fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/gfp.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/page-flags.h>
   8 #include <linux/module.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/swap.h>
  12 #include <linux/writeback.h>
  13 #include <linux/pagevec.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "compat.h"
  17 #include "ctree.h"
  18 #include "btrfs_inode.h"
  19
  20 static struct kmem_cache *extent_state_cache;
  21 static struct kmem_cache *extent_buffer_cache;
  22
  23 static LIST_HEAD(buffers);
  24 static LIST_HEAD(states);
  25
  26 #define LEAK_DEBUG 0
  27 #if LEAK_DEBUG
  28 static DEFINE_SPINLOCK(leak_lock);
  29 #endif
  30
  31 #define BUFFER_LRU_MAX 64
  32
  33 struct tree_entry {
  34         u64 start;
  35         u64 end;
  36         struct rb_node rb_node;
  37 };
  38
  39 struct extent_page_data {
  40         struct bio *bio;
  41         struct extent_io_tree *tree;
  42         get_extent_t *get_extent;
  43
  44         /* tells writepage not to lock the state bits for this range
  45          * it still does the unlocking
  46          */
  47         unsigned int extent_locked:1;
  48
  49         /* tells the submit_bio code to use a WRITE_SYNC */
  50         unsigned int sync_io:1;
  51 };
  52
  53 int __init extent_io_init(void)
  54 {
  55         extent_state_cache = kmem_cache_create("extent_state",
  56                         sizeof(struct extent_state), 0,
  57                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  58         if (!extent_state_cache)
  59                 return -ENOMEM;
  60
  61         extent_buffer_cache = kmem_cache_create("extent_buffers",
  62                         sizeof(struct extent_buffer), 0,
  63                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  64         if (!extent_buffer_cache)
  65                 goto free_state_cache;
  66         return 0;
  67
  68 free_state_cache:
  69         kmem_cache_destroy(extent_state_cache);
  70         return -ENOMEM;
  71 }
  72
  73 void extent_io_exit(void)
  74 {
  75         struct extent_state *state;
  76         struct extent_buffer *eb;
  77
  78         while (!list_empty(&states)) {
  79                 state = list_entry(states.next, struct extent_state, leak_list);
  80                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  81                        "state %lu in tree %p refs %d\n",
  82                        (unsigned long long)state->start,
  83                        (unsigned long long)state->end,
  84                        state->state, state->tree, atomic_read(&state->refs));
  85                 list_del(&state->leak_list);
  86                 kmem_cache_free(extent_state_cache, state);
  87
  88         }
  89
  90         while (!list_empty(&buffers)) {
  91                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  92                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  93                        "refs %d\n", (unsigned long long)eb->start,
  94                        eb->len, atomic_read(&eb->refs));
  95                 list_del(&eb->leak_list);
  96                 kmem_cache_free(extent_buffer_cache, eb);
  97         }
  98         if (extent_state_cache)
  99                 kmem_cache_destroy(extent_state_cache);
 100         if (extent_buffer_cache)
 101                 kmem_cache_destroy(extent_buffer_cache);
 102 }
 103
 104 void extent_io_tree_init(struct extent_io_tree *tree,
 105                           struct address_space *mapping, gfp_t mask)
 106 {
 107         tree->state = RB_ROOT;
 108         tree->buffer = RB_ROOT;
 109         tree->ops = NULL;
 110         tree->dirty_bytes = 0;
 111         spin_lock_init(&tree->lock);
 112         spin_lock_init(&tree->buffer_lock);
 113         tree->mapping = mapping;
 114 }
 115
 116 static struct extent_state *alloc_extent_state(gfp_t mask)
 117 {
 118         struct extent_state *state;
 119 #if LEAK_DEBUG
 120         unsigned long flags;
 121 #endif
 122
 123         state = kmem_cache_alloc(extent_state_cache, mask);
 124         if (!state)
 125                 return state;
 126         state->state = 0;
 127         state->private = 0;
 128         state->tree = NULL;
 129 #if LEAK_DEBUG
 130         spin_lock_irqsave(&leak_lock, flags);
 131         list_add(&state->leak_list, &states);
 132         spin_unlock_irqrestore(&leak_lock, flags);
 133 #endif
 134         atomic_set(&state->refs, 1);
 135         init_waitqueue_head(&state->wq);
 136         return state;
 137 }
 138
 139 static void free_extent_state(struct extent_state *state)
 140 {
 141         if (!state)
 142                 return;
 143         if (atomic_dec_and_test(&state->refs)) {
 144 #if LEAK_DEBUG
 145                 unsigned long flags;
 146 #endif
 147                 WARN_ON(state->tree);
 148 #if LEAK_DEBUG
 149                 spin_lock_irqsave(&leak_lock, flags);
 150                 list_del(&state->leak_list);
 151                 spin_unlock_irqrestore(&leak_lock, flags);
 152 #endif
 153                 kmem_cache_free(extent_state_cache, state);
 154         }
 155 }
 156
 157 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 158                                    struct rb_node *node)
 159 {
 160         struct rb_node **p = &root->rb_node;
 161         struct rb_node *parent = NULL;
 162         struct tree_entry *entry;
 163
 164         while (*p) {
 165                 parent = *p;
 166                 entry = rb_entry(parent, struct tree_entry, rb_node);
 167
 168                 if (offset < entry->start)
 169                         p = &(*p)->rb_left;
 170                 else if (offset > entry->end)
 171                         p = &(*p)->rb_right;
 172                 else
 173                         return parent;
 174         }
 175
 176         entry = rb_entry(node, struct tree_entry, rb_node);
 177         rb_link_node(node, parent, p);
 178         rb_insert_color(node, root);
 179         return NULL;
 180 }
 181
 182 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 183                                      struct rb_node **prev_ret,
 184                                      struct rb_node **next_ret)
 185 {
 186         struct rb_root *root = &tree->state;
 187         struct rb_node *n = root->rb_node;
 188         struct rb_node *prev = NULL;
 189         struct rb_node *orig_prev = NULL;
 190         struct tree_entry *entry;
 191         struct tree_entry *prev_entry = NULL;
 192
 193         while (n) {
 194                 entry = rb_entry(n, struct tree_entry, rb_node);
 195                 prev = n;
 196                 prev_entry = entry;
 197
 198                 if (offset < entry->start)
 199                         n = n->rb_left;
 200                 else if (offset > entry->end)
 201                         n = n->rb_right;
 202                 else
 203                         return n;
 204         }
 205
 206         if (prev_ret) {
 207                 orig_prev = prev;
 208                 while (prev && offset > prev_entry->end) {
 209                         prev = rb_next(prev);
 210                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 211                 }
 212                 *prev_ret = prev;
 213                 prev = orig_prev;
 214         }
 215
 216         if (next_ret) {
 217                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 218                 while (prev && offset < prev_entry->start) {
 219                         prev = rb_prev(prev);
 220                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 221                 }
 222                 *next_ret = prev;
 223         }
 224         return NULL;
 225 }
 226
 227 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 228                                           u64 offset)
 229 {
 230         struct rb_node *prev = NULL;
 231         struct rb_node *ret;
 232
 233         ret = __etree_search(tree, offset, &prev, NULL);
 234         if (!ret)
 235                 return prev;
 236         return ret;
 237 }
 238
 239 static struct extent_buffer *buffer_tree_insert(struct extent_io_tree *tree,
 240                                           u64 offset, struct rb_node *node)
 241 {
 242         struct rb_root *root = &tree->buffer;
 243         struct rb_node **p = &root->rb_node;
 244         struct rb_node *parent = NULL;
 245         struct extent_buffer *eb;
 246
 247         while (*p) {
 248                 parent = *p;
 249                 eb = rb_entry(parent, struct extent_buffer, rb_node);
 250
 251                 if (offset < eb->start)
 252                         p = &(*p)->rb_left;
 253                 else if (offset > eb->start)
 254                         p = &(*p)->rb_right;
 255                 else
 256                         return eb;
 257         }
 258
 259         rb_link_node(node, parent, p);
 260         rb_insert_color(node, root);
 261         return NULL;
 262 }
 263
 264 static struct extent_buffer *buffer_search(struct extent_io_tree *tree,
 265                                            u64 offset)
 266 {
 267         struct rb_root *root = &tree->buffer;
 268         struct rb_node *n = root->rb_node;
 269         struct extent_buffer *eb;
 270
 271         while (n) {
 272                 eb = rb_entry(n, struct extent_buffer, rb_node);
 273                 if (offset < eb->start)
 274                         n = n->rb_left;
 275                 else if (offset > eb->start)
 276                         n = n->rb_right;
 277                 else
 278                         return eb;
 279         }
 280         return NULL;
 281 }
 282
 283 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 284                      struct extent_state *other)
 285 {
 286         if (tree->ops && tree->ops->merge_extent_hook)
 287                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 288                                              other);
 289 }
 290
 291 /*
 292  * utility function to look for merge candidates inside a given range.
 293  * Any extents with matching state are merged together into a single
 294  * extent in the tree.  Extents with EXTENT_IO in their state field
 295  * are not merged because the end_io handlers need to be able to do
 296  * operations on them without sleeping (or doing allocations/splits).
 297  *
 298  * This should be called with the tree lock held.
 299  */
 300 static int merge_state(struct extent_io_tree *tree,
 301                        struct extent_state *state)
 302 {
 303         struct extent_state *other;
 304         struct rb_node *other_node;
 305
 306         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 307                 return 0;
 308
 309         other_node = rb_prev(&state->rb_node);
 310         if (other_node) {
 311                 other = rb_entry(other_node, struct extent_state, rb_node);
 312                 if (other->end == state->start - 1 &&
 313                     other->state == state->state) {
 314                         merge_cb(tree, state, other);
 315                         state->start = other->start;
 316                         other->tree = NULL;
 317                         rb_erase(&other->rb_node, &tree->state);
 318                         free_extent_state(other);
 319                 }
 320         }
 321         other_node = rb_next(&state->rb_node);
 322         if (other_node) {
 323                 other = rb_entry(other_node, struct extent_state, rb_node);
 324                 if (other->start == state->end + 1 &&
 325                     other->state == state->state) {
 326                         merge_cb(tree, state, other);
 327                         other->start = state->start;
 328                         state->tree = NULL;
 329                         rb_erase(&state->rb_node, &tree->state);
 330                         free_extent_state(state);
 331                         state = NULL;
 332                 }
 333         }
 334
 335         return 0;
 336 }
 337
 338 static int set_state_cb(struct extent_io_tree *tree,
 339                          struct extent_state *state,
 340                          unsigned long bits)
 341 {
 342         if (tree->ops && tree->ops->set_bit_hook) {
 343                 return tree->ops->set_bit_hook(tree->mapping->host,
 344                                                state->start, state->end,
 345                                                state->state, bits);
 346         }
 347
 348         return 0;
 349 }
 350
 351 static void clear_state_cb(struct extent_io_tree *tree,
 352                            struct extent_state *state,
 353                            unsigned long bits)
 354 {
 355         if (tree->ops && tree->ops->clear_bit_hook)
 356                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 357 }
 358
 359 /*
 360  * insert an extent_state struct into the tree.  'bits' are set on the
 361  * struct before it is inserted.
 362  *
 363  * This may return -EEXIST if the extent is already there, in which case the
 364  * state struct is freed.
 365  *
 366  * The tree lock is not taken internally.  This is a utility function and
 367  * probably isn't what you want to call (see set/clear_extent_bit).
 368  */
 369 static int insert_state(struct extent_io_tree *tree,
 370                         struct extent_state *state, u64 start, u64 end,
 371                         int bits)
 372 {
 373         struct rb_node *node;
 374         int ret;
 375
 376         if (end < start) {
 377                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
 378                        (unsigned long long)end,
 379                        (unsigned long long)start);
 380                 WARN_ON(1);
 381         }
 382         state->start = start;
 383         state->end = end;
 384         ret = set_state_cb(tree, state, bits);
 385         if (ret)
 386                 return ret;
 387
 388         if (bits & EXTENT_DIRTY)
 389                 tree->dirty_bytes += end - start + 1;
 390         state->state |= bits;
 391         node = tree_insert(&tree->state, end, &state->rb_node);
 392         if (node) {
 393                 struct extent_state *found;
 394                 found = rb_entry(node, struct extent_state, rb_node);
 395                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 396                        "%llu %llu\n", (unsigned long long)found->start,
 397                        (unsigned long long)found->end,
 398                        (unsigned long long)start, (unsigned long long)end);
 399                 free_extent_state(state);
 400                 return -EEXIST;
 401         }
 402         state->tree = tree;
 403         merge_state(tree, state);
 404         return 0;
 405 }
 406
 407 static int split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 408                      u64 split)
 409 {
 410         if (tree->ops && tree->ops->split_extent_hook)
 411                 return tree->ops->split_extent_hook(tree->mapping->host,
 412                                                     orig, split);
 413         return 0;
 414 }
 415
 416 /*
 417  * split a given extent state struct in two, inserting the preallocated
 418  * struct 'prealloc' as the newly created second half.  'split' indicates an
 419  * offset inside 'orig' where it should be split.
 420  *
 421  * Before calling,
 422  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 423  * are two extent state structs in the tree:
 424  * prealloc: [orig->start, split - 1]
 425  * orig: [ split, orig->end ]
 426  *
 427  * The tree locks are not taken by this function. They need to be held
 428  * by the caller.
 429  */
 430 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 431                        struct extent_state *prealloc, u64 split)
 432 {
 433         struct rb_node *node;
 434
 435         split_cb(tree, orig, split);
 436
 437         prealloc->start = orig->start;
 438         prealloc->end = split - 1;
 439         prealloc->state = orig->state;
 440         orig->start = split;
 441
 442         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 443         if (node) {
 444                 free_extent_state(prealloc);
 445                 return -EEXIST;
 446         }
 447         prealloc->tree = tree;
 448         return 0;
 449 }
 450
 451 /*
 452  * utility function to clear some bits in an extent state struct.
 453  * it will optionally wake up any one waiting on this state (wake == 1), or
 454  * forcibly remove the state from the tree (delete == 1).
 455  *
 456  * If no bits are set on the state struct after clearing things, the
 457  * struct is freed and removed from the tree
 458  */
 459 static int clear_state_bit(struct extent_io_tree *tree,
 460                             struct extent_state *state, int bits, int wake,
 461                             int delete)
 462 {
 463         int bits_to_clear = bits & ~EXTENT_DO_ACCOUNTING;
 464         int ret = state->state & bits_to_clear;
 465
 466         if ((bits & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 467                 u64 range = state->end - state->start + 1;
 468                 WARN_ON(range > tree->dirty_bytes);
 469                 tree->dirty_bytes -= range;
 470         }
 471         clear_state_cb(tree, state, bits);
 472         state->state &= ~bits_to_clear;
 473         if (wake)
 474                 wake_up(&state->wq);
 475         if (delete || state->state == 0) {
 476                 if (state->tree) {
 477                         clear_state_cb(tree, state, state->state);
 478                         rb_erase(&state->rb_node, &tree->state);
 479                         state->tree = NULL;
 480                         free_extent_state(state);
 481                 } else {
 482                         WARN_ON(1);
 483                 }
 484         } else {
 485                 merge_state(tree, state);
 486         }
 487         return ret;
 488 }
 489
 490 /*
 491  * clear some bits on a range in the tree.  This may require splitting
 492  * or inserting elements in the tree, so the gfp mask is used to
 493  * indicate which allocations or sleeping are allowed.
 494  *
 495  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 496  * the given range from the tree regardless of state (ie for truncate).
 497  *
 498  * the range [start, end] is inclusive.
 499  *
 500  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 501  * bits were already set, or zero if none of the bits were already set.
 502  */
 503 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 504                      int bits, int wake, int delete,
 505                      struct extent_state **cached_state,
 506                      gfp_t mask)
 507 {
 508         struct extent_state *state;
 509         struct extent_state *cached;
 510         struct extent_state *prealloc = NULL;
 511         struct rb_node *next_node;
 512         struct rb_node *node;
 513         u64 last_end;
 514         int err;
 515         int set = 0;
 516
 517 again:
 518         if (!prealloc && (mask & __GFP_WAIT)) {
 519                 prealloc = alloc_extent_state(mask);
 520                 if (!prealloc)
 521                         return -ENOMEM;
 522         }
 523
 524         spin_lock(&tree->lock);
 525         if (cached_state) {
 526                 cached = *cached_state;
 527                 *cached_state = NULL;
 528                 cached_state = NULL;
 529                 if (cached && cached->tree && cached->start == start) {
 530                         atomic_dec(&cached->refs);
 531                         state = cached;
 532                         goto hit_next;
 533                 }
 534                 free_extent_state(cached);
 535         }
 536         /*
 537          * this search will find the extents that end after
 538          * our range starts
 539          */
 540         node = tree_search(tree, start);
 541         if (!node)
 542                 goto out;
 543         state = rb_entry(node, struct extent_state, rb_node);
 544 hit_next:
 545         if (state->start > end)
 546                 goto out;
 547         WARN_ON(state->end < start);
 548         last_end = state->end;
 549
 550         /*
 551          *     | ---- desired range ---- |
 552          *  | state | or
 553          *  | ------------- state -------------- |
 554          *
 555          * We need to split the extent we found, and may flip
 556          * bits on second half.
 557          *
 558          * If the extent we found extends past our range, we
 559          * just split and search again.  It'll get split again
 560          * the next time though.
 561          *
 562          * If the extent we found is inside our range, we clear
 563          * the desired bit on it.
 564          */
 565
 566         if (state->start < start) {
 567                 if (!prealloc)
 568                         prealloc = alloc_extent_state(GFP_ATOMIC);
 569                 err = split_state(tree, state, prealloc, start);
 570                 BUG_ON(err == -EEXIST);
 571                 prealloc = NULL;
 572                 if (err)
 573                         goto out;
 574                 if (state->end <= end) {
 575                         set |= clear_state_bit(tree, state, bits, wake,
 576                                                delete);
 577                         if (last_end == (u64)-1)
 578                                 goto out;
 579                         start = last_end + 1;
 580                 }
 581                 goto search_again;
 582         }
 583         /*
 584          * | ---- desired range ---- |
 585          *                        | state |
 586          * We need to split the extent, and clear the bit
 587          * on the first half
 588          */
 589         if (state->start <= end && state->end > end) {
 590                 if (!prealloc)
 591                         prealloc = alloc_extent_state(GFP_ATOMIC);
 592                 err = split_state(tree, state, prealloc, end + 1);
 593                 BUG_ON(err == -EEXIST);
 594                 if (wake)
 595                         wake_up(&state->wq);
 596
 597                 set |= clear_state_bit(tree, prealloc, bits, wake, delete);
 598
 599                 prealloc = NULL;
 600                 goto out;
 601         }
 602
 603         if (state->end < end && prealloc && !need_resched())
 604                 next_node = rb_next(&state->rb_node);
 605         else
 606                 next_node = NULL;
 607
 608         set |= clear_state_bit(tree, state, bits, wake, delete);
 609         if (last_end == (u64)-1)
 610                 goto out;
 611         start = last_end + 1;
 612         if (start <= end && next_node) {
 613                 state = rb_entry(next_node, struct extent_state,
 614                                  rb_node);
 615                 if (state->start == start)
 616                         goto hit_next;
 617         }
 618         goto search_again;
 619
 620 out:
 621         spin_unlock(&tree->lock);
 622         if (prealloc)
 623                 free_extent_state(prealloc);
 624
 625         return set;
 626
 627 search_again:
 628         if (start > end)
 629                 goto out;
 630         spin_unlock(&tree->lock);
 631         if (mask & __GFP_WAIT)
 632                 cond_resched();
 633         goto again;
 634 }
 635
 636 static int wait_on_state(struct extent_io_tree *tree,
 637                          struct extent_state *state)
 638                 __releases(tree->lock)
 639                 __acquires(tree->lock)
 640 {
 641         DEFINE_WAIT(wait);
 642         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 643         spin_unlock(&tree->lock);
 644         schedule();
 645         spin_lock(&tree->lock);
 646         finish_wait(&state->wq, &wait);
 647         return 0;
 648 }
 649
 650 /*
 651  * waits for one or more bits to clear on a range in the state tree.
 652  * The range [start, end] is inclusive.
 653  * The tree lock is taken by this function
 654  */
 655 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 656 {
 657         struct extent_state *state;
 658         struct rb_node *node;
 659
 660         spin_lock(&tree->lock);
 661 again:
 662         while (1) {
 663                 /*
 664                  * this search will find all the extents that end after
 665                  * our range starts
 666                  */
 667                 node = tree_search(tree, start);
 668                 if (!node)
 669                         break;
 670
 671                 state = rb_entry(node, struct extent_state, rb_node);
 672
 673                 if (state->start > end)
 674                         goto out;
 675
 676                 if (state->state & bits) {
 677                         start = state->start;
 678                         atomic_inc(&state->refs);
 679                         wait_on_state(tree, state);
 680                         free_extent_state(state);
 681                         goto again;
 682                 }
 683                 start = state->end + 1;
 684
 685                 if (start > end)
 686                         break;
 687
 688                 if (need_resched()) {
 689                         spin_unlock(&tree->lock);
 690                         cond_resched();
 691                         spin_lock(&tree->lock);
 692                 }
 693         }
 694 out:
 695         spin_unlock(&tree->lock);
 696         return 0;
 697 }
 698
 699 static int set_state_bits(struct extent_io_tree *tree,
 700                            struct extent_state *state,
 701                            int bits)
 702 {
 703         int ret;
 704
 705         ret = set_state_cb(tree, state, bits);
 706         if (ret)
 707                 return ret;
 708
 709         if ((bits & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 710                 u64 range = state->end - state->start + 1;
 711                 tree->dirty_bytes += range;
 712         }
 713         state->state |= bits;
 714
 715         return 0;
 716 }
 717
 718 static void cache_state(struct extent_state *state,
 719                         struct extent_state **cached_ptr)
 720 {
 721         if (cached_ptr && !(*cached_ptr)) {
 722                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 723                         *cached_ptr = state;
 724                         atomic_inc(&state->refs);
 725                 }
 726         }
 727 }
 728
 729 /*
 730  * set some bits on a range in the tree.  This may require allocations or
 731  * sleeping, so the gfp mask is used to indicate what is allowed.
 732  *
 733  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 734  * part of the range already has the desired bits set.  The start of the
 735  * existing range is returned in failed_start in this case.
 736  *
 737  * [start, end] is inclusive This takes the tree lock.
 738  */
 739
 740 static int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 741                           int bits, int exclusive_bits, u64 *failed_start,
 742                           struct extent_state **cached_state,
 743                           gfp_t mask)
 744 {
 745         struct extent_state *state;
 746         struct extent_state *prealloc = NULL;
 747         struct rb_node *node;
 748         int err = 0;
 749         u64 last_start;
 750         u64 last_end;
 751
 752 again:
 753         if (!prealloc && (mask & __GFP_WAIT)) {
 754                 prealloc = alloc_extent_state(mask);
 755                 if (!prealloc)
 756                         return -ENOMEM;
 757         }
 758
 759         spin_lock(&tree->lock);
 760         if (cached_state && *cached_state) {
 761                 state = *cached_state;
 762                 if (state->start == start && state->tree) {
 763                         node = &state->rb_node;
 764                         goto hit_next;
 765                 }
 766         }
 767         /*
 768          * this search will find all the extents that end after
 769          * our range starts.
 770          */
 771         node = tree_search(tree, start);
 772         if (!node) {
 773                 err = insert_state(tree, prealloc, start, end, bits);
 774                 prealloc = NULL;
 775                 BUG_ON(err == -EEXIST);
 776                 goto out;
 777         }
 778         state = rb_entry(node, struct extent_state, rb_node);
 779 hit_next:
 780         last_start = state->start;
 781         last_end = state->end;
 782
 783         /*
 784          * | ---- desired range ---- |
 785          * | state |
 786          *
 787          * Just lock what we found and keep going
 788          */
 789         if (state->start == start && state->end <= end) {
 790                 struct rb_node *next_node;
 791                 if (state->state & exclusive_bits) {
 792                         *failed_start = state->start;
 793                         err = -EEXIST;
 794                         goto out;
 795                 }
 796
 797                 err = set_state_bits(tree, state, bits);
 798                 if (err)
 799                         goto out;
 800
 801                 cache_state(state, cached_state);
 802                 merge_state(tree, state);
 803                 if (last_end == (u64)-1)
 804                         goto out;
 805
 806                 start = last_end + 1;
 807                 if (start < end && prealloc && !need_resched()) {
 808                         next_node = rb_next(node);
 809                         if (next_node) {
 810                                 state = rb_entry(next_node, struct extent_state,
 811                                                  rb_node);
 812                                 if (state->start == start)
 813                                         goto hit_next;
 814                         }
 815                 }
 816                 goto search_again;
 817         }
 818
 819         /*
 820          *     | ---- desired range ---- |
 821          * | state |
 822          *   or
 823          * | ------------- state -------------- |
 824          *
 825          * We need to split the extent we found, and may flip bits on
 826          * second half.
 827          *
 828          * If the extent we found extends past our
 829          * range, we just split and search again.  It'll get split
 830          * again the next time though.
 831          *
 832          * If the extent we found is inside our range, we set the
 833          * desired bit on it.
 834          */
 835         if (state->start < start) {
 836                 if (state->state & exclusive_bits) {
 837                         *failed_start = start;
 838                         err = -EEXIST;
 839                         goto out;
 840                 }
 841                 err = split_state(tree, state, prealloc, start);
 842                 BUG_ON(err == -EEXIST);
 843                 prealloc = NULL;
 844                 if (err)
 845                         goto out;
 846                 if (state->end <= end) {
 847                         err = set_state_bits(tree, state, bits);
 848                         if (err)
 849                                 goto out;
 850                         cache_state(state, cached_state);
 851                         merge_state(tree, state);
 852                         if (last_end == (u64)-1)
 853                                 goto out;
 854                         start = last_end + 1;
 855                 }
 856                 goto search_again;
 857         }
 858         /*
 859          * | ---- desired range ---- |
 860          *     | state | or               | state |
 861          *
 862          * There's a hole, we need to insert something in it and
 863          * ignore the extent we found.
 864          */
 865         if (state->start > start) {
 866                 u64 this_end;
 867                 if (end < last_start)
 868                         this_end = end;
 869                 else
 870                         this_end = last_start - 1;
 871                 err = insert_state(tree, prealloc, start, this_end,
 872                                    bits);
 873                 BUG_ON(err == -EEXIST);
 874                 if (err) {
 875                         prealloc = NULL;
 876                         goto out;
 877                 }
 878                 cache_state(prealloc, cached_state);
 879                 prealloc = NULL;
 880                 start = this_end + 1;
 881                 goto search_again;
 882         }
 883         /*
 884          * | ---- desired range ---- |
 885          *                        | state |
 886          * We need to split the extent, and set the bit
 887          * on the first half
 888          */
 889         if (state->start <= end && state->end > end) {
 890                 if (state->state & exclusive_bits) {
 891                         *failed_start = start;
 892                         err = -EEXIST;
 893                         goto out;
 894                 }
 895                 err = split_state(tree, state, prealloc, end + 1);
 896                 BUG_ON(err == -EEXIST);
 897
 898                 err = set_state_bits(tree, prealloc, bits);
 899                 if (err) {
 900                         prealloc = NULL;
 901                         goto out;
 902                 }
 903                 cache_state(prealloc, cached_state);
 904                 merge_state(tree, prealloc);
 905                 prealloc = NULL;
 906                 goto out;
 907         }
 908
 909         goto search_again;
 910
 911 out:
 912         spin_unlock(&tree->lock);
 913         if (prealloc)
 914                 free_extent_state(prealloc);
 915
 916         return err;
 917
 918 search_again:
 919         if (start > end)
 920                 goto out;
 921         spin_unlock(&tree->lock);
 922         if (mask & __GFP_WAIT)
 923                 cond_resched();
 924         goto again;
 925 }
 926
 927 /* wrappers around set/clear extent bit */
 928 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 929                      gfp_t mask)
 930 {
 931         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 932                               NULL, mask);
 933 }
 934
 935 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 936                     int bits, gfp_t mask)
 937 {
 938         return set_extent_bit(tree, start, end, bits, 0, NULL,
 939                               NULL, mask);
 940 }
 941
 942 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 943                       int bits, gfp_t mask)
 944 {
 945         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
 946 }
 947
 948 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
 949                      gfp_t mask)
 950 {
 951         return set_extent_bit(tree, start, end,
 952                               EXTENT_DELALLOC | EXTENT_DIRTY | EXTENT_UPTODATE,
 953                               0, NULL, NULL, mask);
 954 }
 955
 956 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 957                        gfp_t mask)
 958 {
 959         return clear_extent_bit(tree, start, end,
 960                                 EXTENT_DIRTY | EXTENT_DELALLOC |
 961                                 EXTENT_DO_ACCOUNTING, 0, 0,
 962                                 NULL, mask);
 963 }
 964
 965 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 966                      gfp_t mask)
 967 {
 968         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 969                               NULL, mask);
 970 }
 971
 972 static int clear_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 973                        gfp_t mask)
 974 {
 975         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0,
 976                                 NULL, mask);
 977 }
 978
 979 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 980                         gfp_t mask)
 981 {
 982         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, NULL,
 983                               NULL, mask);
 984 }
 985
 986 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
 987                                  u64 end, gfp_t mask)
 988 {
 989         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
 990                                 NULL, mask);
 991 }
 992
 993 int wait_on_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 994 {
 995         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 996 }
 997
 998 /*
 999  * either insert or lock state struct between start and end use mask to tell
1000  * us if waiting is desired.
1001  */
1002 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1003                      int bits, struct extent_state **cached_state, gfp_t mask)
1004 {
1005         int err;
1006         u64 failed_start;
1007         while (1) {
1008                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1009                                      EXTENT_LOCKED, &failed_start,
1010                                      cached_state, mask);
1011                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
1012                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1013                         start = failed_start;
1014                 } else {
1015                         break;
1016                 }
1017                 WARN_ON(start > end);
1018         }
1019         return err;
1020 }
1021
1022 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1023 {
1024         return lock_extent_bits(tree, start, end, 0, NULL, mask);
1025 }
1026
1027 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1028                     gfp_t mask)
1029 {
1030         int err;
1031         u64 failed_start;
1032
1033         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1034                              &failed_start, NULL, mask);
1035         if (err == -EEXIST) {
1036                 if (failed_start > start)
1037                         clear_extent_bit(tree, start, failed_start - 1,
1038                                          EXTENT_LOCKED, 1, 0, NULL, mask);
1039                 return 0;
1040         }
1041         return 1;
1042 }
1043
1044 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1045                          struct extent_state **cached, gfp_t mask)
1046 {
1047         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1048                                 mask);
1049 }
1050
1051 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1052                   gfp_t mask)
1053 {
1054         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1055                                 mask);
1056 }
1057
1058 /*
1059  * helper function to set pages and extents in the tree dirty
1060  */
1061 int set_range_dirty(struct extent_io_tree *tree, u64 start, u64 end)
1062 {
1063         unsigned long index = start >> PAGE_CACHE_SHIFT;
1064         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1065         struct page *page;
1066
1067         while (index <= end_index) {
1068                 page = find_get_page(tree->mapping, index);
1069                 BUG_ON(!page);
1070                 __set_page_dirty_nobuffers(page);
1071                 page_cache_release(page);
1072                 index++;
1073         }
1074         return 0;
1075 }
1076
1077 /*
1078  * helper function to set both pages and extents in the tree writeback
1079  */
1080 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1081 {
1082         unsigned long index = start >> PAGE_CACHE_SHIFT;
1083         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1084         struct page *page;
1085
1086         while (index <= end_index) {
1087                 page = find_get_page(tree->mapping, index);
1088                 BUG_ON(!page);
1089                 set_page_writeback(page);
1090                 page_cache_release(page);
1091                 index++;
1092         }
1093         return 0;
1094 }
1095
1096 /*
1097  * find the first offset in the io tree with 'bits' set. zero is
1098  * returned if we find something, and *start_ret and *end_ret are
1099  * set to reflect the state struct that was found.
1100  *
1101  * If nothing was found, 1 is returned, < 0 on error
1102  */
1103 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1104                           u64 *start_ret, u64 *end_ret, int bits)
1105 {
1106         struct rb_node *node;
1107         struct extent_state *state;
1108         int ret = 1;
1109
1110         spin_lock(&tree->lock);
1111         /*
1112          * this search will find all the extents that end after
1113          * our range starts.
1114          */
1115         node = tree_search(tree, start);
1116         if (!node)
1117                 goto out;
1118
1119         while (1) {
1120                 state = rb_entry(node, struct extent_state, rb_node);
1121                 if (state->end >= start && (state->state & bits)) {
1122                         *start_ret = state->start;
1123                         *end_ret = state->end;
1124                         ret = 0;
1125                         break;
1126                 }
1127                 node = rb_next(node);
1128                 if (!node)
1129                         break;
1130         }
1131 out:
1132         spin_unlock(&tree->lock);
1133         return ret;
1134 }
1135
1136 /* find the first state struct with 'bits' set after 'start', and
1137  * return it.  tree->lock must be held.  NULL will returned if
1138  * nothing was found after 'start'
1139  */
1140 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1141                                                  u64 start, int bits)
1142 {
1143         struct rb_node *node;
1144         struct extent_state *state;
1145
1146         /*
1147          * this search will find all the extents that end after
1148          * our range starts.
1149          */
1150         node = tree_search(tree, start);
1151         if (!node)
1152                 goto out;
1153
1154         while (1) {
1155                 state = rb_entry(node, struct extent_state, rb_node);
1156                 if (state->end >= start && (state->state & bits))
1157                         return state;
1158
1159                 node = rb_next(node);
1160                 if (!node)
1161                         break;
1162         }
1163 out:
1164         return NULL;
1165 }
1166
1167 /*
1168  * find a contiguous range of bytes in the file marked as delalloc, not
1169  * more than 'max_bytes'.  start and end are used to return the range,
1170  *
1171  * 1 is returned if we find something, 0 if nothing was in the tree
1172  */
1173 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1174                                         u64 *start, u64 *end, u64 max_bytes,
1175                                         struct extent_state **cached_state)
1176 {
1177         struct rb_node *node;
1178         struct extent_state *state;
1179         u64 cur_start = *start;
1180         u64 found = 0;
1181         u64 total_bytes = 0;
1182
1183         spin_lock(&tree->lock);
1184
1185         /*
1186          * this search will find all the extents that end after
1187          * our range starts.
1188          */
1189         node = tree_search(tree, cur_start);
1190         if (!node) {
1191                 if (!found)
1192                         *end = (u64)-1;
1193                 goto out;
1194         }
1195
1196         while (1) {
1197                 state = rb_entry(node, struct extent_state, rb_node);
1198                 if (found && (state->start != cur_start ||
1199                               (state->state & EXTENT_BOUNDARY))) {
1200                         goto out;
1201                 }
1202                 if (!(state->state & EXTENT_DELALLOC)) {
1203                         if (!found)
1204                                 *end = state->end;
1205                         goto out;
1206                 }
1207                 if (!found) {
1208                         *start = state->start;
1209                         *cached_state = state;
1210                         atomic_inc(&state->refs);
1211                 }
1212                 found++;
1213                 *end = state->end;
1214                 cur_start = state->end + 1;
1215                 node = rb_next(node);
1216                 if (!node)
1217                         break;
1218                 total_bytes += state->end - state->start + 1;
1219                 if (total_bytes >= max_bytes)
1220                         break;
1221         }
1222 out:
1223         spin_unlock(&tree->lock);
1224         return found;
1225 }
1226
1227 static noinline int __unlock_for_delalloc(struct inode *inode,
1228                                           struct page *locked_page,
1229                                           u64 start, u64 end)
1230 {
1231         int ret;
1232         struct page *pages[16];
1233         unsigned long index = start >> PAGE_CACHE_SHIFT;
1234         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1235         unsigned long nr_pages = end_index - index + 1;
1236         int i;
1237
1238         if (index == locked_page->index && end_index == index)
1239                 return 0;
1240
1241         while (nr_pages > 0) {
1242                 ret = find_get_pages_contig(inode->i_mapping, index,
1243                                      min_t(unsigned long, nr_pages,
1244                                      ARRAY_SIZE(pages)), pages);
1245                 for (i = 0; i < ret; i++) {
1246                         if (pages[i] != locked_page)
1247                                 unlock_page(pages[i]);
1248                         page_cache_release(pages[i]);
1249                 }
1250                 nr_pages -= ret;
1251                 index += ret;
1252                 cond_resched();
1253         }
1254         return 0;
1255 }
1256
1257 static noinline int lock_delalloc_pages(struct inode *inode,
1258                                         struct page *locked_page,
1259                                         u64 delalloc_start,
1260                                         u64 delalloc_end)
1261 {
1262         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1263         unsigned long start_index = index;
1264         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1265         unsigned long pages_locked = 0;
1266         struct page *pages[16];
1267         unsigned long nrpages;
1268         int ret;
1269         int i;
1270
1271         /* the caller is responsible for locking the start index */
1272         if (index == locked_page->index && index == end_index)
1273                 return 0;
1274
1275         /* skip the page at the start index */
1276         nrpages = end_index - index + 1;
1277         while (nrpages > 0) {
1278                 ret = find_get_pages_contig(inode->i_mapping, index,
1279                                      min_t(unsigned long,
1280                                      nrpages, ARRAY_SIZE(pages)), pages);
1281                 if (ret == 0) {
1282                         ret = -EAGAIN;
1283                         goto done;
1284                 }
1285                 /* now we have an array of pages, lock them all */
1286                 for (i = 0; i < ret; i++) {
1287                         /*
1288                          * the caller is taking responsibility for
1289                          * locked_page
1290                          */
1291                         if (pages[i] != locked_page) {
1292                                 lock_page(pages[i]);
1293                                 if (!PageDirty(pages[i]) ||
1294                                     pages[i]->mapping != inode->i_mapping) {
1295                                         ret = -EAGAIN;
1296                                         unlock_page(pages[i]);
1297                                         page_cache_release(pages[i]);
1298                                         goto done;
1299                                 }
1300                         }
1301                         page_cache_release(pages[i]);
1302                         pages_locked++;
1303                 }
1304                 nrpages -= ret;
1305                 index += ret;
1306                 cond_resched();
1307         }
1308         ret = 0;
1309 done:
1310         if (ret && pages_locked) {
1311                 __unlock_for_delalloc(inode, locked_page,
1312                               delalloc_start,
1313                               ((u64)(start_index + pages_locked - 1)) <<
1314                               PAGE_CACHE_SHIFT);
1315         }
1316         return ret;
1317 }
1318
1319 /*
1320  * find a contiguous range of bytes in the file marked as delalloc, not
1321  * more than 'max_bytes'.  start and end are used to return the range,
1322  *
1323  * 1 is returned if we find something, 0 if nothing was in the tree
1324  */
1325 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1326                                              struct extent_io_tree *tree,
1327                                              struct page *locked_page,
1328                                              u64 *start, u64 *end,
1329                                              u64 max_bytes)
1330 {
1331         u64 delalloc_start;
1332         u64 delalloc_end;
1333         u64 found;
1334         struct extent_state *cached_state = NULL;
1335         int ret;
1336         int loops = 0;
1337
1338 again:
1339         /* step one, find a bunch of delalloc bytes starting at start */
1340         delalloc_start = *start;
1341         delalloc_end = 0;
1342         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1343                                     max_bytes, &cached_state);
1344         if (!found || delalloc_end <= *start) {
1345                 *start = delalloc_start;
1346                 *end = delalloc_end;
1347                 free_extent_state(cached_state);
1348                 return found;
1349         }
1350
1351         /*
1352          * start comes from the offset of locked_page.  We have to lock
1353          * pages in order, so we can't process delalloc bytes before
1354          * locked_page
1355          */
1356         if (delalloc_start < *start)
1357                 delalloc_start = *start;
1358
1359         /*
1360          * make sure to limit the number of pages we try to lock down
1361          * if we're looping.
1362          */
1363         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1364                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1365
1366         /* step two, lock all the pages after the page that has start */
1367         ret = lock_delalloc_pages(inode, locked_page,
1368                                   delalloc_start, delalloc_end);
1369         if (ret == -EAGAIN) {
1370                 /* some of the pages are gone, lets avoid looping by
1371                  * shortening the size of the delalloc range we're searching
1372                  */
1373                 free_extent_state(cached_state);
1374                 if (!loops) {
1375                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1376                         max_bytes = PAGE_CACHE_SIZE - offset;
1377                         loops = 1;
1378                         goto again;
1379                 } else {
1380                         found = 0;
1381                         goto out_failed;
1382                 }
1383         }
1384         BUG_ON(ret);
1385
1386         /* step three, lock the state bits for the whole range */
1387         lock_extent_bits(tree, delalloc_start, delalloc_end,
1388                          0, &cached_state, GFP_NOFS);
1389
1390         /* then test to make sure it is all still delalloc */
1391         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1392                              EXTENT_DELALLOC, 1, cached_state);
1393         if (!ret) {
1394                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1395                                      &cached_state, GFP_NOFS);
1396                 __unlock_for_delalloc(inode, locked_page,
1397                               delalloc_start, delalloc_end);
1398                 cond_resched();
1399                 goto again;
1400         }
1401         free_extent_state(cached_state);
1402         *start = delalloc_start;
1403         *end = delalloc_end;
1404 out_failed:
1405         return found;
1406 }
1407
1408 int extent_clear_unlock_delalloc(struct inode *inode,
1409                                 struct extent_io_tree *tree,
1410                                 u64 start, u64 end, struct page *locked_page,
1411                                 unsigned long op)
1412 {
1413         int ret;
1414         struct page *pages[16];
1415         unsigned long index = start >> PAGE_CACHE_SHIFT;
1416         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1417         unsigned long nr_pages = end_index - index + 1;
1418         int i;
1419         int clear_bits = 0;
1420
1421         if (op & EXTENT_CLEAR_UNLOCK)
1422                 clear_bits |= EXTENT_LOCKED;
1423         if (op & EXTENT_CLEAR_DIRTY)
1424                 clear_bits |= EXTENT_DIRTY;
1425
1426         if (op & EXTENT_CLEAR_DELALLOC)
1427                 clear_bits |= EXTENT_DELALLOC;
1428
1429         if (op & EXTENT_CLEAR_ACCOUNTING)
1430                 clear_bits |= EXTENT_DO_ACCOUNTING;
1431
1432         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1433         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1434                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1435                     EXTENT_SET_PRIVATE2)))
1436                 return 0;
1437
1438         while (nr_pages > 0) {
1439                 ret = find_get_pages_contig(inode->i_mapping, index,
1440                                      min_t(unsigned long,
1441                                      nr_pages, ARRAY_SIZE(pages)), pages);
1442                 for (i = 0; i < ret; i++) {
1443
1444                         if (op & EXTENT_SET_PRIVATE2)
1445                                 SetPagePrivate2(pages[i]);
1446
1447                         if (pages[i] == locked_page) {
1448                                 page_cache_release(pages[i]);
1449                                 continue;
1450                         }
1451                         if (op & EXTENT_CLEAR_DIRTY)
1452                                 clear_page_dirty_for_io(pages[i]);
1453                         if (op & EXTENT_SET_WRITEBACK)
1454                                 set_page_writeback(pages[i]);
1455                         if (op & EXTENT_END_WRITEBACK)
1456                                 end_page_writeback(pages[i]);
1457                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1458                                 unlock_page(pages[i]);
1459                         page_cache_release(pages[i]);
1460                 }
1461                 nr_pages -= ret;
1462                 index += ret;
1463                 cond_resched();
1464         }
1465         return 0;
1466 }
1467
1468 /*
1469  * count the number of bytes in the tree that have a given bit(s)
1470  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1471  * cached.  The total number found is returned.
1472  */
1473 u64 count_range_bits(struct extent_io_tree *tree,
1474                      u64 *start, u64 search_end, u64 max_bytes,
1475                      unsigned long bits)
1476 {
1477         struct rb_node *node;
1478         struct extent_state *state;
1479         u64 cur_start = *start;
1480         u64 total_bytes = 0;
1481         int found = 0;
1482
1483         if (search_end <= cur_start) {
1484                 WARN_ON(1);
1485                 return 0;
1486         }
1487
1488         spin_lock(&tree->lock);
1489         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1490                 total_bytes = tree->dirty_bytes;
1491                 goto out;
1492         }
1493         /*
1494          * this search will find all the extents that end after
1495          * our range starts.
1496          */
1497         node = tree_search(tree, cur_start);
1498         if (!node)
1499                 goto out;
1500
1501         while (1) {
1502                 state = rb_entry(node, struct extent_state, rb_node);
1503                 if (state->start > search_end)
1504                         break;
1505                 if (state->end >= cur_start && (state->state & bits)) {
1506                         total_bytes += min(search_end, state->end) + 1 -
1507                                        max(cur_start, state->start);
1508                         if (total_bytes >= max_bytes)
1509                                 break;
1510                         if (!found) {
1511                                 *start = state->start;
1512                                 found = 1;
1513                         }
1514                 }
1515                 node = rb_next(node);
1516                 if (!node)
1517                         break;
1518         }
1519 out:
1520         spin_unlock(&tree->lock);
1521         return total_bytes;
1522 }
1523
1524 /*
1525  * set the private field for a given byte offset in the tree.  If there isn't
1526  * an extent_state there already, this does nothing.
1527  */
1528 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1529 {
1530         struct rb_node *node;
1531         struct extent_state *state;
1532         int ret = 0;
1533
1534         spin_lock(&tree->lock);
1535         /*
1536          * this search will find all the extents that end after
1537          * our range starts.
1538          */
1539         node = tree_search(tree, start);
1540         if (!node) {
1541                 ret = -ENOENT;
1542                 goto out;
1543         }
1544         state = rb_entry(node, struct extent_state, rb_node);
1545         if (state->start != start) {
1546                 ret = -ENOENT;
1547                 goto out;
1548         }
1549         state->private = private;
1550 out:
1551         spin_unlock(&tree->lock);
1552         return ret;
1553 }
1554
1555 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1556 {
1557         struct rb_node *node;
1558         struct extent_state *state;
1559         int ret = 0;
1560
1561         spin_lock(&tree->lock);
1562         /*
1563          * this search will find all the extents that end after
1564          * our range starts.
1565          */
1566         node = tree_search(tree, start);
1567         if (!node) {
1568                 ret = -ENOENT;
1569                 goto out;
1570         }
1571         state = rb_entry(node, struct extent_state, rb_node);
1572         if (state->start != start) {
1573                 ret = -ENOENT;
1574                 goto out;
1575         }
1576         *private = state->private;
1577 out:
1578         spin_unlock(&tree->lock);
1579         return ret;
1580 }
1581
1582 /*
1583  * searches a range in the state tree for a given mask.
1584  * If 'filled' == 1, this returns 1 only if every extent in the tree
1585  * has the bits set.  Otherwise, 1 is returned if any bit in the
1586  * range is found set.
1587  */
1588 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1589                    int bits, int filled, struct extent_state *cached)
1590 {
1591         struct extent_state *state = NULL;
1592         struct rb_node *node;
1593         int bitset = 0;
1594
1595         spin_lock(&tree->lock);
1596         if (cached && cached->tree && cached->start == start)
1597                 node = &cached->rb_node;
1598         else
1599                 node = tree_search(tree, start);
1600         while (node && start <= end) {
1601                 state = rb_entry(node, struct extent_state, rb_node);
1602
1603                 if (filled && state->start > start) {
1604                         bitset = 0;
1605                         break;
1606                 }
1607
1608                 if (state->start > end)
1609                         break;
1610
1611                 if (state->state & bits) {
1612                         bitset = 1;
1613                         if (!filled)
1614                                 break;
1615                 } else if (filled) {
1616                         bitset = 0;
1617                         break;
1618                 }
1619
1620                 if (state->end == (u64)-1)
1621                         break;
1622
1623                 start = state->end + 1;
1624                 if (start > end)
1625                         break;
1626                 node = rb_next(node);
1627                 if (!node) {
1628                         if (filled)
1629                                 bitset = 0;
1630                         break;
1631                 }
1632         }
1633         spin_unlock(&tree->lock);
1634         return bitset;
1635 }
1636
1637 /*
1638  * helper function to set a given page up to date if all the
1639  * extents in the tree for that page are up to date
1640  */
1641 static int check_page_uptodate(struct extent_io_tree *tree,
1642                                struct page *page)
1643 {
1644         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1645         u64 end = start + PAGE_CACHE_SIZE - 1;
1646         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1647                 SetPageUptodate(page);
1648         return 0;
1649 }
1650
1651 /*
1652  * helper function to unlock a page if all the extents in the tree
1653  * for that page are unlocked
1654  */
1655 static int check_page_locked(struct extent_io_tree *tree,
1656                              struct page *page)
1657 {
1658         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1659         u64 end = start + PAGE_CACHE_SIZE - 1;
1660         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1661                 unlock_page(page);
1662         return 0;
1663 }
1664
1665 /*
1666  * helper function to end page writeback if all the extents
1667  * in the tree for that page are done with writeback
1668  */
1669 static int check_page_writeback(struct extent_io_tree *tree,
1670                              struct page *page)
1671 {
1672         end_page_writeback(page);
1673         return 0;
1674 }
1675
1676 /* lots and lots of room for performance fixes in the end_bio funcs */
1677
1678 /*
1679  * after a writepage IO is done, we need to:
1680  * clear the uptodate bits on error
1681  * clear the writeback bits in the extent tree for this IO
1682  * end_page_writeback if the page has no more pending IO
1683  *
1684  * Scheduling is not allowed, so the extent state tree is expected
1685  * to have one and only one object corresponding to this IO.
1686  */
1687 static void end_bio_extent_writepage(struct bio *bio, int err)
1688 {
1689         int uptodate = err == 0;
1690         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1691         struct extent_io_tree *tree;
1692         u64 start;
1693         u64 end;
1694         int whole_page;
1695         int ret;
1696
1697         do {
1698                 struct page *page = bvec->bv_page;
1699                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1700
1701                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1702                          bvec->bv_offset;
1703                 end = start + bvec->bv_len - 1;
1704
1705                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1706                         whole_page = 1;
1707                 else
1708                         whole_page = 0;
1709
1710                 if (--bvec >= bio->bi_io_vec)
1711                         prefetchw(&bvec->bv_page->flags);
1712                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1713                         ret = tree->ops->writepage_end_io_hook(page, start,
1714                                                        end, NULL, uptodate);
1715                         if (ret)
1716                                 uptodate = 0;
1717                 }
1718
1719                 if (!uptodate && tree->ops &&
1720                     tree->ops->writepage_io_failed_hook) {
1721                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1722                                                          start, end, NULL);
1723                         if (ret == 0) {
1724                                 uptodate = (err == 0);
1725                                 continue;
1726                         }
1727                 }
1728
1729                 if (!uptodate) {
1730                         clear_extent_uptodate(tree, start, end, GFP_NOFS);
1731                         ClearPageUptodate(page);
1732                         SetPageError(page);
1733                 }
1734
1735                 if (whole_page)
1736                         end_page_writeback(page);
1737                 else
1738                         check_page_writeback(tree, page);
1739         } while (bvec >= bio->bi_io_vec);
1740
1741         bio_put(bio);
1742 }
1743
1744 /*
1745  * after a readpage IO is done, we need to:
1746  * clear the uptodate bits on error
1747  * set the uptodate bits if things worked
1748  * set the page up to date if all extents in the tree are uptodate
1749  * clear the lock bit in the extent tree
1750  * unlock the page if there are no other extents locked for it
1751  *
1752  * Scheduling is not allowed, so the extent state tree is expected
1753  * to have one and only one object corresponding to this IO.
1754  */
1755 static void end_bio_extent_readpage(struct bio *bio, int err)
1756 {
1757         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1758         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
1759         struct bio_vec *bvec = bio->bi_io_vec;
1760         struct extent_io_tree *tree;
1761         u64 start;
1762         u64 end;
1763         int whole_page;
1764         int ret;
1765
1766         if (err)
1767                 uptodate = 0;
1768
1769         do {
1770                 struct page *page = bvec->bv_page;
1771                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1772
1773                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1774                         bvec->bv_offset;
1775                 end = start + bvec->bv_len - 1;
1776
1777                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1778                         whole_page = 1;
1779                 else
1780                         whole_page = 0;
1781
1782                 if (++bvec <= bvec_end)
1783                         prefetchw(&bvec->bv_page->flags);
1784
1785                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1786                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1787                                                               NULL);
1788                         if (ret)
1789                                 uptodate = 0;
1790                 }
1791                 if (!uptodate && tree->ops &&
1792                     tree->ops->readpage_io_failed_hook) {
1793                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1794                                                          start, end, NULL);
1795                         if (ret == 0) {
1796                                 uptodate =
1797                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1798                                 if (err)
1799                                         uptodate = 0;
1800                                 continue;
1801                         }
1802                 }
1803
1804                 if (uptodate) {
1805                         set_extent_uptodate(tree, start, end,
1806                                             GFP_ATOMIC);
1807                 }
1808                 unlock_extent(tree, start, end, GFP_ATOMIC);
1809
1810                 if (whole_page) {
1811                         if (uptodate) {
1812                                 SetPageUptodate(page);
1813                         } else {
1814                                 ClearPageUptodate(page);
1815                                 SetPageError(page);
1816                         }
1817                         unlock_page(page);
1818                 } else {
1819                         if (uptodate) {
1820                                 check_page_uptodate(tree, page);
1821                         } else {
1822                                 ClearPageUptodate(page);
1823                                 SetPageError(page);
1824                         }
1825                         check_page_locked(tree, page);
1826                 }
1827         } while (bvec <= bvec_end);
1828
1829         bio_put(bio);
1830 }
1831
1832 /*
1833  * IO done from prepare_write is pretty simple, we just unlock
1834  * the structs in the extent tree when done, and set the uptodate bits
1835  * as appropriate.
1836  */
1837 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1838 {
1839         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1840         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1841         struct extent_io_tree *tree;
1842         u64 start;
1843         u64 end;
1844
1845         do {
1846                 struct page *page = bvec->bv_page;
1847                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1848
1849                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1850                         bvec->bv_offset;
1851                 end = start + bvec->bv_len - 1;
1852
1853                 if (--bvec >= bio->bi_io_vec)
1854                         prefetchw(&bvec->bv_page->flags);
1855
1856                 if (uptodate) {
1857                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1858                 } else {
1859                         ClearPageUptodate(page);
1860                         SetPageError(page);
1861                 }
1862
1863                 unlock_extent(tree, start, end, GFP_ATOMIC);
1864
1865         } while (bvec >= bio->bi_io_vec);
1866
1867         bio_put(bio);
1868 }
1869
1870 static struct bio *
1871 extent_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1872                  gfp_t gfp_flags)
1873 {
1874         struct bio *bio;
1875
1876         bio = bio_alloc(gfp_flags, nr_vecs);
1877
1878         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1879                 while (!bio && (nr_vecs /= 2))
1880                         bio = bio_alloc(gfp_flags, nr_vecs);
1881         }
1882
1883         if (bio) {
1884                 bio->bi_size = 0;
1885                 bio->bi_bdev = bdev;
1886                 bio->bi_sector = first_sector;
1887         }
1888         return bio;
1889 }
1890
1891 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1892                           unsigned long bio_flags)
1893 {
1894         int ret = 0;
1895         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1896         struct page *page = bvec->bv_page;
1897         struct extent_io_tree *tree = bio->bi_private;
1898         u64 start;
1899         u64 end;
1900
1901         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1902         end = start + bvec->bv_len - 1;
1903
1904         bio->bi_private = NULL;
1905
1906         bio_get(bio);
1907
1908         if (tree->ops && tree->ops->submit_bio_hook)
1909                 tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1910                                            mirror_num, bio_flags);
1911         else
1912                 submit_bio(rw, bio);
1913         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1914                 ret = -EOPNOTSUPP;
1915         bio_put(bio);
1916         return ret;
1917 }
1918
1919 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1920                               struct page *page, sector_t sector,
1921                               size_t size, unsigned long offset,
1922                               struct block_device *bdev,
1923                               struct bio **bio_ret,
1924                               unsigned long max_pages,
1925                               bio_end_io_t end_io_func,
1926                               int mirror_num,
1927                               unsigned long prev_bio_flags,
1928                               unsigned long bio_flags)
1929 {
1930         int ret = 0;
1931         struct bio *bio;
1932         int nr;
1933         int contig = 0;
1934         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1935         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1936         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1937
1938         if (bio_ret && *bio_ret) {
1939                 bio = *bio_ret;
1940                 if (old_compressed)
1941                         contig = bio->bi_sector == sector;
1942                 else
1943                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
1944                                 sector;
1945
1946                 if (prev_bio_flags != bio_flags || !contig ||
1947                     (tree->ops && tree->ops->merge_bio_hook &&
1948                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
1949                                                bio_flags)) ||
1950                     bio_add_page(bio, page, page_size, offset) < page_size) {
1951                         ret = submit_one_bio(rw, bio, mirror_num,
1952                                              prev_bio_flags);
1953                         bio = NULL;
1954                 } else {
1955                         return 0;
1956                 }
1957         }
1958         if (this_compressed)
1959                 nr = BIO_MAX_PAGES;
1960         else
1961                 nr = bio_get_nr_vecs(bdev);
1962
1963         bio = extent_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1964
1965         bio_add_page(bio, page, page_size, offset);
1966         bio->bi_end_io = end_io_func;
1967         bio->bi_private = tree;
1968
1969         if (bio_ret)
1970                 *bio_ret = bio;
1971         else
1972                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1973
1974         return ret;
1975 }
1976
1977 void set_page_extent_mapped(struct page *page)
1978 {
1979         if (!PagePrivate(page)) {
1980                 SetPagePrivate(page);
1981                 page_cache_get(page);
1982                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1983         }
1984 }
1985
1986 static void set_page_extent_head(struct page *page, unsigned long len)
1987 {
1988         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1989 }
1990
1991 /*
1992  * basic readpage implementation.  Locked extent state structs are inserted
1993  * into the tree that are removed when the IO is done (by the end_io
1994  * handlers)
1995  */
1996 static int __extent_read_full_page(struct extent_io_tree *tree,
1997                                    struct page *page,
1998                                    get_extent_t *get_extent,
1999                                    struct bio **bio, int mirror_num,
2000                                    unsigned long *bio_flags)
2001 {
2002         struct inode *inode = page->mapping->host;
2003         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2004         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2005         u64 end;
2006         u64 cur = start;
2007         u64 extent_offset;
2008         u64 last_byte = i_size_read(inode);
2009         u64 block_start;
2010         u64 cur_end;
2011         sector_t sector;
2012         struct extent_map *em;
2013         struct block_device *bdev;
2014         int ret;
2015         int nr = 0;
2016         size_t page_offset = 0;
2017         size_t iosize;
2018         size_t disk_io_size;
2019         size_t blocksize = inode->i_sb->s_blocksize;
2020         unsigned long this_bio_flag = 0;
2021
2022         set_page_extent_mapped(page);
2023
2024         end = page_end;
2025         lock_extent(tree, start, end, GFP_NOFS);
2026
2027         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2028                 char *userpage;
2029                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2030
2031                 if (zero_offset) {
2032                         iosize = PAGE_CACHE_SIZE - zero_offset;
2033                         userpage = kmap_atomic(page, KM_USER0);
2034                         memset(userpage + zero_offset, 0, iosize);
2035                         flush_dcache_page(page);
2036                         kunmap_atomic(userpage, KM_USER0);
2037                 }
2038         }
2039         while (cur <= end) {
2040                 if (cur >= last_byte) {
2041                         char *userpage;
2042                         iosize = PAGE_CACHE_SIZE - page_offset;
2043                         userpage = kmap_atomic(page, KM_USER0);
2044                         memset(userpage + page_offset, 0, iosize);
2045                         flush_dcache_page(page);
2046                         kunmap_atomic(userpage, KM_USER0);
2047                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2048                                             GFP_NOFS);
2049                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2050                         break;
2051                 }
2052                 em = get_extent(inode, page, page_offset, cur,
2053                                 end - cur + 1, 0);
2054                 if (IS_ERR(em) || !em) {
2055                         SetPageError(page);
2056                         unlock_extent(tree, cur, end, GFP_NOFS);
2057                         break;
2058                 }
2059                 extent_offset = cur - em->start;
2060                 BUG_ON(extent_map_end(em) <= cur);
2061                 BUG_ON(end < cur);
2062
2063                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2064                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2065
2066                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2067                 cur_end = min(extent_map_end(em) - 1, end);
2068                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2069                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2070                         disk_io_size = em->block_len;
2071                         sector = em->block_start >> 9;
2072                 } else {
2073                         sector = (em->block_start + extent_offset) >> 9;
2074                         disk_io_size = iosize;
2075                 }
2076                 bdev = em->bdev;
2077                 block_start = em->block_start;
2078                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2079                         block_start = EXTENT_MAP_HOLE;
2080                 free_extent_map(em);
2081                 em = NULL;
2082
2083                 /* we've found a hole, just zero and go on */
2084                 if (block_start == EXTENT_MAP_HOLE) {
2085                         char *userpage;
2086                         userpage = kmap_atomic(page, KM_USER0);
2087                         memset(userpage + page_offset, 0, iosize);
2088                         flush_dcache_page(page);
2089                         kunmap_atomic(userpage, KM_USER0);
2090
2091                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2092                                             GFP_NOFS);
2093                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2094                         cur = cur + iosize;
2095                         page_offset += iosize;
2096                         continue;
2097                 }
2098                 /* the get_extent function already copied into the page */
2099                 if (test_range_bit(tree, cur, cur_end,
2100                                    EXTENT_UPTODATE, 1, NULL)) {
2101                         check_page_uptodate(tree, page);
2102                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2103                         cur = cur + iosize;
2104                         page_offset += iosize;
2105                         continue;
2106                 }
2107                 /* we have an inline extent but it didn't get marked up
2108                  * to date.  Error out
2109                  */
2110                 if (block_start == EXTENT_MAP_INLINE) {
2111                         SetPageError(page);
2112                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2113                         cur = cur + iosize;
2114                         page_offset += iosize;
2115                         continue;
2116                 }
2117
2118                 ret = 0;
2119                 if (tree->ops && tree->ops->readpage_io_hook) {
2120                         ret = tree->ops->readpage_io_hook(page, cur,
2121                                                           cur + iosize - 1);
2122                 }
2123                 if (!ret) {
2124                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2125                         pnr -= page->index;
2126                         ret = submit_extent_page(READ, tree, page,
2127                                          sector, disk_io_size, page_offset,
2128                                          bdev, bio, pnr,
2129                                          end_bio_extent_readpage, mirror_num,
2130                                          *bio_flags,
2131                                          this_bio_flag);
2132                         nr++;
2133                         *bio_flags = this_bio_flag;
2134                 }
2135                 if (ret)
2136                         SetPageError(page);
2137                 cur = cur + iosize;
2138                 page_offset += iosize;
2139         }
2140         if (!nr) {
2141                 if (!PageError(page))
2142                         SetPageUptodate(page);
2143                 unlock_page(page);
2144         }
2145         return 0;
2146 }
2147
2148 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2149                             get_extent_t *get_extent)
2150 {
2151         struct bio *bio = NULL;
2152         unsigned long bio_flags = 0;
2153         int ret;
2154
2155         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2156                                       &bio_flags);
2157         if (bio)
2158                 submit_one_bio(READ, bio, 0, bio_flags);
2159         return ret;
2160 }
2161
2162 static noinline void update_nr_written(struct page *page,
2163                                       struct writeback_control *wbc,
2164                                       unsigned long nr_written)
2165 {
2166         wbc->nr_to_write -= nr_written;
2167         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2168             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2169                 page->mapping->writeback_index = page->index + nr_written;
2170 }
2171
2172 /*
2173  * the writepage semantics are similar to regular writepage.  extent
2174  * records are inserted to lock ranges in the tree, and as dirty areas
2175  * are found, they are marked writeback.  Then the lock bits are removed
2176  * and the end_io handler clears the writeback ranges
2177  */
2178 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2179                               void *data)
2180 {
2181         struct inode *inode = page->mapping->host;
2182         struct extent_page_data *epd = data;
2183         struct extent_io_tree *tree = epd->tree;
2184         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2185         u64 delalloc_start;
2186         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2187         u64 end;
2188         u64 cur = start;
2189         u64 extent_offset;
2190         u64 last_byte = i_size_read(inode);
2191         u64 block_start;
2192         u64 iosize;
2193         u64 unlock_start;
2194         sector_t sector;
2195         struct extent_state *cached_state = NULL;
2196         struct extent_map *em;
2197         struct block_device *bdev;
2198         int ret;
2199         int nr = 0;
2200         size_t pg_offset = 0;
2201         size_t blocksize;
2202         loff_t i_size = i_size_read(inode);
2203         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2204         u64 nr_delalloc;
2205         u64 delalloc_end;
2206         int page_started;
2207         int compressed;
2208         int write_flags;
2209         unsigned long nr_written = 0;
2210
2211         if (wbc->sync_mode == WB_SYNC_ALL)
2212                 write_flags = WRITE_SYNC_PLUG;
2213         else
2214                 write_flags = WRITE;
2215
2216         WARN_ON(!PageLocked(page));
2217         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2218         if (page->index > end_index ||
2219            (page->index == end_index && !pg_offset)) {
2220                 page->mapping->a_ops->invalidatepage(page, 0);
2221                 unlock_page(page);
2222                 return 0;
2223         }
2224
2225         if (page->index == end_index) {
2226                 char *userpage;
2227
2228                 userpage = kmap_atomic(page, KM_USER0);
2229                 memset(userpage + pg_offset, 0,
2230                        PAGE_CACHE_SIZE - pg_offset);
2231                 kunmap_atomic(userpage, KM_USER0);
2232                 flush_dcache_page(page);
2233         }
2234         pg_offset = 0;
2235
2236         set_page_extent_mapped(page);
2237
2238         delalloc_start = start;
2239         delalloc_end = 0;
2240         page_started = 0;
2241         if (!epd->extent_locked) {
2242                 u64 delalloc_to_write = 0;
2243                 /*
2244                  * make sure the wbc mapping index is at least updated
2245                  * to this page.
2246                  */
2247                 update_nr_written(page, wbc, 0);
2248
2249                 while (delalloc_end < page_end) {
2250                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2251                                                        page,
2252                                                        &delalloc_start,
2253                                                        &delalloc_end,
2254                                                        128 * 1024 * 1024);
2255                         if (nr_delalloc == 0) {
2256                                 delalloc_start = delalloc_end + 1;
2257                                 continue;
2258                         }
2259                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2260                                                  delalloc_end, &page_started,
2261                                                  &nr_written);
2262                         /*
2263                          * delalloc_end is already one less than the total
2264                          * length, so we don't subtract one from
2265                          * PAGE_CACHE_SIZE
2266                          */
2267                         delalloc_to_write += (delalloc_end - delalloc_start +
2268                                               PAGE_CACHE_SIZE) >>
2269                                               PAGE_CACHE_SHIFT;
2270                         delalloc_start = delalloc_end + 1;
2271                 }
2272                 if (wbc->nr_to_write < delalloc_to_write) {
2273                         int thresh = 8192;
2274
2275                         if (delalloc_to_write < thresh * 2)
2276                                 thresh = delalloc_to_write;
2277                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2278                                                  thresh);
2279                 }
2280
2281                 /* did the fill delalloc function already unlock and start
2282                  * the IO?
2283                  */
2284                 if (page_started) {
2285                         ret = 0;
2286                         /*
2287                          * we've unlocked the page, so we can't update
2288                          * the mapping's writeback index, just update
2289                          * nr_to_write.
2290                          */
2291                         wbc->nr_to_write -= nr_written;
2292                         goto done_unlocked;
2293                 }
2294         }
2295         if (tree->ops && tree->ops->writepage_start_hook) {
2296                 ret = tree->ops->writepage_start_hook(page, start,
2297                                                       page_end);
2298                 if (ret == -EAGAIN) {
2299                         redirty_page_for_writepage(wbc, page);
2300                         update_nr_written(page, wbc, nr_written);
2301                         unlock_page(page);
2302                         ret = 0;
2303                         goto done_unlocked;
2304                 }
2305         }
2306
2307         /*
2308          * we don't want to touch the inode after unlocking the page,
2309          * so we update the mapping writeback index now
2310          */
2311         update_nr_written(page, wbc, nr_written + 1);
2312
2313         end = page_end;
2314         if (last_byte <= start) {
2315                 if (tree->ops && tree->ops->writepage_end_io_hook)
2316                         tree->ops->writepage_end_io_hook(page, start,
2317                                                          page_end, NULL, 1);
2318                 unlock_start = page_end + 1;
2319                 goto done;
2320         }
2321
2322         blocksize = inode->i_sb->s_blocksize;
2323
2324         while (cur <= end) {
2325                 if (cur >= last_byte) {
2326                         if (tree->ops && tree->ops->writepage_end_io_hook)
2327                                 tree->ops->writepage_end_io_hook(page, cur,
2328                                                          page_end, NULL, 1);
2329                         unlock_start = page_end + 1;
2330                         break;
2331                 }
2332                 em = epd->get_extent(inode, page, pg_offset, cur,
2333                                      end - cur + 1, 1);
2334                 if (IS_ERR(em) || !em) {
2335                         SetPageError(page);
2336                         break;
2337                 }
2338
2339                 extent_offset = cur - em->start;
2340                 BUG_ON(extent_map_end(em) <= cur);
2341                 BUG_ON(end < cur);
2342                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2343                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2344                 sector = (em->block_start + extent_offset) >> 9;
2345                 bdev = em->bdev;
2346                 block_start = em->block_start;
2347                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2348                 free_extent_map(em);
2349                 em = NULL;
2350
2351                 /*
2352                  * compressed and inline extents are written through other
2353                  * paths in the FS
2354                  */
2355                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2356                     block_start == EXTENT_MAP_INLINE) {
2357                         /*
2358                          * end_io notification does not happen here for
2359                          * compressed extents
2360                          */
2361                         if (!compressed && tree->ops &&
2362                             tree->ops->writepage_end_io_hook)
2363                                 tree->ops->writepage_end_io_hook(page, cur,
2364                                                          cur + iosize - 1,
2365                                                          NULL, 1);
2366                         else if (compressed) {
2367                                 /* we don't want to end_page_writeback on
2368                                  * a compressed extent.  this happens
2369                                  * elsewhere
2370                                  */
2371                                 nr++;
2372                         }
2373
2374                         cur += iosize;
2375                         pg_offset += iosize;
2376                         unlock_start = cur;
2377                         continue;
2378                 }
2379                 /* leave this out until we have a page_mkwrite call */
2380                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2381                                    EXTENT_DIRTY, 0, NULL)) {
2382                         cur = cur + iosize;
2383                         pg_offset += iosize;
2384                         continue;
2385                 }
2386
2387                 if (tree->ops && tree->ops->writepage_io_hook) {
2388                         ret = tree->ops->writepage_io_hook(page, cur,
2389                                                 cur + iosize - 1);
2390                 } else {
2391                         ret = 0;
2392                 }
2393                 if (ret) {
2394                         SetPageError(page);
2395                 } else {
2396                         unsigned long max_nr = end_index + 1;
2397
2398                         set_range_writeback(tree, cur, cur + iosize - 1);
2399                         if (!PageWriteback(page)) {
2400                                 printk(KERN_ERR "btrfs warning page %lu not "
2401                                        "writeback, cur %llu end %llu\n",
2402                                        page->index, (unsigned long long)cur,
2403                                        (unsigned long long)end);
2404                         }
2405
2406                         ret = submit_extent_page(write_flags, tree, page,
2407                                                  sector, iosize, pg_offset,
2408                                                  bdev, &epd->bio, max_nr,
2409                                                  end_bio_extent_writepage,
2410                                                  0, 0, 0);
2411                         if (ret)
2412                                 SetPageError(page);
2413                 }
2414                 cur = cur + iosize;
2415                 pg_offset += iosize;
2416                 nr++;
2417         }
2418 done:
2419         if (nr == 0) {
2420                 /* make sure the mapping tag for page dirty gets cleared */
2421                 set_page_writeback(page);
2422                 end_page_writeback(page);
2423         }
2424         unlock_page(page);
2425
2426 done_unlocked:
2427
2428         /* drop our reference on any cached states */
2429         free_extent_state(cached_state);
2430         return 0;
2431 }
2432
2433 /**
2434  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2435  * @mapping: address space structure to write
2436  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2437  * @writepage: function called for each page
2438  * @data: data passed to writepage function
2439  *
2440  * If a page is already under I/O, write_cache_pages() skips it, even
2441  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2442  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2443  * and msync() need to guarantee that all the data which was dirty at the time
2444  * the call was made get new I/O started against them.  If wbc->sync_mode is
2445  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2446  * existing IO to complete.
2447  */
2448 static int extent_write_cache_pages(struct extent_io_tree *tree,
2449                              struct address_space *mapping,
2450                              struct writeback_control *wbc,
2451                              writepage_t writepage, void *data,
2452                              void (*flush_fn)(void *))
2453 {
2454         int ret = 0;
2455         int done = 0;
2456         int nr_to_write_done = 0;
2457         struct pagevec pvec;
2458         int nr_pages;
2459         pgoff_t index;
2460         pgoff_t end;            /* Inclusive */
2461         int scanned = 0;
2462         int range_whole = 0;
2463
2464         pagevec_init(&pvec, 0);
2465         if (wbc->range_cyclic) {
2466                 index = mapping->writeback_index; /* Start from prev offset */
2467                 end = -1;
2468         } else {
2469                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2470                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2471                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2472                         range_whole = 1;
2473                 scanned = 1;
2474         }
2475 retry:
2476         while (!done && !nr_to_write_done && (index <= end) &&
2477                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2478                               PAGECACHE_TAG_DIRTY, min(end - index,
2479                                   (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2480                 unsigned i;
2481
2482                 scanned = 1;
2483                 for (i = 0; i < nr_pages; i++) {
2484                         struct page *page = pvec.pages[i];
2485
2486                         /*
2487                          * At this point we hold neither mapping->tree_lock nor
2488                          * lock on the page itself: the page may be truncated or
2489                          * invalidated (changing page->mapping to NULL), or even
2490                          * swizzled back from swapper_space to tmpfs file
2491                          * mapping
2492                          */
2493                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2494                                 tree->ops->write_cache_pages_lock_hook(page);
2495                         else
2496                                 lock_page(page);
2497
2498                         if (unlikely(page->mapping != mapping)) {
2499                                 unlock_page(page);
2500                                 continue;
2501                         }
2502
2503                         if (!wbc->range_cyclic && page->index > end) {
2504                                 done = 1;
2505                                 unlock_page(page);
2506                                 continue;
2507                         }
2508
2509                         if (wbc->sync_mode != WB_SYNC_NONE) {
2510                                 if (PageWriteback(page))
2511                                         flush_fn(data);
2512                                 wait_on_page_writeback(page);
2513                         }
2514
2515                         if (PageWriteback(page) ||
2516                             !clear_page_dirty_for_io(page)) {
2517                                 unlock_page(page);
2518                                 continue;
2519                         }
2520
2521                         ret = (*writepage)(page, wbc, data);
2522
2523                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2524                                 unlock_page(page);
2525                                 ret = 0;
2526                         }
2527                         if (ret)
2528                                 done = 1;
2529
2530                         /*
2531                          * the filesystem may choose to bump up nr_to_write.
2532                          * We have to make sure to honor the new nr_to_write
2533                          * at any time
2534                          */
2535                         nr_to_write_done = wbc->nr_to_write <= 0;
2536                 }
2537                 pagevec_release(&pvec);
2538                 cond_resched();
2539         }
2540         if (!scanned && !done) {
2541                 /*
2542                  * We hit the last page and there is more work to be done: wrap
2543                  * back to the start of the file
2544                  */
2545                 scanned = 1;
2546                 index = 0;
2547                 goto retry;
2548         }
2549         return ret;
2550 }
2551
2552 static void flush_epd_write_bio(struct extent_page_data *epd)
2553 {
2554         if (epd->bio) {
2555                 if (epd->sync_io)
2556                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2557                 else
2558                         submit_one_bio(WRITE, epd->bio, 0, 0);
2559                 epd->bio = NULL;
2560         }
2561 }
2562
2563 static noinline void flush_write_bio(void *data)
2564 {
2565         struct extent_page_data *epd = data;
2566         flush_epd_write_bio(epd);
2567 }
2568
2569 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2570                           get_extent_t *get_extent,
2571                           struct writeback_control *wbc)
2572 {
2573         int ret;
2574         struct address_space *mapping = page->mapping;
2575         struct extent_page_data epd = {
2576                 .bio = NULL,
2577                 .tree = tree,
2578                 .get_extent = get_extent,
2579                 .extent_locked = 0,
2580                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2581         };
2582         struct writeback_control wbc_writepages = {
2583                 .bdi            = wbc->bdi,
2584                 .sync_mode      = wbc->sync_mode,
2585                 .older_than_this = NULL,
2586                 .nr_to_write    = 64,
2587                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2588                 .range_end      = (loff_t)-1,
2589         };
2590
2591         ret = __extent_writepage(page, wbc, &epd);
2592
2593         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2594                                  __extent_writepage, &epd, flush_write_bio);
2595         flush_epd_write_bio(&epd);
2596         return ret;
2597 }
2598
2599 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2600                               u64 start, u64 end, get_extent_t *get_extent,
2601                               int mode)
2602 {
2603         int ret = 0;
2604         struct address_space *mapping = inode->i_mapping;
2605         struct page *page;
2606         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2607                 PAGE_CACHE_SHIFT;
2608
2609         struct extent_page_data epd = {
2610                 .bio = NULL,
2611                 .tree = tree,
2612                 .get_extent = get_extent,
2613                 .extent_locked = 1,
2614                 .sync_io = mode == WB_SYNC_ALL,
2615         };
2616         struct writeback_control wbc_writepages = {
2617                 .bdi            = inode->i_mapping->backing_dev_info,
2618                 .sync_mode      = mode,
2619                 .older_than_this = NULL,
2620                 .nr_to_write    = nr_pages * 2,
2621                 .range_start    = start,
2622                 .range_end      = end + 1,
2623         };
2624
2625         while (start <= end) {
2626                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2627                 if (clear_page_dirty_for_io(page))
2628                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2629                 else {
2630                         if (tree->ops && tree->ops->writepage_end_io_hook)
2631                                 tree->ops->writepage_end_io_hook(page, start,
2632                                                  start + PAGE_CACHE_SIZE - 1,
2633                                                  NULL, 1);
2634                         unlock_page(page);
2635                 }
2636                 page_cache_release(page);
2637                 start += PAGE_CACHE_SIZE;
2638         }
2639
2640         flush_epd_write_bio(&epd);
2641         return ret;
2642 }
2643
2644 int extent_writepages(struct extent_io_tree *tree,
2645                       struct address_space *mapping,
2646                       get_extent_t *get_extent,
2647                       struct writeback_control *wbc)
2648 {
2649         int ret = 0;
2650         struct extent_page_data epd = {
2651                 .bio = NULL,
2652                 .tree = tree,
2653                 .get_extent = get_extent,
2654                 .extent_locked = 0,
2655                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2656         };
2657
2658         ret = extent_write_cache_pages(tree, mapping, wbc,
2659                                        __extent_writepage, &epd,
2660                                        flush_write_bio);
2661         flush_epd_write_bio(&epd);
2662         return ret;
2663 }
2664
2665 int extent_readpages(struct extent_io_tree *tree,
2666                      struct address_space *mapping,
2667                      struct list_head *pages, unsigned nr_pages,
2668                      get_extent_t get_extent)
2669 {
2670         struct bio *bio = NULL;
2671         unsigned page_idx;
2672         struct pagevec pvec;
2673         unsigned long bio_flags = 0;
2674
2675         pagevec_init(&pvec, 0);
2676         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2677                 struct page *page = list_entry(pages->prev, struct page, lru);
2678
2679                 prefetchw(&page->flags);
2680                 list_del(&page->lru);
2681                 /*
2682                  * what we want to do here is call add_to_page_cache_lru,
2683                  * but that isn't exported, so we reproduce it here
2684                  */
2685                 if (!add_to_page_cache(page, mapping,
2686                                         page->index, GFP_KERNEL)) {
2687
2688                         /* open coding of lru_cache_add, also not exported */
2689                         page_cache_get(page);
2690                         if (!pagevec_add(&pvec, page))
2691                                 __pagevec_lru_add_file(&pvec);
2692                         __extent_read_full_page(tree, page, get_extent,
2693                                                 &bio, 0, &bio_flags);
2694                 }
2695                 page_cache_release(page);
2696         }
2697         if (pagevec_count(&pvec))
2698                 __pagevec_lru_add_file(&pvec);
2699         BUG_ON(!list_empty(pages));
2700         if (bio)
2701                 submit_one_bio(READ, bio, 0, bio_flags);
2702         return 0;
2703 }
2704
2705 /*
2706  * basic invalidatepage code, this waits on any locked or writeback
2707  * ranges corresponding to the page, and then deletes any extent state
2708  * records from the tree
2709  */
2710 int extent_invalidatepage(struct extent_io_tree *tree,
2711                           struct page *page, unsigned long offset)
2712 {
2713         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2714         u64 end = start + PAGE_CACHE_SIZE - 1;
2715         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2716
2717         start += (offset + blocksize - 1) & ~(blocksize - 1);
2718         if (start > end)
2719                 return 0;
2720
2721         lock_extent(tree, start, end, GFP_NOFS);
2722         wait_on_page_writeback(page);
2723         clear_extent_bit(tree, start, end,
2724                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2725                          EXTENT_DO_ACCOUNTING,
2726                          1, 1, NULL, GFP_NOFS);
2727         return 0;
2728 }
2729
2730 /*
2731  * simple commit_write call, set_range_dirty is used to mark both
2732  * the pages and the extent records as dirty
2733  */
2734 int extent_commit_write(struct extent_io_tree *tree,
2735                         struct inode *inode, struct page *page,
2736                         unsigned from, unsigned to)
2737 {
2738         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2739
2740         set_page_extent_mapped(page);
2741         set_page_dirty(page);
2742
2743         if (pos > inode->i_size) {
2744                 i_size_write(inode, pos);
2745                 mark_inode_dirty(inode);
2746         }
2747         return 0;
2748 }
2749
2750 int extent_prepare_write(struct extent_io_tree *tree,
2751                          struct inode *inode, struct page *page,
2752                          unsigned from, unsigned to, get_extent_t *get_extent)
2753 {
2754         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2755         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2756         u64 block_start;
2757         u64 orig_block_start;
2758         u64 block_end;
2759         u64 cur_end;
2760         struct extent_map *em;
2761         unsigned blocksize = 1 << inode->i_blkbits;
2762         size_t page_offset = 0;
2763         size_t block_off_start;
2764         size_t block_off_end;
2765         int err = 0;
2766         int iocount = 0;
2767         int ret = 0;
2768         int isnew;
2769
2770         set_page_extent_mapped(page);
2771
2772         block_start = (page_start + from) & ~((u64)blocksize - 1);
2773         block_end = (page_start + to - 1) | (blocksize - 1);
2774         orig_block_start = block_start;
2775
2776         lock_extent(tree, page_start, page_end, GFP_NOFS);
2777         while (block_start <= block_end) {
2778                 em = get_extent(inode, page, page_offset, block_start,
2779                                 block_end - block_start + 1, 1);
2780                 if (IS_ERR(em) || !em)
2781                         goto err;
2782
2783                 cur_end = min(block_end, extent_map_end(em) - 1);
2784                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2785                 block_off_end = block_off_start + blocksize;
2786                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2787
2788                 if (!PageUptodate(page) && isnew &&
2789                     (block_off_end > to || block_off_start < from)) {
2790                         void *kaddr;
2791
2792                         kaddr = kmap_atomic(page, KM_USER0);
2793                         if (block_off_end > to)
2794                                 memset(kaddr + to, 0, block_off_end - to);
2795                         if (block_off_start < from)
2796                                 memset(kaddr + block_off_start, 0,
2797                                        from - block_off_start);
2798                         flush_dcache_page(page);
2799                         kunmap_atomic(kaddr, KM_USER0);
2800                 }
2801                 if ((em->block_start != EXTENT_MAP_HOLE &&
2802                      em->block_start != EXTENT_MAP_INLINE) &&
2803                     !isnew && !PageUptodate(page) &&
2804                     (block_off_end > to || block_off_start < from) &&
2805                     !test_range_bit(tree, block_start, cur_end,
2806                                     EXTENT_UPTODATE, 1, NULL)) {
2807                         u64 sector;
2808                         u64 extent_offset = block_start - em->start;
2809                         size_t iosize;
2810                         sector = (em->block_start + extent_offset) >> 9;
2811                         iosize = (cur_end - block_start + blocksize) &
2812                                 ~((u64)blocksize - 1);
2813                         /*
2814                          * we've already got the extent locked, but we
2815                          * need to split the state such that our end_bio
2816                          * handler can clear the lock.
2817                          */
2818                         set_extent_bit(tree, block_start,
2819                                        block_start + iosize - 1,
2820                                        EXTENT_LOCKED, 0, NULL, NULL, GFP_NOFS);
2821                         ret = submit_extent_page(READ, tree, page,
2822                                          sector, iosize, page_offset, em->bdev,
2823                                          NULL, 1,
2824                                          end_bio_extent_preparewrite, 0,
2825                                          0, 0);
2826                         iocount++;
2827                         block_start = block_start + iosize;
2828                 } else {
2829                         set_extent_uptodate(tree, block_start, cur_end,
2830                                             GFP_NOFS);
2831                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2832                         block_start = cur_end + 1;
2833                 }
2834                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2835                 free_extent_map(em);
2836         }
2837         if (iocount) {
2838                 wait_extent_bit(tree, orig_block_start,
2839                                 block_end, EXTENT_LOCKED);
2840         }
2841         check_page_uptodate(tree, page);
2842 err:
2843         /* FIXME, zero out newly allocated blocks on error */
2844         return err;
2845 }
2846
2847 /*
2848  * a helper for releasepage, this tests for areas of the page that
2849  * are locked or under IO and drops the related state bits if it is safe
2850  * to drop the page.
2851  */
2852 int try_release_extent_state(struct extent_map_tree *map,
2853                              struct extent_io_tree *tree, struct page *page,
2854                              gfp_t mask)
2855 {
2856         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2857         u64 end = start + PAGE_CACHE_SIZE - 1;
2858         int ret = 1;
2859
2860         if (test_range_bit(tree, start, end,
2861                            EXTENT_IOBITS, 0, NULL))
2862                 ret = 0;
2863         else {
2864                 if ((mask & GFP_NOFS) == GFP_NOFS)
2865                         mask = GFP_NOFS;
2866                 /*
2867                  * at this point we can safely clear everything except the
2868                  * locked bit and the nodatasum bit
2869                  */
2870                 clear_extent_bit(tree, start, end,
2871                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2872                                  0, 0, NULL, mask);
2873         }
2874         return ret;
2875 }
2876
2877 /*
2878  * a helper for releasepage.  As long as there are no locked extents
2879  * in the range corresponding to the page, both state records and extent
2880  * map records are removed
2881  */
2882 int try_release_extent_mapping(struct extent_map_tree *map,
2883                                struct extent_io_tree *tree, struct page *page,
2884                                gfp_t mask)
2885 {
2886         struct extent_map *em;
2887         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2888         u64 end = start + PAGE_CACHE_SIZE - 1;
2889
2890         if ((mask & __GFP_WAIT) &&
2891             page->mapping->host->i_size > 16 * 1024 * 1024) {
2892                 u64 len;
2893                 while (start <= end) {
2894                         len = end - start + 1;
2895                         write_lock(&map->lock);
2896                         em = lookup_extent_mapping(map, start, len);
2897                         if (!em || IS_ERR(em)) {
2898                                 write_unlock(&map->lock);
2899                                 break;
2900                         }
2901                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2902                             em->start != start) {
2903                                 write_unlock(&map->lock);
2904                                 free_extent_map(em);
2905                                 break;
2906                         }
2907                         if (!test_range_bit(tree, em->start,
2908                                             extent_map_end(em) - 1,
2909                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
2910                                             0, NULL)) {
2911                                 remove_extent_mapping(map, em);
2912                                 /* once for the rb tree */
2913                                 free_extent_map(em);
2914                         }
2915                         start = extent_map_end(em);
2916                         write_unlock(&map->lock);
2917
2918                         /* once for us */
2919                         free_extent_map(em);
2920                 }
2921         }
2922         return try_release_extent_state(map, tree, page, mask);
2923 }
2924
2925 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2926                 get_extent_t *get_extent)
2927 {
2928         struct inode *inode = mapping->host;
2929         u64 start = iblock << inode->i_blkbits;
2930         sector_t sector = 0;
2931         size_t blksize = (1 << inode->i_blkbits);
2932         struct extent_map *em;
2933
2934         lock_extent(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2935                     GFP_NOFS);
2936         em = get_extent(inode, NULL, 0, start, blksize, 0);
2937         unlock_extent(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2938                       GFP_NOFS);
2939         if (!em || IS_ERR(em))
2940                 return 0;
2941
2942         if (em->block_start > EXTENT_MAP_LAST_BYTE)
2943                 goto out;
2944
2945         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2946 out:
2947         free_extent_map(em);
2948         return sector;
2949 }
2950
2951 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2952                 __u64 start, __u64 len, get_extent_t *get_extent)
2953 {
2954         int ret;
2955         u64 off = start;
2956         u64 max = start + len;
2957         u32 flags = 0;
2958         u64 disko = 0;
2959         struct extent_map *em = NULL;
2960         int end = 0;
2961         u64 em_start = 0, em_len = 0;
2962         unsigned long emflags;
2963         ret = 0;
2964
2965         if (len == 0)
2966                 return -EINVAL;
2967
2968         lock_extent(&BTRFS_I(inode)->io_tree, start, start + len,
2969                 GFP_NOFS);
2970         em = get_extent(inode, NULL, 0, off, max - off, 0);
2971         if (!em)
2972                 goto out;
2973         if (IS_ERR(em)) {
2974                 ret = PTR_ERR(em);
2975                 goto out;
2976         }
2977         while (!end) {
2978                 off = em->start + em->len;
2979                 if (off >= max)
2980                         end = 1;
2981
2982                 em_start = em->start;
2983                 em_len = em->len;
2984
2985                 disko = 0;
2986                 flags = 0;
2987
2988                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
2989                         end = 1;
2990                         flags |= FIEMAP_EXTENT_LAST;
2991                 } else if (em->block_start == EXTENT_MAP_HOLE) {
2992                         flags |= FIEMAP_EXTENT_UNWRITTEN;
2993                 } else if (em->block_start == EXTENT_MAP_INLINE) {
2994                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
2995                                   FIEMAP_EXTENT_NOT_ALIGNED);
2996                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
2997                         flags |= (FIEMAP_EXTENT_DELALLOC |
2998                                   FIEMAP_EXTENT_UNKNOWN);
2999                 } else {
3000                         disko = em->block_start;
3001                 }
3002                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3003                         flags |= FIEMAP_EXTENT_ENCODED;
3004
3005                 emflags = em->flags;
3006                 free_extent_map(em);
3007                 em = NULL;
3008
3009                 if (!end) {
3010                         em = get_extent(inode, NULL, 0, off, max - off, 0);
3011                         if (!em)
3012                                 goto out;
3013                         if (IS_ERR(em)) {
3014                                 ret = PTR_ERR(em);
3015                                 goto out;
3016                         }
3017                         emflags = em->flags;
3018                 }
3019                 if (test_bit(EXTENT_FLAG_VACANCY, &emflags)) {
3020                         flags |= FIEMAP_EXTENT_LAST;
3021                         end = 1;
3022                 }
3023
3024                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3025                                         em_len, flags);
3026                 if (ret)
3027                         goto out_free;
3028         }
3029 out_free:
3030         free_extent_map(em);
3031 out:
3032         unlock_extent(&BTRFS_I(inode)->io_tree, start, start + len,
3033                         GFP_NOFS);
3034         return ret;
3035 }
3036
3037 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
3038                                               unsigned long i)
3039 {
3040         struct page *p;
3041         struct address_space *mapping;
3042
3043         if (i == 0)
3044                 return eb->first_page;
3045         i += eb->start >> PAGE_CACHE_SHIFT;
3046         mapping = eb->first_page->mapping;
3047         if (!mapping)
3048                 return NULL;
3049
3050         /*
3051          * extent_buffer_page is only called after pinning the page
3052          * by increasing the reference count.  So we know the page must
3053          * be in the radix tree.
3054          */
3055         rcu_read_lock();
3056         p = radix_tree_lookup(&mapping->page_tree, i);
3057         rcu_read_unlock();
3058
3059         return p;
3060 }
3061
3062 static inline unsigned long num_extent_pages(u64 start, u64 len)
3063 {
3064         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3065                 (start >> PAGE_CACHE_SHIFT);
3066 }
3067
3068 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3069                                                    u64 start,
3070                                                    unsigned long len,
3071                                                    gfp_t mask)
3072 {
3073         struct extent_buffer *eb = NULL;
3074 #if LEAK_DEBUG
3075         unsigned long flags;
3076 #endif
3077
3078         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3079         eb->start = start;
3080         eb->len = len;
3081         spin_lock_init(&eb->lock);
3082         init_waitqueue_head(&eb->lock_wq);
3083
3084 #if LEAK_DEBUG
3085         spin_lock_irqsave(&leak_lock, flags);
3086         list_add(&eb->leak_list, &buffers);
3087         spin_unlock_irqrestore(&leak_lock, flags);
3088 #endif
3089         atomic_set(&eb->refs, 1);
3090
3091         return eb;
3092 }
3093
3094 static void __free_extent_buffer(struct extent_buffer *eb)
3095 {
3096 #if LEAK_DEBUG
3097         unsigned long flags;
3098         spin_lock_irqsave(&leak_lock, flags);
3099         list_del(&eb->leak_list);
3100         spin_unlock_irqrestore(&leak_lock, flags);
3101 #endif
3102         kmem_cache_free(extent_buffer_cache, eb);
3103 }
3104
3105 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3106                                           u64 start, unsigned long len,
3107                                           struct page *page0,
3108                                           gfp_t mask)
3109 {
3110         unsigned long num_pages = num_extent_pages(start, len);
3111         unsigned long i;
3112         unsigned long index = start >> PAGE_CACHE_SHIFT;
3113         struct extent_buffer *eb;
3114         struct extent_buffer *exists = NULL;
3115         struct page *p;
3116         struct address_space *mapping = tree->mapping;
3117         int uptodate = 1;
3118
3119         spin_lock(&tree->buffer_lock);
3120         eb = buffer_search(tree, start);
3121         if (eb) {
3122                 atomic_inc(&eb->refs);
3123                 spin_unlock(&tree->buffer_lock);
3124                 mark_page_accessed(eb->first_page);
3125                 return eb;
3126         }
3127         spin_unlock(&tree->buffer_lock);
3128
3129         eb = __alloc_extent_buffer(tree, start, len, mask);
3130         if (!eb)
3131                 return NULL;
3132
3133         if (page0) {
3134                 eb->first_page = page0;
3135                 i = 1;
3136                 index++;
3137                 page_cache_get(page0);
3138                 mark_page_accessed(page0);
3139                 set_page_extent_mapped(page0);
3140                 set_page_extent_head(page0, len);
3141                 uptodate = PageUptodate(page0);
3142         } else {
3143                 i = 0;
3144         }
3145         for (; i < num_pages; i++, index++) {
3146                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
3147                 if (!p) {
3148                         WARN_ON(1);
3149                         goto free_eb;
3150                 }
3151                 set_page_extent_mapped(p);
3152                 mark_page_accessed(p);
3153                 if (i == 0) {
3154                         eb->first_page = p;
3155                         set_page_extent_head(p, len);
3156                 } else {
3157                         set_page_private(p, EXTENT_PAGE_PRIVATE);
3158                 }
3159                 if (!PageUptodate(p))
3160                         uptodate = 0;
3161                 unlock_page(p);
3162         }
3163         if (uptodate)
3164                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3165
3166         spin_lock(&tree->buffer_lock);
3167         exists = buffer_tree_insert(tree, start, &eb->rb_node);
3168         if (exists) {
3169                 /* add one reference for the caller */
3170                 atomic_inc(&exists->refs);
3171                 spin_unlock(&tree->buffer_lock);
3172                 goto free_eb;
3173         }
3174         /* add one reference for the tree */
3175         atomic_inc(&eb->refs);
3176         spin_unlock(&tree->buffer_lock);
3177         return eb;
3178
3179 free_eb:
3180         if (!atomic_dec_and_test(&eb->refs))
3181                 return exists;
3182         for (index = 1; index < i; index++)
3183                 page_cache_release(extent_buffer_page(eb, index));
3184         page_cache_release(extent_buffer_page(eb, 0));
3185         __free_extent_buffer(eb);
3186         return exists;
3187 }
3188
3189 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3190                                          u64 start, unsigned long len,
3191                                           gfp_t mask)
3192 {
3193         struct extent_buffer *eb;
3194
3195         spin_lock(&tree->buffer_lock);
3196         eb = buffer_search(tree, start);
3197         if (eb)
3198                 atomic_inc(&eb->refs);
3199         spin_unlock(&tree->buffer_lock);
3200
3201         if (eb)
3202                 mark_page_accessed(eb->first_page);
3203
3204         return eb;
3205 }
3206
3207 void free_extent_buffer(struct extent_buffer *eb)
3208 {
3209         if (!eb)
3210                 return;
3211
3212         if (!atomic_dec_and_test(&eb->refs))
3213                 return;
3214
3215         WARN_ON(1);
3216 }
3217
3218 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3219                               struct extent_buffer *eb)
3220 {
3221         unsigned long i;
3222         unsigned long num_pages;
3223         struct page *page;
3224
3225         num_pages = num_extent_pages(eb->start, eb->len);
3226
3227         for (i = 0; i < num_pages; i++) {
3228                 page = extent_buffer_page(eb, i);
3229                 if (!PageDirty(page))
3230                         continue;
3231
3232                 lock_page(page);
3233                 if (i == 0)
3234                         set_page_extent_head(page, eb->len);
3235                 else
3236                         set_page_private(page, EXTENT_PAGE_PRIVATE);
3237
3238                 clear_page_dirty_for_io(page);
3239                 spin_lock_irq(&page->mapping->tree_lock);
3240                 if (!PageDirty(page)) {
3241                         radix_tree_tag_clear(&page->mapping->page_tree,
3242                                                 page_index(page),
3243                                                 PAGECACHE_TAG_DIRTY);
3244                 }
3245                 spin_unlock_irq(&page->mapping->tree_lock);
3246                 unlock_page(page);
3247         }
3248         return 0;
3249 }
3250
3251 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
3252                                     struct extent_buffer *eb)
3253 {
3254         return wait_on_extent_writeback(tree, eb->start,
3255                                         eb->start + eb->len - 1);
3256 }
3257
3258 int set_extent_buffer_dirty(struct extent_io_tree *tree,
3259                              struct extent_buffer *eb)
3260 {
3261         unsigned long i;
3262         unsigned long num_pages;
3263         int was_dirty = 0;
3264
3265         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3266         num_pages = num_extent_pages(eb->start, eb->len);
3267         for (i = 0; i < num_pages; i++)
3268                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3269         return was_dirty;
3270 }
3271
3272 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3273                                 struct extent_buffer *eb)
3274 {
3275         unsigned long i;
3276         struct page *page;
3277         unsigned long num_pages;
3278
3279         num_pages = num_extent_pages(eb->start, eb->len);
3280         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3281
3282         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3283                               GFP_NOFS);
3284         for (i = 0; i < num_pages; i++) {
3285                 page = extent_buffer_page(eb, i);
3286                 if (page)
3287                         ClearPageUptodate(page);
3288         }
3289         return 0;
3290 }
3291
3292 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3293                                 struct extent_buffer *eb)
3294 {
3295         unsigned long i;
3296         struct page *page;
3297         unsigned long num_pages;
3298
3299         num_pages = num_extent_pages(eb->start, eb->len);
3300
3301         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3302                             GFP_NOFS);
3303         for (i = 0; i < num_pages; i++) {
3304                 page = extent_buffer_page(eb, i);
3305                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3306                     ((i == num_pages - 1) &&
3307                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3308                         check_page_uptodate(tree, page);
3309                         continue;
3310                 }
3311                 SetPageUptodate(page);
3312         }
3313         return 0;
3314 }
3315
3316 int extent_range_uptodate(struct extent_io_tree *tree,
3317                           u64 start, u64 end)
3318 {
3319         struct page *page;
3320         int ret;
3321         int pg_uptodate = 1;
3322         int uptodate;
3323         unsigned long index;
3324
3325         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL);
3326         if (ret)
3327                 return 1;
3328         while (start <= end) {
3329                 index = start >> PAGE_CACHE_SHIFT;
3330                 page = find_get_page(tree->mapping, index);
3331                 uptodate = PageUptodate(page);
3332                 page_cache_release(page);
3333                 if (!uptodate) {
3334                         pg_uptodate = 0;
3335                         break;
3336                 }
3337                 start += PAGE_CACHE_SIZE;
3338         }
3339         return pg_uptodate;
3340 }
3341
3342 int extent_buffer_uptodate(struct extent_io_tree *tree,
3343                            struct extent_buffer *eb)
3344 {
3345         int ret = 0;
3346         unsigned long num_pages;
3347         unsigned long i;
3348         struct page *page;
3349         int pg_uptodate = 1;
3350
3351         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3352                 return 1;
3353
3354         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3355                            EXTENT_UPTODATE, 1, NULL);
3356         if (ret)
3357                 return ret;
3358
3359         num_pages = num_extent_pages(eb->start, eb->len);
3360         for (i = 0; i < num_pages; i++) {
3361                 page = extent_buffer_page(eb, i);
3362                 if (!PageUptodate(page)) {
3363                         pg_uptodate = 0;
3364                         break;
3365                 }
3366         }
3367         return pg_uptodate;
3368 }
3369
3370 int read_extent_buffer_pages(struct extent_io_tree *tree,
3371                              struct extent_buffer *eb,
3372                              u64 start, int wait,
3373                              get_extent_t *get_extent, int mirror_num)
3374 {
3375         unsigned long i;
3376         unsigned long start_i;
3377         struct page *page;
3378         int err;
3379         int ret = 0;
3380         int locked_pages = 0;
3381         int all_uptodate = 1;
3382         int inc_all_pages = 0;
3383         unsigned long num_pages;
3384         struct bio *bio = NULL;
3385         unsigned long bio_flags = 0;
3386
3387         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3388                 return 0;
3389
3390         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3391                            EXTENT_UPTODATE, 1, NULL)) {
3392                 return 0;
3393         }
3394
3395         if (start) {
3396                 WARN_ON(start < eb->start);
3397                 start_i = (start >> PAGE_CACHE_SHIFT) -
3398                         (eb->start >> PAGE_CACHE_SHIFT);
3399         } else {
3400                 start_i = 0;
3401         }
3402
3403         num_pages = num_extent_pages(eb->start, eb->len);
3404         for (i = start_i; i < num_pages; i++) {
3405                 page = extent_buffer_page(eb, i);
3406                 if (!wait) {
3407                         if (!trylock_page(page))
3408                                 goto unlock_exit;
3409                 } else {
3410                         lock_page(page);
3411                 }
3412                 locked_pages++;
3413                 if (!PageUptodate(page))
3414                         all_uptodate = 0;
3415         }
3416         if (all_uptodate) {
3417                 if (start_i == 0)
3418                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3419                 goto unlock_exit;
3420         }
3421
3422         for (i = start_i; i < num_pages; i++) {
3423                 page = extent_buffer_page(eb, i);
3424                 if (inc_all_pages)
3425                         page_cache_get(page);
3426                 if (!PageUptodate(page)) {
3427                         if (start_i == 0)
3428                                 inc_all_pages = 1;
3429                         ClearPageError(page);
3430                         err = __extent_read_full_page(tree, page,
3431                                                       get_extent, &bio,
3432                                                       mirror_num, &bio_flags);
3433                         if (err)
3434                                 ret = err;
3435                 } else {
3436                         unlock_page(page);
3437                 }
3438         }
3439
3440         if (bio)
3441                 submit_one_bio(READ, bio, mirror_num, bio_flags);
3442
3443         if (ret || !wait)
3444                 return ret;
3445
3446         for (i = start_i; i < num_pages; i++) {
3447                 page = extent_buffer_page(eb, i);
3448                 wait_on_page_locked(page);
3449                 if (!PageUptodate(page))
3450                         ret = -EIO;
3451         }
3452
3453         if (!ret)
3454                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3455         return ret;
3456
3457 unlock_exit:
3458         i = start_i;
3459         while (locked_pages > 0) {
3460                 page = extent_buffer_page(eb, i);
3461                 i++;
3462                 unlock_page(page);
3463                 locked_pages--;
3464         }
3465         return ret;
3466 }
3467
3468 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3469                         unsigned long start,
3470                         unsigned long len)
3471 {
3472         size_t cur;
3473         size_t offset;
3474         struct page *page;
3475         char *kaddr;
3476         char *dst = (char *)dstv;
3477         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3478         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3479
3480         WARN_ON(start > eb->len);
3481         WARN_ON(start + len > eb->start + eb->len);
3482
3483         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3484
3485         while (len > 0) {
3486                 page = extent_buffer_page(eb, i);
3487
3488                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3489                 kaddr = kmap_atomic(page, KM_USER1);
3490                 memcpy(dst, kaddr + offset, cur);
3491                 kunmap_atomic(kaddr, KM_USER1);
3492
3493                 dst += cur;
3494                 len -= cur;
3495                 offset = 0;
3496                 i++;
3497         }
3498 }
3499
3500 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3501                                unsigned long min_len, char **token, char **map,
3502                                unsigned long *map_start,
3503                                unsigned long *map_len, int km)
3504 {
3505         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3506         char *kaddr;
3507         struct page *p;
3508         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3509         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3510         unsigned long end_i = (start_offset + start + min_len - 1) >>
3511                 PAGE_CACHE_SHIFT;
3512
3513         if (i != end_i)
3514                 return -EINVAL;
3515
3516         if (i == 0) {
3517                 offset = start_offset;
3518                 *map_start = 0;
3519         } else {
3520                 offset = 0;
3521                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3522         }
3523
3524         if (start + min_len > eb->len) {
3525                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3526                        "wanted %lu %lu\n", (unsigned long long)eb->start,
3527                        eb->len, start, min_len);
3528                 WARN_ON(1);
3529         }
3530
3531         p = extent_buffer_page(eb, i);
3532         kaddr = kmap_atomic(p, km);
3533         *token = kaddr;
3534         *map = kaddr + offset;
3535         *map_len = PAGE_CACHE_SIZE - offset;
3536         return 0;
3537 }
3538
3539 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3540                       unsigned long min_len,
3541                       char **token, char **map,
3542                       unsigned long *map_start,
3543                       unsigned long *map_len, int km)
3544 {
3545         int err;
3546         int save = 0;
3547         if (eb->map_token) {
3548                 unmap_extent_buffer(eb, eb->map_token, km);
3549                 eb->map_token = NULL;
3550                 save = 1;
3551         }
3552         err = map_private_extent_buffer(eb, start, min_len, token, map,
3553                                        map_start, map_len, km);
3554         if (!err && save) {
3555                 eb->map_token = *token;
3556                 eb->kaddr = *map;
3557                 eb->map_start = *map_start;
3558                 eb->map_len = *map_len;
3559         }
3560         return err;
3561 }
3562
3563 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3564 {
3565         kunmap_atomic(token, km);
3566 }
3567
3568 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3569                           unsigned long start,
3570                           unsigned long len)
3571 {
3572         size_t cur;
3573         size_t offset;
3574         struct page *page;
3575         char *kaddr;
3576         char *ptr = (char *)ptrv;
3577         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3578         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3579         int ret = 0;
3580
3581         WARN_ON(start > eb->len);
3582         WARN_ON(start + len > eb->start + eb->len);
3583
3584         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3585
3586         while (len > 0) {
3587                 page = extent_buffer_page(eb, i);
3588
3589                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3590
3591                 kaddr = kmap_atomic(page, KM_USER0);
3592                 ret = memcmp(ptr, kaddr + offset, cur);
3593                 kunmap_atomic(kaddr, KM_USER0);
3594                 if (ret)
3595                         break;
3596
3597                 ptr += cur;
3598                 len -= cur;
3599                 offset = 0;
3600                 i++;
3601         }
3602         return ret;
3603 }
3604
3605 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3606                          unsigned long start, unsigned long len)
3607 {
3608         size_t cur;
3609         size_t offset;
3610         struct page *page;
3611         char *kaddr;
3612         char *src = (char *)srcv;
3613         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3614         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3615
3616         WARN_ON(start > eb->len);
3617         WARN_ON(start + len > eb->start + eb->len);
3618
3619         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3620
3621         while (len > 0) {
3622                 page = extent_buffer_page(eb, i);
3623                 WARN_ON(!PageUptodate(page));
3624
3625                 cur = min(len, PAGE_CACHE_SIZE - offset);
3626                 kaddr = kmap_atomic(page, KM_USER1);
3627                 memcpy(kaddr + offset, src, cur);
3628                 kunmap_atomic(kaddr, KM_USER1);
3629
3630                 src += cur;
3631                 len -= cur;
3632                 offset = 0;
3633                 i++;
3634         }
3635 }
3636
3637 void memset_extent_buffer(struct extent_buffer *eb, char c,
3638                           unsigned long start, unsigned long len)
3639 {
3640         size_t cur;
3641         size_t offset;
3642         struct page *page;
3643         char *kaddr;
3644         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3645         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3646
3647         WARN_ON(start > eb->len);
3648         WARN_ON(start + len > eb->start + eb->len);
3649
3650         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3651
3652         while (len > 0) {
3653                 page = extent_buffer_page(eb, i);
3654                 WARN_ON(!PageUptodate(page));
3655
3656                 cur = min(len, PAGE_CACHE_SIZE - offset);
3657                 kaddr = kmap_atomic(page, KM_USER0);
3658                 memset(kaddr + offset, c, cur);
3659                 kunmap_atomic(kaddr, KM_USER0);
3660
3661                 len -= cur;
3662                 offset = 0;
3663                 i++;
3664         }
3665 }
3666
3667 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3668                         unsigned long dst_offset, unsigned long src_offset,
3669                         unsigned long len)
3670 {
3671         u64 dst_len = dst->len;
3672         size_t cur;
3673         size_t offset;
3674         struct page *page;
3675         char *kaddr;
3676         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3677         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3678
3679         WARN_ON(src->len != dst_len);
3680
3681         offset = (start_offset + dst_offset) &
3682                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3683
3684         while (len > 0) {
3685                 page = extent_buffer_page(dst, i);
3686                 WARN_ON(!PageUptodate(page));
3687
3688                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3689
3690                 kaddr = kmap_atomic(page, KM_USER0);
3691                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3692                 kunmap_atomic(kaddr, KM_USER0);
3693
3694                 src_offset += cur;
3695                 len -= cur;
3696                 offset = 0;
3697                 i++;
3698         }
3699 }
3700
3701 static void move_pages(struct page *dst_page, struct page *src_page,
3702                        unsigned long dst_off, unsigned long src_off,
3703                        unsigned long len)
3704 {
3705         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3706         if (dst_page == src_page) {
3707                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3708         } else {
3709                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3710                 char *p = dst_kaddr + dst_off + len;
3711                 char *s = src_kaddr + src_off + len;
3712
3713                 while (len--)
3714                         *--p = *--s;
3715
3716                 kunmap_atomic(src_kaddr, KM_USER1);
3717         }
3718         kunmap_atomic(dst_kaddr, KM_USER0);
3719 }
3720
3721 static void copy_pages(struct page *dst_page, struct page *src_page,
3722                        unsigned long dst_off, unsigned long src_off,
3723                        unsigned long len)
3724 {
3725         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3726         char *src_kaddr;
3727
3728         if (dst_page != src_page)
3729                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3730         else
3731                 src_kaddr = dst_kaddr;
3732
3733         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3734         kunmap_atomic(dst_kaddr, KM_USER0);
3735         if (dst_page != src_page)
3736                 kunmap_atomic(src_kaddr, KM_USER1);
3737 }
3738
3739 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3740                            unsigned long src_offset, unsigned long len)
3741 {
3742         size_t cur;
3743         size_t dst_off_in_page;
3744         size_t src_off_in_page;
3745         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3746         unsigned long dst_i;
3747         unsigned long src_i;
3748
3749         if (src_offset + len > dst->len) {
3750                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3751                        "len %lu dst len %lu\n", src_offset, len, dst->len);
3752                 BUG_ON(1);
3753         }
3754         if (dst_offset + len > dst->len) {
3755                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3756                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
3757                 BUG_ON(1);
3758         }
3759
3760         while (len > 0) {
3761                 dst_off_in_page = (start_offset + dst_offset) &
3762                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3763                 src_off_in_page = (start_offset + src_offset) &
3764                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3765
3766                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3767                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3768
3769                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3770                                                src_off_in_page));
3771                 cur = min_t(unsigned long, cur,
3772                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3773
3774                 copy_pages(extent_buffer_page(dst, dst_i),
3775                            extent_buffer_page(dst, src_i),
3776                            dst_off_in_page, src_off_in_page, cur);
3777
3778                 src_offset += cur;
3779                 dst_offset += cur;
3780                 len -= cur;
3781         }
3782 }
3783
3784 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3785                            unsigned long src_offset, unsigned long len)
3786 {
3787         size_t cur;
3788         size_t dst_off_in_page;
3789         size_t src_off_in_page;
3790         unsigned long dst_end = dst_offset + len - 1;
3791         unsigned long src_end = src_offset + len - 1;
3792         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3793         unsigned long dst_i;
3794         unsigned long src_i;
3795
3796         if (src_offset + len > dst->len) {
3797                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3798                        "len %lu len %lu\n", src_offset, len, dst->len);
3799                 BUG_ON(1);
3800         }
3801         if (dst_offset + len > dst->len) {
3802                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3803                        "len %lu len %lu\n", dst_offset, len, dst->len);
3804                 BUG_ON(1);
3805         }
3806         if (dst_offset < src_offset) {
3807                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3808                 return;
3809         }
3810         while (len > 0) {
3811                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3812                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3813
3814                 dst_off_in_page = (start_offset + dst_end) &
3815                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3816                 src_off_in_page = (start_offset + src_end) &
3817                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3818
3819                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3820                 cur = min(cur, dst_off_in_page + 1);
3821                 move_pages(extent_buffer_page(dst, dst_i),
3822                            extent_buffer_page(dst, src_i),
3823                            dst_off_in_page - cur + 1,
3824                            src_off_in_page - cur + 1, cur);
3825
3826                 dst_end -= cur;
3827                 src_end -= cur;
3828                 len -= cur;
3829         }
3830 }
3831
3832 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3833 {
3834         u64 start = page_offset(page);
3835         struct extent_buffer *eb;
3836         int ret = 1;
3837         unsigned long i;
3838         unsigned long num_pages;
3839
3840         spin_lock(&tree->buffer_lock);
3841         eb = buffer_search(tree, start);
3842         if (!eb)
3843                 goto out;
3844
3845         if (atomic_read(&eb->refs) > 1) {
3846                 ret = 0;
3847                 goto out;
3848         }
3849         if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3850                 ret = 0;
3851                 goto out;
3852         }
3853         /* at this point we can safely release the extent buffer */
3854         num_pages = num_extent_pages(eb->start, eb->len);
3855         for (i = 0; i < num_pages; i++)
3856                 page_cache_release(extent_buffer_page(eb, i));
3857         rb_erase(&eb->rb_node, &tree->buffer);
3858         __free_extent_buffer(eb);
3859 out:
3860         spin_unlock(&tree->buffer_lock);
3861         return ret;
3862 }