fs/btrfs/extent_io.c

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