr,
2317                                                  end_bio_extent_writepage,
2318                                                  0, 0, 0);
2319                         if (ret)
2320                                 SetPageError(page);
2321                 }
2322                 cur = cur + iosize;
2323                 pg_offset += iosize;
2324                 nr++;
2325         }
2326 done:
2327         if (nr == 0) {
2328                 /* make sure the mapping tag for page dirty gets cleared */
2329                 set_page_writeback(page);
2330                 end_page_writeback(page);
2331         }
2332         if (unlock_start <= page_end)
2333                 unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
2334         unlock_page(page);
2335
2336 done_unlocked:
2337
2338         return 0;
2339 }
2340
2341 /**
2342  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2343  * @mapping: address space structure to write
2344  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2345  * @writepage: function called for each page
2346  * @data: data passed to writepage function
2347  *
2348  * If a page is already under I/O, write_cache_pages() skips it, even
2349  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2350  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2351  * and msync() need to guarantee that all the data which was dirty at the time
2352  * the call was made get new I/O started against them.  If wbc->sync_mode is
2353  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2354  * existing IO to complete.
2355  */
2356 static int extent_write_cache_pages(struct extent_io_tree *tree,
2357                              struct address_space *mapping,
2358                              struct writeback_control *wbc,
2359                              writepage_t writepage, void *data,
2360                              void (*flush_fn)(void *))
2361 {
2362         int ret = 0;
2363         int done = 0;
2364         struct pagevec pvec;
2365         int nr_pages;
2366         pgoff_t index;
2367         pgoff_t end;            /* Inclusive */
2368         int scanned = 0;
2369         int range_whole = 0;
2370
2371         pagevec_init(&pvec, 0);
2372         if (wbc->range_cyclic) {
2373                 index = mapping->writeback_index; /* Start from prev offset */
2374                 end = -1;
2375         } else {
2376                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2377                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2378                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2379                         range_whole = 1;
2380                 scanned = 1;
2381         }
2382 retry:
2383         while (!done && (index <= end) &&
2384                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2385                               PAGECACHE_TAG_DIRTY, min(end - index,
2386                                   (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2387                 unsigned i;
2388
2389                 scanned = 1;
2390                 for (i = 0; i < nr_pages; i++) {
2391                         struct page *page = pvec.pages[i];
2392
2393                         /*
2394                          * At this point we hold neither mapping->tree_lock nor
2395                          * lock on the page itself: the page may be truncated or
2396                          * invalidated (changing page->mapping to NULL), or even
2397                          * swizzled back from swapper_space to tmpfs file
2398                          * mapping
2399                          */
2400                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2401                                 tree->ops->write_cache_pages_lock_hook(page);
2402                         else
2403                                 lock_page(page);
2404
2405                         if (unlikely(page->mapping != mapping)) {
2406                                 unlock_page(page);
2407                                 continue;
2408                         }
2409
2410                         if (!wbc->range_cyclic && page->index > end) {
2411                                 done = 1;
2412                                 unlock_page(page);
2413                                 continue;
2414                         }
2415
2416                         if (wbc->sync_mode != WB_SYNC_NONE) {
2417                                 if (PageWriteback(page))
2418                                         flush_fn(data);
2419                                 wait_on_page_writeback(page);
2420                         }
2421
2422                         if (PageWriteback(page) ||
2423                             !clear_page_dirty_for_io(page)) {
2424                                 unlock_page(page);
2425                                 continue;
2426                         }
2427
2428                         ret = (*writepage)(page, wbc, data);
2429
2430                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2431                                 unlock_page(page);
2432                                 ret = 0;
2433                         }
2434                         if (ret || wbc->nr_to_write <= 0)
2435                                 done = 1;
2436                 }
2437                 pagevec_release(&pvec);
2438                 cond_resched();
2439         }
2440         if (!scanned && !done) {
2441                 /*
2442                  * We hit the last page and there is more work to be done: wrap
2443                  * back to the start of the file
2444                  */
2445                 scanned = 1;
2446                 index = 0;
2447                 goto retry;
2448         }
2449         return ret;
2450 }
2451
2452 static void flush_epd_write_bio(struct extent_page_data *epd)
2453 {
2454         if (epd->bio) {
2455                 if (epd->sync_io)
2456                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2457                 else
2458                         submit_one_bio(WRITE, epd->bio, 0, 0);
2459                 epd->bio = NULL;
2460         }
2461 }
2462
2463 static noinline void flush_write_bio(void *data)
2464 {
2465         struct extent_page_data *epd = data;
2466         flush_epd_write_bio(epd);
2467 }
2468
2469 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2470                           get_extent_t *get_extent,
2471                           struct writeback_control *wbc)
2472 {
2473         int ret;
2474         struct address_space *mapping = page->mapping;
2475         struct extent_page_data epd = {
2476                 .bio = NULL,
2477                 .tree = tree,
2478                 .get_extent = get_extent,
2479                 .extent_locked = 0,
2480                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2481         };
2482         struct writeback_control wbc_writepages = {
2483                 .bdi            = wbc->bdi,
2484                 .sync_mode      = wbc->sync_mode,
2485                 .older_than_this = NULL,
2486                 .nr_to_write    = 64,
2487                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2488                 .range_end      = (loff_t)-1,
2489         };
2490
2491         ret = __extent_writepage(page, wbc, &epd);
2492
2493         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2494                                  __extent_writepage, &epd, flush_write_bio);
2495         flush_epd_write_bio(&epd);
2496         return ret;
2497 }
2498
2499 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2500                               u64 start, u64 end, get_extent_t *get_extent,
2501                               int mode)
2502 {
2503         int ret = 0;
2504         struct address_space *mapping = inode->i_mapping;
2505         struct page *page;
2506         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2507                 PAGE_CACHE_SHIFT;
2508
2509         struct extent_page_data epd = {
2510                 .bio = NULL,
2511                 .tree = tree,
2512                 .get_extent = get_extent,
2513                 .extent_locked = 1,
2514                 .sync_io = mode == WB_SYNC_ALL,
2515         };
2516         struct writeback_control wbc_writepages = {
2517                 .bdi            = inode->i_mapping->backing_dev_info,
2518                 .sync_mode      = mode,
2519                 .older_than_this = NULL,
2520                 .nr_to_write    = nr_pages * 2,
2521                 .range_start    = start,
2522                 .range_end      = end + 1,
2523         };
2524
2525         while (start <= end) {
2526                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2527                 if (clear_page_dirty_for_io(page))
2528                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2529                 else {
2530                         if (tree->ops && tree->ops->writepage_end_io_hook)
2531                                 tree->ops->writepage_end_io_hook(page, start,
2532                                                  start + PAGE_CACHE_SIZE - 1,
2533                                                  NULL, 1);
2534                         unlock_page(page);
2535                 }
2536                 page_cache_release(page);
2537                 start += PAGE_CACHE_SIZE;
2538         }
2539
2540         flush_epd_write_bio(&epd);
2541         return ret;
2542 }
2543
2544 int extent_writepages(struct extent_io_tree *tree,
2545                       struct address_space *mapping,
2546                       get_extent_t *get_extent,
2547                       struct writeback_control *wbc)
2548 {
2549         int ret = 0;
2550         struct extent_page_data epd = {
2551                 .bio = NULL,
2552                 .tree = tree,
2553                 .get_extent = get_extent,
2554                 .extent_locked = 0,
2555                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2556         };
2557
2558         ret = extent_write_cache_pages(tree, mapping, wbc,
2559                                        __extent_writepage, &epd,
2560                                        flush_write_bio);
2561         flush_epd_write_bio(&epd);
2562         return ret;
2563 }
2564
2565 int extent_readpages(struct extent_io_tree *tree,
2566                      struct address_space *mapping,
2567                      struct list_head *pages, unsigned nr_pages,
2568                      get_extent_t get_extent)
2569 {
2570         struct bio *bio = NULL;
2571         unsigned page_idx;
2572         struct pagevec pvec;
2573         unsigned long bio_flags = 0;
2574
2575         pagevec_init(&pvec, 0);
2576         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2577                 struct page *page = list_entry(pages->prev, struct page, lru);
2578
2579                 prefetchw(&page->flags);
2580                 list_del(&page->lru);
2581                 /*
2582                  * what we want to do here is call add_to_page_cache_lru,
2583                  * but that isn't exported, so we reproduce it here
2584                  */
2585                 if (!add_to_page_cache(page, mapping,
2586                                         page->index, GFP_KERNEL)) {
2587
2588                         /* open coding of lru_cache_add, also not exported */
2589                         page_cache_get(page);
2590                         if (!pagevec_add(&pvec, page))
2591                                 __pagevec_lru_add_file(&pvec);
2592                         __extent_read_full_page(tree, page, get_extent,
2593                                                 &bio, 0, &bio_flags);
2594                 }
2595                 page_cache_release(page);
2596         }
2597         if (pagevec_count(&pvec))
2598                 __pagevec_lru_add_file(&pvec);
2599         BUG_ON(!list_empty(pages));
2600         if (bio)
2601                 submit_one_bio(READ, bio, 0, bio_flags);
2602         return 0;
2603 }
2604
2605 /*
2606  * basic invalidatepage code, this waits on any locked or writeback
2607  * ranges corresponding to the page, and then deletes any extent state
2608  * records from the tree
2609  */
2610 int extent_invalidatepage(struct extent_io_tree *tree,
2611                           struct page *page, unsigned long offset)
2612 {
2613         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2614         u64 end = start + PAGE_CACHE_SIZE - 1;
2615         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2616
2617         start += (offset + blocksize - 1) & ~(blocksize - 1);
2618         if (start > end)
2619                 return 0;
2620
2621         lock_extent(tree, start, end, GFP_NOFS);
2622         wait_on_extent_writeback(tree, start, end);
2623         clear_extent_bit(tree, start, end,
2624                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC,
2625                          1, 1, GFP_NOFS);
2626         return 0;
2627 }
2628
2629 /*
2630  * simple commit_write call, set_range_dirty is used to mark both
2631  * the pages and the extent records as dirty
2632  */
2633 int extent_commit_write(struct extent_io_tree *tree,
2634                         struct inode *inode, struct page *page,
2635                         unsigned from, unsigned to)
2636 {
2637         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2638
2639         set_page_extent_mapped(page);
2640         set_page_dirty(page);
2641
2642         if (pos > inode->i_size) {
2643                 i_size_write(inode, pos);
2644                 mark_inode_dirty(inode);
2645         }
2646         return 0;
2647 }
2648
2649 int extent_prepare_write(struct extent_io_tree *tree,
2650                          struct inode *inode, struct page *page,
2651                          unsigned from, unsigned to, get_extent_t *get_extent)
2652 {
2653         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2654         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2655         u64 block_start;
2656         u64 orig_block_start;
2657         u64 block_end;
2658         u64 cur_end;
2659         struct extent_map *em;
2660         unsigned blocksize = 1 << inode->i_blkbits;
2661         size_t page_offset = 0;
2662         size_t block_off_start;
2663         size_t block_off_end;
2664         int err = 0;
2665         int iocount = 0;
2666         int ret = 0;
2667         int isnew;
2668
2669         set_page_extent_mapped(page);
2670
2671         block_start = (page_start + from) & ~((u64)blocksize - 1);
2672         block_end = (page_start + to - 1) | (blocksize - 1);
2673         orig_block_start = block_start;
2674
2675         lock_extent(tree, page_start, page_end, GFP_NOFS);
2676         while (block_start <= block_end) {
2677                 em = get_extent(inode, page, page_offset, block_start,
2678                                 block_end - block_start + 1, 1);
2679                 if (IS_ERR(em) || !em)
2680                         goto err;
2681
2682                 cur_end = min(block_end, extent_map_end(em) - 1);
2683                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2684                 block_off_end = block_off_start + blocksize;
2685                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2686
2687                 if (!PageUptodate(page) && isnew &&
2688                     (block_off_end > to || block_off_start < from)) {
2689                         void *kaddr;
2690
2691                         kaddr = kmap_atomic(page, KM_USER0);
2692                         if (block_off_end > to)
2693                                 memset(kaddr + to, 0, block_off_end - to);
2694                         if (block_off_start < from)
2695                                 memset(kaddr + block_off_start, 0,
2696                                        from - block_off_start);
2697                         flush_dcache_page(page);
2698                         kunmap_atomic(kaddr, KM_USER0);
2699                 }
2700                 if ((em->block_start != EXTENT_MAP_HOLE &&
2701                      em->block_start != EXTENT_MAP_INLINE) &&
2702                     !isnew && !PageUptodate(page) &&
2703                     (block_off_end > to || block_off_start < from) &&
2704                     !test_range_bit(tree, block_start, cur_end,
2705                                     EXTENT_UPTODATE, 1)) {
2706                         u64 sector;
2707                         u64 extent_offset = block_start - em->start;
2708                         size_t iosize;
2709                         sector = (em->block_start + extent_offset) >> 9;
2710                         iosize = (cur_end - block_start + blocksize) &
2711                                 ~((u64)blocksize - 1);
2712                         /*
2713                          * we've already got the extent locked, but we
2714                          * need to split the state such that our end_bio
2715                          * handler can clear the lock.
2716                          */
2717                         set_extent_bit(tree, block_start,
2718                                        block_start + iosize - 1,
2719                                        EXTENT_LOCKED, 0, NULL, GFP_NOFS);
2720                         ret = submit_extent_page(READ, tree, page,
2721                                          sector, iosize, page_offset, em->bdev,
2722                                          NULL, 1,
2723                                          end_bio_extent_preparewrite, 0,
2724                                          0, 0);
2725                         iocount++;
2726                         block_start = block_start + iosize;
2727                 } else {
2728                         set_extent_uptodate(tree, block_start, cur_end,
2729                                             GFP_NOFS);
2730                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2731                         block_start = cur_end + 1;
2732                 }
2733                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2734                 free_extent_map(em);
2735         }
2736         if (iocount) {
2737                 wait_extent_bit(tree, orig_block_start,
2738                                 block_end, EXTENT_LOCKED);
2739         }
2740         check_page_uptodate(tree, page);
2741 err:
2742         /* FIXME, zero out newly allocated blocks on error */
2743         return err;
2744 }
2745
2746 /*
2747  * a helper for releasepage, this tests for areas of the page that
2748  * are locked or under IO and drops the related state bits if it is safe
2749  * to drop the page.
2750  */
2751 int try_release_extent_state(struct extent_map_tree *map,
2752                              struct extent_io_tree *tree, struct page *page,
2753                              gfp_t mask)
2754 {
2755         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2756         u64 end = start + PAGE_CACHE_SIZE - 1;
2757         int ret = 1;
2758
2759         if (test_range_bit(tree, start, end,
2760                            EXTENT_IOBITS | EXTENT_ORDERED, 0))
2761                 ret = 0;
2762         else {
2763                 if ((mask & GFP_NOFS) == GFP_NOFS)
2764                         mask = GFP_NOFS;
2765                 clear_extent_bit(tree, start, end, EXTENT_UPTODATE,
2766                                  1, 1, mask);
2767         }
2768         return ret;
2769 }
2770
2771 /*
2772  * a helper for releasepage.  As long as there are no locked extents
2773  * in the range corresponding to the page, both state records and extent
2774  * map records are removed
2775  */
2776 int try_release_extent_mapping(struct extent_map_tree *map,
2777                                struct extent_io_tree *tree, struct page *page,
2778                                gfp_t mask)
2779 {
2780         struct extent_map *em;
2781         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2782         u64 end = start + PAGE_CACHE_SIZE - 1;
2783
2784         if ((mask & __GFP_WAIT) &&
2785             page->mapping->host->i_size > 16 * 1024 * 1024) {
2786                 u64 len;
2787                 while (start <= end) {
2788                         len = end - start + 1;
2789                         spin_lock(&map->lock);
2790                         em = lookup_extent_mapping(map, start, len);
2791                         if (!em || IS_ERR(em)) {
2792                                 spin_unlock(&map->lock);
2793                                 break;
2794                         }
2795                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2796                             em->start != start) {
2797                                 spin_unlock(&map->lock);
2798                                 free_extent_map(em);
2799                                 break;
2800                         }
2801                         if (!test_range_bit(tree, em->start,
2802                                             extent_map_end(em) - 1,
2803                                             EXTENT_LOCKED | EXTENT_WRITEBACK |
2804                                             EXTENT_ORDERED,
2805                                             0)) {
2806                                 remove_extent_mapping(map, em);
2807                                 /* once for the rb tree */
2808                                 free_extent_map(em);
2809                         }
2810                         start = extent_map_end(em);
2811                         spin_unlock(&map->lock);
2812
2813                         /* once for us */
2814                         free_extent_map(em);
2815                 }
2816         }
2817         return try_release_extent_state(map, tree, page, mask);
2818 }
2819
2820 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2821                 get_extent_t *get_extent)
2822 {
2823         struct inode *inode = mapping->host;
2824         u64 start = iblock << inode->i_blkbits;
2825         sector_t sector = 0;
2826         size_t blksize = (1 << inode->i_blkbits);
2827         struct extent_map *em;
2828
2829         lock_extent(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2830                     GFP_NOFS);
2831         em = get_extent(inode, NULL, 0, start, blksize, 0);
2832         unlock_extent(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2833                       GFP_NOFS);
2834         if (!em || IS_ERR(em))
2835                 return 0;
2836
2837         if (em->block_start > EXTENT_MAP_LAST_BYTE)
2838                 goto out;
2839
2840         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2841 out:
2842         free_extent_map(em);
2843         return sector;
2844 }
2845
2846 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2847                 __u64 start, __u64 len, get_extent_t *get_extent)
2848 {
2849         int ret;
2850         u64 off = start;
2851         u64 max = start + len;
2852         u32 flags = 0;
2853         u64 disko = 0;
2854         struct extent_map *em = NULL;
2855         int end = 0;
2856         u64 em_start = 0, em_len = 0;
2857         unsigned long emflags;
2858         ret = 0;
2859
2860         if (len == 0)
2861                 return -EINVAL;
2862
2863         lock_extent(&BTRFS_I(inode)->io_tree, start, start + len,
2864                 GFP_NOFS);
2865         em = get_extent(inode, NULL, 0, off, max - off, 0);
2866         if (!em)
2867                 goto out;
2868         if (IS_ERR(em)) {
2869                 ret = PTR_ERR(em);
2870                 goto out;
2871         }
2872         while (!end) {
2873                 off = em->start + em->len;
2874                 if (off >= max)
2875                         end = 1;
2876
2877                 em_start = em->start;
2878                 em_len = em->len;
2879
2880                 disko = 0;
2881                 flags = 0;
2882
2883                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
2884                         end = 1;
2885                         flags |= FIEMAP_EXTENT_LAST;
2886                 } else if (em->block_start == EXTENT_MAP_HOLE) {
2887                         flags |= FIEMAP_EXTENT_UNWRITTEN;
2888                 } else if (em->block_start == EXTENT_MAP_INLINE) {
2889                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
2890                                   FIEMAP_EXTENT_NOT_ALIGNED);
2891                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
2892                         flags |= (FIEMAP_EXTENT_DELALLOC |
2893                                   FIEMAP_EXTENT_UNKNOWN);
2894                 } else {
2895                         disko = em->block_start;
2896                 }
2897                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2898                         flags |= FIEMAP_EXTENT_ENCODED;
2899
2900                 emflags = em->flags;
2901                 free_extent_map(em);
2902                 em = NULL;
2903
2904                 if (!end) {
2905                         em = get_extent(inode, NULL, 0, off, max - off, 0);
2906                         if (!em)
2907                                 goto out;
2908                         if (IS_ERR(em)) {
2909                                 ret = PTR_ERR(em);
2910                                 goto out;
2911                         }
2912                         emflags = em->flags;
2913                 }
2914                 if (test_bit(EXTENT_FLAG_VACANCY, &emflags)) {
2915                         flags |= FIEMAP_EXTENT_LAST;
2916                         end = 1;
2917                 }
2918
2919                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
2920                                         em_len, flags);
2921                 if (ret)
2922                         goto out_free;
2923         }
2924 out_free:
2925         free_extent_map(em);
2926 out:
2927         unlock_extent(&BTRFS_I(inode)->io_tree, start, start + len,
2928                         GFP_NOFS);
2929         return ret;
2930 }
2931
2932 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2933                                               unsigned long i)
2934 {
2935         struct page *p;
2936         struct address_space *mapping;
2937
2938         if (i == 0)
2939                 return eb->first_page;
2940         i += eb->start >> PAGE_CACHE_SHIFT;
2941         mapping = eb->first_page->mapping;
2942         if (!mapping)
2943                 return NULL;
2944
2945         /*
2946          * extent_buffer_page is only called after pinning the page
2947          * by increasing the reference count.  So we know the page must
2948          * be in the radix tree.
2949          */
2950         rcu_read_lock();
2951         p = radix_tree_lookup(&mapping->page_tree, i);
2952         rcu_read_unlock();
2953
2954         return p;
2955 }
2956
2957 static inline unsigned long num_extent_pages(u64 start, u64 len)
2958 {
2959         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
2960                 (start >> PAGE_CACHE_SHIFT);
2961 }
2962
2963 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
2964                                                    u64 start,
2965                                                    unsigned long len,
2966                                                    gfp_t mask)
2967 {
2968         struct extent_buffer *eb = NULL;
2969 #if LEAK_DEBUG
2970         unsigned long flags;
2971 #endif
2972
2973         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
2974         eb->start = start;
2975         eb->len = len;
2976         spin_lock_init(&eb->lock);
2977         init_waitqueue_head(&eb->lock_wq);
2978
2979 #if LEAK_DEBUG
2980         spin_lock_irqsave(&leak_lock, flags);
2981         list_add(&eb->leak_list, &buffers);
2982         spin_unlock_irqrestore(&leak_lock, flags);
2983 #endif
2984         atomic_set(&eb->refs, 1);
2985
2986         return eb;
2987 }
2988
2989 static void __free_extent_buffer(struct extent_buffer *eb)
2990 {
2991 #if LEAK_DEBUG
2992         unsigned long flags;
2993         spin_lock_irqsave(&leak_lock, flags);
2994         list_del(&eb->leak_list);
2995         spin_unlock_irqrestore(&leak_lock, flags);
2996 #endif
2997         kmem_cache_free(extent_buffer_cache, eb);
2998 }
2999
3000 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3001                                           u64 start, unsigned long len,
3002                                           struct page *page0,
3003                                           gfp_t mask)
3004 {
3005         unsigned long num_pages = num_extent_pages(start, len);
3006         unsigned long i;
3007         unsigned long index = start >> PAGE_CACHE_SHIFT;
3008         struct extent_buffer *eb;
3009         struct extent_buffer *exists = NULL;
3010         struct page *p;
3011         struct address_space *mapping = tree->mapping;
3012         int uptodate = 1;
3013
3014         spin_lock(&tree->buffer_lock);
3015         eb = buffer_search(tree, start);
3016         if (eb) {
3017                 atomic_inc(&eb->refs);
3018                 spin_unlock(&tree->buffer_lock);
3019                 mark_page_accessed(eb->first_page);
3020                 return eb;
3021         }
3022         spin_unlock(&tree->buffer_lock);
3023
3024         eb = __alloc_extent_buffer(tree, start, len, mask);
3025         if (!eb)
3026                 return NULL;
3027
3028         if (page0) {
3029                 eb->first_page = page0;
3030                 i = 1;
3031                 index++;
3032                 page_cache_get(page0);
3033                 mark_page_accessed(page0);
3034                 set_page_extent_mapped(page0);
3035                 set_page_extent_head(page0, len);
3036                 uptodate = PageUptodate(page0);
3037         } else {
3038                 i = 0;
3039         }
3040         for (; i < num_pages; i++, index++) {
3041                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
3042                 if (!p) {
3043                         WARN_ON(1);
3044                         goto free_eb;
3045                 }
3046                 set_page_extent_mapped(p);
3047                 mark_page_accessed(p);
3048                 if (i == 0) {
3049                         eb->first_page = p;
3050                         set_page_extent_head(p, len);
3051                 } else {
3052                         set_page_private(p, EXTENT_PAGE_PRIVATE);
3053                 }
3054                 if (!PageUptodate(p))
3055                         uptodate = 0;
3056                 unlock_page(p);
3057         }
3058         if (uptodate)
3059                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3060
3061         spin_lock(&tree->buffer_lock);
3062         exists = buffer_tree_insert(tree, start, &eb->rb_node);
3063         if (exists) {
3064                 /* add one reference for the caller */
3065                 atomic_inc(&exists->refs);
3066                 spin_unlock(&tree->buffer_lock);
3067                 goto free_eb;
3068         }
3069         spin_unlock(&tree->buffer_lock);
3070
3071         /* add one reference for the tree */
3072         atomic_inc(&eb->refs);
3073         return eb;
3074
3075 free_eb:
3076         if (!atomic_dec_and_test(&eb->refs))
3077                 return exists;
3078         for (index = 1; index < i; index++)
3079                 page_cache_release(extent_buffer_page(eb, index));
3080         page_cache_release(extent_buffer_page(eb, 0));
3081         __free_extent_buffer(eb);
3082         return exists;
3083 }
3084
3085 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3086                                          u64 start, unsigned long len,
3087                                           gfp_t mask)
3088 {
3089         struct extent_buffer *eb;
3090
3091         spin_lock(&tree->buffer_lock);
3092         eb = buffer_search(tree, start);
3093         if (eb)
3094                 atomic_inc(&eb->refs);
3095         spin_unlock(&tree->buffer_lock);
3096
3097         if (eb)
3098                 mark_page_accessed(eb->first_page);
3099
3100         return eb;
3101 }
3102
3103 void free_extent_buffer(struct extent_buffer *eb)
3104 {
3105         if (!eb)
3106                 return;
3107
3108         if (!atomic_dec_and_test(&eb->refs))
3109                 return;
3110
3111         WARN_ON(1);
3112 }
3113
3114 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3115                               struct extent_buffer *eb)
3116 {
3117         unsigned long i;
3118         unsigned long num_pages;
3119         struct page *page;
3120
3121         num_pages = num_extent_pages(eb->start, eb->len);
3122
3123         for (i = 0; i < num_pages; i++) {
3124                 page = extent_buffer_page(eb, i);
3125                 if (!PageDirty(page))
3126                         continue;
3127
3128                 lock_page(page);
3129                 if (i == 0)
3130                         set_page_extent_head(page, eb->len);
3131                 else
3132                         set_page_private(page, EXTENT_PAGE_PRIVATE);
3133
3134                 clear_page_dirty_for_io(page);
3135                 spin_lock_irq(&page->mapping->tree_lock);
3136                 if (!PageDirty(page)) {
3137                         radix_tree_tag_clear(&page->mapping->page_tree,
3138                                                 page_index(page),
3139                                                 PAGECACHE_TAG_DIRTY);
3140                 }
3141                 spin_unlock_irq(&page->mapping->tree_lock);
3142                 unlock_page(page);
3143         }
3144         return 0;
3145 }
3146
3147 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
3148                                     struct extent_buffer *eb)
3149 {
3150         return wait_on_extent_writeback(tree, eb->start,
3151                                         eb->start + eb->len - 1);
3152 }
3153
3154 int set_extent_buffer_dirty(struct extent_io_tree *tree,
3155                              struct extent_buffer *eb)
3156 {
3157         unsigned long i;
3158         unsigned long num_pages;
3159         int was_dirty = 0;
3160
3161         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3162         num_pages = num_extent_pages(eb->start, eb->len);
3163         for (i = 0; i < num_pages; i++)
3164                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3165         return was_dirty;
3166 }
3167
3168 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3169                                 struct extent_buffer *eb)
3170 {
3171         unsigned long i;
3172         struct page *page;
3173         unsigned long num_pages;
3174
3175         num_pages = num_extent_pages(eb->start, eb->len);
3176         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3177
3178         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3179                               GFP_NOFS);
3180         for (i = 0; i < num_pages; i++) {
3181                 page = extent_buffer_page(eb, i);
3182                 if (page)
3183                         ClearPageUptodate(page);
3184         }
3185         return 0;
3186 }
3187
3188 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3189                                 struct extent_buffer *eb)
3190 {
3191         unsigned long i;
3192         struct page *page;
3193         unsigned long num_pages;
3194
3195         num_pages = num_extent_pages(eb->start, eb->len);
3196
3197         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3198                             GFP_NOFS);
3199         for (i = 0; i < num_pages; i++) {
3200                 page = extent_buffer_page(eb, i);
3201                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3202                     ((i == num_pages - 1) &&
3203                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3204                         check_page_uptodate(tree, page);
3205                         continue;
3206                 }
3207                 SetPageUptodate(page);
3208         }
3209         return 0;
3210 }
3211
3212 int extent_range_uptodate(struct extent_io_tree *tree,
3213                           u64 start, u64 end)
3214 {
3215         struct page *page;
3216         int ret;
3217         int pg_uptodate = 1;
3218         int uptodate;
3219         unsigned long index;
3220
3221         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1);
3222         if (ret)
3223                 return 1;
3224         while (start <= end) {
3225                 index = start >> PAGE_CACHE_SHIFT;
3226                 page = find_get_page(tree->mapping, index);
3227                 uptodate = PageUptodate(page);
3228                 page_cache_release(page);
3229                 if (!uptodate) {
3230                         pg_uptodate = 0;
3231                         break;
3232                 }
3233                 start += PAGE_CACHE_SIZE;
3234         }
3235         return pg_uptodate;
3236 }
3237
3238 int extent_buffer_uptodate(struct extent_io_tree *tree,
3239                            struct extent_buffer *eb)
3240 {
3241         int ret = 0;
3242         unsigned long num_pages;
3243         unsigned long i;
3244         struct page *page;
3245         int pg_uptodate = 1;
3246
3247         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3248                 return 1;
3249
3250         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3251                            EXTENT_UPTODATE, 1);
3252         if (ret)
3253                 return ret;
3254
3255         num_pages = num_extent_pages(eb->start, eb->len);
3256         for (i = 0; i < num_pages; i++) {
3257                 page = extent_buffer_page(eb, i);
3258                 if (!PageUptodate(page)) {
3259                         pg_uptodate = 0;
3260                         break;
3261                 }
3262         }
3263         return pg_uptodate;
3264 }
3265
3266 int read_extent_buffer_pages(struct extent_io_tree *tree,
3267                              struct extent_buffer *eb,
3268                              u64 start, int wait,
3269                              get_extent_t *get_extent, int mirror_num)
3270 {
3271         unsigned long i;
3272         unsigned long start_i;
3273         struct page *page;
3274         int err;
3275         int ret = 0;
3276         int locked_pages = 0;
3277         int all_uptodate = 1;
3278         int inc_all_pages = 0;
3279         unsigned long num_pages;
3280         struct bio *bio = NULL;
3281         unsigned long bio_flags = 0;
3282
3283         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3284                 return 0;
3285
3286         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3287                            EXTENT_UPTODATE, 1)) {
3288                 return 0;
3289         }
3290
3291         if (start) {
3292                 WARN_ON(start < eb->start);
3293                 start_i = (start >> PAGE_CACHE_SHIFT) -
3294                         (eb->start >> PAGE_CACHE_SHIFT);
3295         } else {
3296                 start_i = 0;
3297         }
3298
3299         num_pages = num_extent_pages(eb->start, eb->len);
3300         for (i = start_i; i < num_pages; i++) {
3301                 page = extent_buffer_page(eb, i);
3302                 if (!wait) {
3303                         if (!trylock_page(page))
3304                                 goto unlock_exit;
3305                 } else {
3306                         lock_page(page);
3307                 }
3308                 locked_pages++;
3309                 if (!PageUptodate(page))
3310                         all_uptodate = 0;
3311         }
3312         if (all_uptodate) {
3313                 if (start_i == 0)
3314                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3315                 goto unlock_exit;
3316         }
3317
3318         for (i = start_i; i < num_pages; i++) {
3319                 page = extent_buffer_page(eb, i);
3320                 if (inc_all_pages)
3321                         page_cache_get(page);
3322                 if (!PageUptodate(page)) {
3323                         if (start_i == 0)
3324                                 inc_all_pages = 1;
3325                         ClearPageError(page);
3326                         err = __extent_read_full_page(tree, page,
3327                                                       get_extent, &bio,
3328                                                       mirror_num, &bio_flags);
3329                         if (err)
3330                                 ret = err;
3331                 } else {
3332                         unlock_page(page);
3333                 }
3334         }
3335
3336         if (bio)
3337                 submit_one_bio(READ, bio, mirror_num, bio_flags);
3338
3339         if (ret || !wait)
3340                 return ret;
3341
3342         for (i = start_i; i < num_pages; i++) {
3343                 page = extent_buffer_page(eb, i);
3344                 wait_on_page_locked(page);
3345                 if (!PageUptodate(page))
3346                         ret = -EIO;
3347         }
3348
3349         if (!ret)
3350                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3351         return ret;
3352
3353 unlock_exit:
3354         i = start_i;
3355         while (locked_pages > 0) {
3356                 page = extent_buffer_page(eb, i);
3357                 i++;
3358                 unlock_page(page);
3359                 locked_pages--;
3360         }
3361         return ret;
3362 }
3363
3364 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3365                         unsigned long start,
3366                         unsigned long len)
3367 {
3368         size_t cur;
3369         size_t offset;
3370         struct page *page;
3371         char *kaddr;
3372         char *dst = (char *)dstv;
3373         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3374         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3375
3376         WARN_ON(start > eb->len);
3377         WARN_ON(start + len > eb->start + eb->len);
3378
3379         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3380
3381         while (len > 0) {
3382                 page = extent_buffer_page(eb, i);
3383
3384                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3385                 kaddr = kmap_atomic(page, KM_USER1);
3386                 memcpy(dst, kaddr + offset, cur);
3387                 kunmap_atomic(kaddr, KM_USER1);
3388
3389                 dst += cur;
3390                 len -= cur;
3391                 offset = 0;
3392                 i++;
3393         }
3394 }
3395
3396 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3397                                unsigned long min_len, char **token, char **map,
3398                                unsigned long *map_start,
3399                                unsigned long *map_len, int km)
3400 {
3401         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3402         char *kaddr;
3403         struct page *p;
3404         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3405         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3406         unsigned long end_i = (start_offset + start + min_len - 1) >>
3407                 PAGE_CACHE_SHIFT;
3408
3409         if (i != end_i)
3410                 return -EINVAL;
3411
3412         if (i == 0) {
3413                 offset = start_offset;
3414                 *map_start = 0;
3415         } else {
3416                 offset = 0;
3417                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3418         }
3419
3420         if (start + min_len > eb->len) {
3421                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3422                        "wanted %lu %lu\n", (unsigned long long)eb->start,
3423                        eb->len, start, min_len);
3424                 WARN_ON(1);
3425         }
3426
3427         p = extent_buffer_page(eb, i);
3428         kaddr = kmap_atomic(p, km);
3429         *token = kaddr;
3430         *map = kaddr + offset;
3431         *map_len = PAGE_CACHE_SIZE - offset;
3432         return 0;
3433 }
3434
3435 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3436                       unsigned long min_len,
3437                       char **token, char **map,
3438                       unsigned long *map_start,
3439                       unsigned long *map_len, int km)
3440 {
3441         int err;
3442         int save = 0;
3443         if (eb->map_token) {
3444                 unmap_extent_buffer(eb, eb->map_token, km);
3445                 eb->map_token = NULL;
3446                 save = 1;
3447         }
3448         err = map_private_extent_buffer(eb, start, min_len, token, map,
3449                                        map_start, map_len, km);
3450         if (!err && save) {
3451                 eb->map_token = *token;
3452                 eb->kaddr = *map;
3453                 eb->map_start = *map_start;
3454                 eb->map_len = *map_len;
3455         }
3456         return err;
3457 }
3458
3459 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3460 {
3461         kunmap_atomic(token, km);
3462 }
3463
3464 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3465                           unsigned long start,
3466                           unsigned long len)
3467 {
3468         size_t cur;
3469         size_t offset;
3470         struct page *page;
3471         char *kaddr;
3472         char *ptr = (char *)ptrv;
3473         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3474         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3475         int ret = 0;
3476
3477         WARN_ON(start > eb->len);
3478         WARN_ON(start + len > eb->start + eb->len);
3479
3480         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3481
3482         while (len > 0) {
3483                 page = extent_buffer_page(eb, i);
3484
3485                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3486
3487                 kaddr = kmap_atomic(page, KM_USER0);
3488                 ret = memcmp(ptr, kaddr + offset, cur);
3489                 kunmap_atomic(kaddr, KM_USER0);
3490                 if (ret)
3491                         break;
3492
3493                 ptr += cur;
3494                 len -= cur;
3495                 offset = 0;
3496                 i++;
3497         }
3498         return ret;
3499 }
3500
3501 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3502                          unsigned long start, unsigned long len)
3503 {
3504         size_t cur;
3505         size_t offset;
3506         struct page *page;
3507         char *kaddr;
3508         char *src = (char *)srcv;
3509         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3510         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3511
3512         WARN_ON(start > eb->len);
3513         WARN_ON(start + len > eb->start + eb->len);
3514
3515         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3516
3517         while (len > 0) {
3518                 page = extent_buffer_page(eb, i);
3519                 WARN_ON(!PageUptodate(page));
3520
3521                 cur = min(len, PAGE_CACHE_SIZE - offset);
3522                 kaddr = kmap_atomic(page, KM_USER1);
3523                 memcpy(kaddr + offset, src, cur);
3524                 kunmap_atomic(kaddr, KM_USER1);
3525
3526                 src += cur;
3527                 len -= cur;
3528                 offset = 0;
3529                 i++;
3530         }
3531 }
3532
3533 void memset_extent_buffer(struct extent_buffer *eb, char c,
3534                           unsigned long start, unsigned long len)
3535 {
3536         size_t cur;
3537         size_t offset;
3538         struct page *page;
3539         char *kaddr;
3540         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3541         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3542
3543         WARN_ON(start > eb->len);
3544         WARN_ON(start + len > eb->start + eb->len);
3545
3546         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3547
3548         while (len > 0) {
3549                 page = extent_buffer_page(eb, i);
3550                 WARN_ON(!PageUptodate(page));
3551
3552                 cur = min(len, PAGE_CACHE_SIZE - offset);
3553                 kaddr = kmap_atomic(page, KM_USER0);
3554                 memset(kaddr + offset, c, cur);
3555                 kunmap_atomic(kaddr, KM_USER0);
3556
3557                 len -= cur;
3558                 offset = 0;
3559                 i++;
3560         }
3561 }
3562
3563 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3564                         unsigned long dst_offset, unsigned long src_offset,
3565                         unsigned long len)
3566 {
3567         u64 dst_len = dst->len;
3568         size_t cur;
3569         size_t offset;
3570         struct page *page;
3571         char *kaddr;
3572         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3573         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3574
3575         WARN_ON(src->len != dst_len);
3576
3577         offset = (start_offset + dst_offset) &
3578                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3579
3580         while (len > 0) {
3581                 page = extent_buffer_page(dst, i);
3582                 WARN_ON(!PageUptodate(page));
3583
3584                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3585
3586                 kaddr = kmap_atomic(page, KM_USER0);
3587                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3588                 kunmap_atomic(kaddr, KM_USER0);
3589
3590                 src_offset += cur;
3591                 len -= cur;
3592                 offset = 0;
3593                 i++;
3594         }
3595 }
3596
3597 static void move_pages(struct page *dst_page, struct page *src_page,
3598                        unsigned long dst_off, unsigned long src_off,
3599                        unsigned long len)
3600 {
3601         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3602         if (dst_page == src_page) {
3603                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3604         } else {
3605                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3606                 char *p = dst_kaddr + dst_off + len;
3607                 char *s = src_kaddr + src_off + len;
3608
3609                 while (len--)
3610                         *--p = *--s;
3611
3612                 kunmap_atomic(src_kaddr, KM_USER1);
3613         }
3614         kunmap_atomic(dst_kaddr, KM_USER0);
3615 }
3616
3617 static void copy_pages(struct page *dst_page, struct page *src_page,
3618                        unsigned long dst_off, unsigned long src_off,
3619                        unsigned long len)
3620 {
3621         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3622         char *src_kaddr;
3623
3624         if (dst_page != src_page)
3625                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3626         else
3627                 src_kaddr = dst_kaddr;
3628
3629         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3630         kunmap_atomic(dst_kaddr, KM_USER0);
3631         if (dst_page != src_page)
3632                 kunmap_atomic(src_kaddr, KM_USER1);
3633 }
3634
3635 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3636                            unsigned long src_offset, unsigned long len)
3637 {
3638         size_t cur;
3639         size_t dst_off_in_page;
3640         size_t src_off_in_page;
3641         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3642         unsigned long dst_i;
3643         unsigned long src_i;
3644
3645         if (src_offset + len > dst->len) {
3646                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3647                        "len %lu dst len %lu\n", src_offset, len, dst->len);
3648                 BUG_ON(1);
3649         }
3650         if (dst_offset + len > dst->len) {
3651                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3652                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
3653                 BUG_ON(1);
3654         }
3655
3656         while (len > 0) {
3657                 dst_off_in_page = (start_offset + dst_offset) &
3658                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3659                 src_off_in_page = (start_offset + src_offset) &
3660                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3661
3662                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3663                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3664
3665                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3666                                                src_off_in_page));
3667                 cur = min_t(unsigned long, cur,
3668                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3669
3670                 copy_pages(extent_buffer_page(dst, dst_i),
3671                            extent_buffer_page(dst, src_i),
3672                            dst_off_in_page, src_off_in_page, cur);
3673
3674                 src_offset += cur;
3675                 dst_offset += cur;
3676                 len -= cur;
3677         }
3678 }
3679
3680 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3681                            unsigned long src_offset, unsigned long len)
3682 {
3683         size_t cur;
3684         size_t dst_off_in_page;
3685         size_t src_off_in_page;
3686         unsigned long dst_end = dst_offset + len - 1;
3687         unsigned long src_end = src_offset + len - 1;
3688         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3689         unsigned long dst_i;
3690         unsigned long src_i;
3691
3692         if (src_offset + len > dst->len) {
3693                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3694                        "len %lu len %lu\n", src_offset, len, dst->len);
3695                 BUG_ON(1);
3696         }
3697         if (dst_offset + len > dst->len) {
3698                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3699                        "len %lu len %lu\n", dst_offset, len, dst->len);
3700                 BUG_ON(1);
3701         }
3702         if (dst_offset < src_offset) {
3703                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3704                 return;
3705         }
3706         while (len > 0) {
3707                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3708                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3709
3710                 dst_off_in_page = (start_offset + dst_end) &
3711                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3712                 src_off_in_page = (start_offset + src_end) &
3713                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3714
3715                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3716                 cur = min(cur, dst_off_in_page + 1);
3717                 move_pages(extent_buffer_page(dst, dst_i),
3718                            extent_buffer_page(dst, src_i),
3719                            dst_off_in_page - cur + 1,
3720                            src_off_in_page - cur + 1, cur);
3721
3722                 dst_end -= cur;
3723                 src_end -= cur;
3724                 len -= cur;
3725         }
3726 }
3727
3728 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3729 {
3730         u64 start = page_offset(page);
3731         struct extent_buffer *eb;
3732         int ret = 1;
3733         unsigned long i;
3734         unsigned long num_pages;
3735
3736         spin_lock(&tree->buffer_lock);
3737         eb = buffer_search(tree, start);
3738         if (!eb)
3739                 goto out;
3740
3741         if (atomic_read(&eb->refs) > 1) {
3742                 ret = 0;
3743                 goto out;
3744         }
3745         if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3746                 ret = 0;
3747                 goto out;
3748         }
3749         /* at this point we can safely release the extent buffer */
3750         num_pages = num_extent_pages(eb->start, eb->len);
3751         for (i = 0; i < num_pages; i++)
3752                 page_cache_release(extent_buffer_page(eb, i));
3753         rb_erase(&eb->rb_node, &tree->buffer);
3754         __free_extent_buffer(eb);
3755 out:
3756         spin_unlock(&tree->buffer_lock);
3757         return ret;
3758 }