fs/btrfs/extent_io.c

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