Pileus Git - ~andy/linux/blob - fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/gfp.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/page-flags.h>
   8 #include <linux/module.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/swap.h>
  12 #include <linux/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 | EXTENT_BOUNDARY))
 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                               (state->state & EXTENT_BOUNDARY))) {
1075                         goto out;
1076                 }
1077                 if (!(state->state & EXTENT_DELALLOC)) {
1078                         if (!found)
1079                                 *end = state->end;
1080                         goto out;
1081                 }
1082                 if (!found && !(state->state & EXTENT_BOUNDARY)) {
1083                         struct extent_state *prev_state;
1084                         struct rb_node *prev_node = node;
1085                         while(1) {
1086                                 prev_node = rb_prev(prev_node);
1087                                 if (!prev_node)
1088                                         break;
1089                                 prev_state = rb_entry(prev_node,
1090                                                       struct extent_state,
1091                                                       rb_node);
1092                                 if ((prev_state->end + 1 != state->start) ||
1093                                     !(prev_state->state & EXTENT_DELALLOC))
1094                                         break;
1095                                 if ((cur_start - prev_state->start) * 2 >
1096                                      max_bytes)
1097                                         break;
1098                                 state = prev_state;
1099                                 node = prev_node;
1100                         }
1101                 }
1102                 if (state->state & EXTENT_LOCKED) {
1103                         DEFINE_WAIT(wait);
1104                         atomic_inc(&state->refs);
1105                         prepare_to_wait(&state->wq, &wait,
1106                                         TASK_UNINTERRUPTIBLE);
1107                         spin_unlock_irq(&tree->lock);
1108                         schedule();
1109                         spin_lock_irq(&tree->lock);
1110                         finish_wait(&state->wq, &wait);
1111                         free_extent_state(state);
1112                         goto search_again;
1113                 }
1114                 set_state_cb(tree, state, EXTENT_LOCKED);
1115                 state->state |= EXTENT_LOCKED;
1116                 if (!found)
1117                         *start = state->start;
1118                 found++;
1119                 *end = state->end;
1120                 cur_start = state->end + 1;
1121                 node = rb_next(node);
1122                 if (!node)
1123                         break;
1124                 total_bytes += state->end - state->start + 1;
1125                 if (total_bytes >= max_bytes)
1126                         break;
1127         }
1128 out:
1129         spin_unlock_irq(&tree->lock);
1130         return found;
1131 }
1132
1133 u64 count_range_bits(struct extent_io_tree *tree,
1134                      u64 *start, u64 search_end, u64 max_bytes,
1135                      unsigned long bits)
1136 {
1137         struct rb_node *node;
1138         struct extent_state *state;
1139         u64 cur_start = *start;
1140         u64 total_bytes = 0;
1141         int found = 0;
1142
1143         if (search_end <= cur_start) {
1144                 printk("search_end %Lu start %Lu\n", search_end, cur_start);
1145                 WARN_ON(1);
1146                 return 0;
1147         }
1148
1149         spin_lock_irq(&tree->lock);
1150         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1151                 total_bytes = tree->dirty_bytes;
1152                 goto out;
1153         }
1154         /*
1155          * this search will find all the extents that end after
1156          * our range starts.
1157          */
1158         node = tree_search(tree, cur_start);
1159         if (!node) {
1160                 goto out;
1161         }
1162
1163         while(1) {
1164                 state = rb_entry(node, struct extent_state, rb_node);
1165                 if (state->start > search_end)
1166                         break;
1167                 if (state->end >= cur_start && (state->state & bits)) {
1168                         total_bytes += min(search_end, state->end) + 1 -
1169                                        max(cur_start, state->start);
1170                         if (total_bytes >= max_bytes)
1171                                 break;
1172                         if (!found) {
1173                                 *start = state->start;
1174                                 found = 1;
1175                         }
1176                 }
1177                 node = rb_next(node);
1178                 if (!node)
1179                         break;
1180         }
1181 out:
1182         spin_unlock_irq(&tree->lock);
1183         return total_bytes;
1184 }
1185 /*
1186  * helper function to lock both pages and extents in the tree.
1187  * pages must be locked first.
1188  */
1189 int lock_range(struct extent_io_tree *tree, u64 start, u64 end)
1190 {
1191         unsigned long index = start >> PAGE_CACHE_SHIFT;
1192         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1193         struct page *page;
1194         int err;
1195
1196         while (index <= end_index) {
1197                 page = grab_cache_page(tree->mapping, index);
1198                 if (!page) {
1199                         err = -ENOMEM;
1200                         goto failed;
1201                 }
1202                 if (IS_ERR(page)) {
1203                         err = PTR_ERR(page);
1204                         goto failed;
1205                 }
1206                 index++;
1207         }
1208         lock_extent(tree, start, end, GFP_NOFS);
1209         return 0;
1210
1211 failed:
1212         /*
1213          * we failed above in getting the page at 'index', so we undo here
1214          * up to but not including the page at 'index'
1215          */
1216         end_index = index;
1217         index = start >> PAGE_CACHE_SHIFT;
1218         while (index < end_index) {
1219                 page = find_get_page(tree->mapping, index);
1220                 unlock_page(page);
1221                 page_cache_release(page);
1222                 index++;
1223         }
1224         return err;
1225 }
1226 EXPORT_SYMBOL(lock_range);
1227
1228 /*
1229  * helper function to unlock both pages and extents in the tree.
1230  */
1231 int unlock_range(struct extent_io_tree *tree, u64 start, u64 end)
1232 {
1233         unsigned long index = start >> PAGE_CACHE_SHIFT;
1234         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1235         struct page *page;
1236
1237         while (index <= end_index) {
1238                 page = find_get_page(tree->mapping, index);
1239                 unlock_page(page);
1240                 page_cache_release(page);
1241                 index++;
1242         }
1243         unlock_extent(tree, start, end, GFP_NOFS);
1244         return 0;
1245 }
1246 EXPORT_SYMBOL(unlock_range);
1247
1248 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1249 {
1250         struct rb_node *node;
1251         struct extent_state *state;
1252         int ret = 0;
1253
1254         spin_lock_irq(&tree->lock);
1255         /*
1256          * this search will find all the extents that end after
1257          * our range starts.
1258          */
1259         node = tree_search(tree, start);
1260         if (!node) {
1261                 ret = -ENOENT;
1262                 goto out;
1263         }
1264         state = rb_entry(node, struct extent_state, rb_node);
1265         if (state->start != start) {
1266                 ret = -ENOENT;
1267                 goto out;
1268         }
1269         state->private = private;
1270 out:
1271         spin_unlock_irq(&tree->lock);
1272         return ret;
1273 }
1274
1275 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1276 {
1277         struct rb_node *node;
1278         struct extent_state *state;
1279         int ret = 0;
1280
1281         spin_lock_irq(&tree->lock);
1282         /*
1283          * this search will find all the extents that end after
1284          * our range starts.
1285          */
1286         node = tree_search(tree, start);
1287         if (!node) {
1288                 ret = -ENOENT;
1289                 goto out;
1290         }
1291         state = rb_entry(node, struct extent_state, rb_node);
1292         if (state->start != start) {
1293                 ret = -ENOENT;
1294                 goto out;
1295         }
1296         *private = state->private;
1297 out:
1298         spin_unlock_irq(&tree->lock);
1299         return ret;
1300 }
1301
1302 /*
1303  * searches a range in the state tree for a given mask.
1304  * If 'filled' == 1, this returns 1 only if every extent in the tree
1305  * has the bits set.  Otherwise, 1 is returned if any bit in the
1306  * range is found set.
1307  */
1308 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1309                    int bits, int filled)
1310 {
1311         struct extent_state *state = NULL;
1312         struct rb_node *node;
1313         int bitset = 0;
1314         unsigned long flags;
1315
1316         spin_lock_irqsave(&tree->lock, flags);
1317         node = tree_search(tree, start);
1318         while (node && start <= end) {
1319                 state = rb_entry(node, struct extent_state, rb_node);
1320
1321                 if (filled && state->start > start) {
1322                         bitset = 0;
1323                         break;
1324                 }
1325
1326                 if (state->start > end)
1327                         break;
1328
1329                 if (state->state & bits) {
1330                         bitset = 1;
1331                         if (!filled)
1332                                 break;
1333                 } else if (filled) {
1334                         bitset = 0;
1335                         break;
1336                 }
1337                 start = state->end + 1;
1338                 if (start > end)
1339                         break;
1340                 node = rb_next(node);
1341                 if (!node) {
1342                         if (filled)
1343                                 bitset = 0;
1344                         break;
1345                 }
1346         }
1347         spin_unlock_irqrestore(&tree->lock, flags);
1348         return bitset;
1349 }
1350 EXPORT_SYMBOL(test_range_bit);
1351
1352 /*
1353  * helper function to set a given page up to date if all the
1354  * extents in the tree for that page are up to date
1355  */
1356 static int check_page_uptodate(struct extent_io_tree *tree,
1357                                struct page *page)
1358 {
1359         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1360         u64 end = start + PAGE_CACHE_SIZE - 1;
1361         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1))
1362                 SetPageUptodate(page);
1363         return 0;
1364 }
1365
1366 /*
1367  * helper function to unlock a page if all the extents in the tree
1368  * for that page are unlocked
1369  */
1370 static int check_page_locked(struct extent_io_tree *tree,
1371                              struct page *page)
1372 {
1373         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1374         u64 end = start + PAGE_CACHE_SIZE - 1;
1375         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0))
1376                 unlock_page(page);
1377         return 0;
1378 }
1379
1380 /*
1381  * helper function to end page writeback if all the extents
1382  * in the tree for that page are done with writeback
1383  */
1384 static int check_page_writeback(struct extent_io_tree *tree,
1385                              struct page *page)
1386 {
1387         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1388         u64 end = start + PAGE_CACHE_SIZE - 1;
1389         if (!test_range_bit(tree, start, end, EXTENT_WRITEBACK, 0))
1390                 end_page_writeback(page);
1391         return 0;
1392 }
1393
1394 /* lots and lots of room for performance fixes in the end_bio funcs */
1395
1396 /*
1397  * after a writepage IO is done, we need to:
1398  * clear the uptodate bits on error
1399  * clear the writeback bits in the extent tree for this IO
1400  * end_page_writeback if the page has no more pending IO
1401  *
1402  * Scheduling is not allowed, so the extent state tree is expected
1403  * to have one and only one object corresponding to this IO.
1404  */
1405 static void end_bio_extent_writepage(struct bio *bio, int err)
1406 {
1407         int uptodate = err == 0;
1408         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1409         struct extent_io_tree *tree;
1410         u64 start;
1411         u64 end;
1412         int whole_page;
1413         int ret;
1414
1415         do {
1416                 struct page *page = bvec->bv_page;
1417                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1418
1419                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1420                          bvec->bv_offset;
1421                 end = start + bvec->bv_len - 1;
1422
1423                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1424                         whole_page = 1;
1425                 else
1426                         whole_page = 0;
1427
1428                 if (--bvec >= bio->bi_io_vec)
1429                         prefetchw(&bvec->bv_page->flags);
1430                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1431                         ret = tree->ops->writepage_end_io_hook(page, start,
1432                                                        end, NULL, uptodate);
1433                         if (ret)
1434                                 uptodate = 0;
1435                 }
1436
1437                 if (!uptodate && tree->ops &&
1438                     tree->ops->writepage_io_failed_hook) {
1439                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1440                                                          start, end, NULL);
1441                         if (ret == 0) {
1442                                 uptodate = (err == 0);
1443                                 continue;
1444                         }
1445                 }
1446
1447                 if (!uptodate) {
1448                         clear_extent_uptodate(tree, start, end, GFP_ATOMIC);
1449                         ClearPageUptodate(page);
1450                         SetPageError(page);
1451                 }
1452
1453                 clear_extent_writeback(tree, start, end, GFP_ATOMIC);
1454
1455                 if (whole_page)
1456                         end_page_writeback(page);
1457                 else
1458                         check_page_writeback(tree, page);
1459         } while (bvec >= bio->bi_io_vec);
1460
1461         bio_put(bio);
1462 }
1463
1464 /*
1465  * after a readpage IO is done, we need to:
1466  * clear the uptodate bits on error
1467  * set the uptodate bits if things worked
1468  * set the page up to date if all extents in the tree are uptodate
1469  * clear the lock bit in the extent tree
1470  * unlock the page if there are no other extents locked for it
1471  *
1472  * Scheduling is not allowed, so the extent state tree is expected
1473  * to have one and only one object corresponding to this IO.
1474  */
1475 static void end_bio_extent_readpage(struct bio *bio, int err)
1476 {
1477         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1478         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1479         struct extent_io_tree *tree;
1480         u64 start;
1481         u64 end;
1482         int whole_page;
1483         int ret;
1484
1485         do {
1486                 struct page *page = bvec->bv_page;
1487                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1488
1489                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1490                         bvec->bv_offset;
1491                 end = start + bvec->bv_len - 1;
1492
1493                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1494                         whole_page = 1;
1495                 else
1496                         whole_page = 0;
1497
1498                 if (--bvec >= bio->bi_io_vec)
1499                         prefetchw(&bvec->bv_page->flags);
1500
1501                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1502                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1503                                                               NULL);
1504                         if (ret)
1505                                 uptodate = 0;
1506                 }
1507                 if (!uptodate && tree->ops &&
1508                     tree->ops->readpage_io_failed_hook) {
1509                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1510                                                          start, end, NULL);
1511                         if (ret == 0) {
1512                                 uptodate =
1513                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1514                                 continue;
1515                         }
1516                 }
1517
1518                 if (uptodate)
1519                         set_extent_uptodate(tree, start, end,
1520                                             GFP_ATOMIC);
1521                 unlock_extent(tree, start, end, GFP_ATOMIC);
1522
1523                 if (whole_page) {
1524                         if (uptodate) {
1525                                 SetPageUptodate(page);
1526                         } else {
1527                                 ClearPageUptodate(page);
1528                                 SetPageError(page);
1529                         }
1530                         unlock_page(page);
1531                 } else {
1532                         if (uptodate) {
1533                                 check_page_uptodate(tree, page);
1534                         } else {
1535                                 ClearPageUptodate(page);
1536                                 SetPageError(page);
1537                         }
1538                         check_page_locked(tree, page);
1539                 }
1540         } while (bvec >= bio->bi_io_vec);
1541
1542         bio_put(bio);
1543 }
1544
1545 /*
1546  * IO done from prepare_write is pretty simple, we just unlock
1547  * the structs in the extent tree when done, and set the uptodate bits
1548  * as appropriate.
1549  */
1550 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1551 {
1552         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1553         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1554         struct extent_io_tree *tree;
1555         u64 start;
1556         u64 end;
1557
1558         do {
1559                 struct page *page = bvec->bv_page;
1560                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1561
1562                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1563                         bvec->bv_offset;
1564                 end = start + bvec->bv_len - 1;
1565
1566                 if (--bvec >= bio->bi_io_vec)
1567                         prefetchw(&bvec->bv_page->flags);
1568
1569                 if (uptodate) {
1570                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1571                 } else {
1572                         ClearPageUptodate(page);
1573                         SetPageError(page);
1574                 }
1575
1576                 unlock_extent(tree, start, end, GFP_ATOMIC);
1577
1578         } while (bvec >= bio->bi_io_vec);
1579
1580         bio_put(bio);
1581 }
1582
1583 static struct bio *
1584 extent_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1585                  gfp_t gfp_flags)
1586 {
1587         struct bio *bio;
1588
1589         bio = bio_alloc(gfp_flags, nr_vecs);
1590
1591         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1592                 while (!bio && (nr_vecs /= 2))
1593                         bio = bio_alloc(gfp_flags, nr_vecs);
1594         }
1595
1596         if (bio) {
1597                 bio->bi_size = 0;
1598                 bio->bi_bdev = bdev;
1599                 bio->bi_sector = first_sector;
1600         }
1601         return bio;
1602 }
1603
1604 static int submit_one_bio(int rw, struct bio *bio, int mirror_num)
1605 {
1606         int ret = 0;
1607         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1608         struct page *page = bvec->bv_page;
1609         struct extent_io_tree *tree = bio->bi_private;
1610         struct rb_node *node;
1611         struct extent_state *state;
1612         u64 start;
1613         u64 end;
1614
1615         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1616         end = start + bvec->bv_len - 1;
1617
1618         spin_lock_irq(&tree->lock);
1619         node = __etree_search(tree, start, NULL, NULL);
1620         BUG_ON(!node);
1621         state = rb_entry(node, struct extent_state, rb_node);
1622         while(state->end < end) {
1623                 node = rb_next(node);
1624                 state = rb_entry(node, struct extent_state, rb_node);
1625         }
1626         BUG_ON(state->end != end);
1627         spin_unlock_irq(&tree->lock);
1628
1629         bio->bi_private = NULL;
1630
1631         bio_get(bio);
1632
1633         if (tree->ops && tree->ops->submit_bio_hook)
1634                 tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1635                                            mirror_num);
1636         else
1637                 submit_bio(rw, bio);
1638         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1639                 ret = -EOPNOTSUPP;
1640         bio_put(bio);
1641         return ret;
1642 }
1643
1644 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1645                               struct page *page, sector_t sector,
1646                               size_t size, unsigned long offset,
1647                               struct block_device *bdev,
1648                               struct bio **bio_ret,
1649                               unsigned long max_pages,
1650                               bio_end_io_t end_io_func,
1651                               int mirror_num)
1652 {
1653         int ret = 0;
1654         struct bio *bio;
1655         int nr;
1656
1657         if (bio_ret && *bio_ret) {
1658                 bio = *bio_ret;
1659                 if (bio->bi_sector + (bio->bi_size >> 9) != sector ||
1660                     (tree->ops && tree->ops->merge_bio_hook &&
1661                      tree->ops->merge_bio_hook(page, offset, size, bio)) ||
1662                     bio_add_page(bio, page, size, offset) < size) {
1663                         ret = submit_one_bio(rw, bio, mirror_num);
1664                         bio = NULL;
1665                 } else {
1666                         return 0;
1667                 }
1668         }
1669         nr = bio_get_nr_vecs(bdev);
1670         bio = extent_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1671         if (!bio) {
1672                 printk("failed to allocate bio nr %d\n", nr);
1673         }
1674
1675
1676         bio_add_page(bio, page, size, offset);
1677         bio->bi_end_io = end_io_func;
1678         bio->bi_private = tree;
1679
1680         if (bio_ret) {
1681                 *bio_ret = bio;
1682         } else {
1683                 ret = submit_one_bio(rw, bio, mirror_num);
1684         }
1685
1686         return ret;
1687 }
1688
1689 void set_page_extent_mapped(struct page *page)
1690 {
1691         if (!PagePrivate(page)) {
1692                 SetPagePrivate(page);
1693                 page_cache_get(page);
1694                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1695         }
1696 }
1697
1698 void set_page_extent_head(struct page *page, unsigned long len)
1699 {
1700         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1701 }
1702
1703 /*
1704  * basic readpage implementation.  Locked extent state structs are inserted
1705  * into the tree that are removed when the IO is done (by the end_io
1706  * handlers)
1707  */
1708 static int __extent_read_full_page(struct extent_io_tree *tree,
1709                                    struct page *page,
1710                                    get_extent_t *get_extent,
1711                                    struct bio **bio, int mirror_num)
1712 {
1713         struct inode *inode = page->mapping->host;
1714         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1715         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1716         u64 end;
1717         u64 cur = start;
1718         u64 extent_offset;
1719         u64 last_byte = i_size_read(inode);
1720         u64 block_start;
1721         u64 cur_end;
1722         sector_t sector;
1723         struct extent_map *em;
1724         struct block_device *bdev;
1725         int ret;
1726         int nr = 0;
1727         size_t page_offset = 0;
1728         size_t iosize;
1729         size_t blocksize = inode->i_sb->s_blocksize;
1730
1731         set_page_extent_mapped(page);
1732
1733         end = page_end;
1734         lock_extent(tree, start, end, GFP_NOFS);
1735
1736         while (cur <= end) {
1737                 if (cur >= last_byte) {
1738                         char *userpage;
1739                         iosize = PAGE_CACHE_SIZE - page_offset;
1740                         userpage = kmap_atomic(page, KM_USER0);
1741                         memset(userpage + page_offset, 0, iosize);
1742                         flush_dcache_page(page);
1743                         kunmap_atomic(userpage, KM_USER0);
1744                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1745                                             GFP_NOFS);
1746                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1747                         break;
1748                 }
1749                 em = get_extent(inode, page, page_offset, cur,
1750                                 end - cur + 1, 0);
1751                 if (IS_ERR(em) || !em) {
1752                         SetPageError(page);
1753                         unlock_extent(tree, cur, end, GFP_NOFS);
1754                         break;
1755                 }
1756                 extent_offset = cur - em->start;
1757                 if (extent_map_end(em) <= cur) {
1758 printk("bad mapping em [%Lu %Lu] cur %Lu\n", em->start, extent_map_end(em), cur);
1759                 }
1760                 BUG_ON(extent_map_end(em) <= cur);
1761                 if (end < cur) {
1762 printk("2bad mapping end %Lu cur %Lu\n", end, cur);
1763                 }
1764                 BUG_ON(end < cur);
1765
1766                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1767                 cur_end = min(extent_map_end(em) - 1, end);
1768                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1769                 sector = (em->block_start + extent_offset) >> 9;
1770                 bdev = em->bdev;
1771                 block_start = em->block_start;
1772                 free_extent_map(em);
1773                 em = NULL;
1774
1775                 /* we've found a hole, just zero and go on */
1776                 if (block_start == EXTENT_MAP_HOLE) {
1777                         char *userpage;
1778                         userpage = kmap_atomic(page, KM_USER0);
1779                         memset(userpage + page_offset, 0, iosize);
1780                         flush_dcache_page(page);
1781                         kunmap_atomic(userpage, KM_USER0);
1782
1783                         set_extent_uptodate(tree, cur, cur + iosize - 1,
1784                                             GFP_NOFS);
1785                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1786                         cur = cur + iosize;
1787                         page_offset += iosize;
1788                         continue;
1789                 }
1790                 /* the get_extent function already copied into the page */
1791                 if (test_range_bit(tree, cur, cur_end, EXTENT_UPTODATE, 1)) {
1792                         check_page_uptodate(tree, page);
1793                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1794                         cur = cur + iosize;
1795                         page_offset += iosize;
1796                         continue;
1797                 }
1798                 /* we have an inline extent but it didn't get marked up
1799                  * to date.  Error out
1800                  */
1801                 if (block_start == EXTENT_MAP_INLINE) {
1802                         SetPageError(page);
1803                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
1804                         cur = cur + iosize;
1805                         page_offset += iosize;
1806                         continue;
1807                 }
1808
1809                 ret = 0;
1810                 if (tree->ops && tree->ops->readpage_io_hook) {
1811                         ret = tree->ops->readpage_io_hook(page, cur,
1812                                                           cur + iosize - 1);
1813                 }
1814                 if (!ret) {
1815                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
1816                         pnr -= page->index;
1817                         ret = submit_extent_page(READ, tree, page,
1818                                          sector, iosize, page_offset,
1819                                          bdev, bio, pnr,
1820                                          end_bio_extent_readpage, mirror_num);
1821                         nr++;
1822                 }
1823                 if (ret)
1824                         SetPageError(page);
1825                 cur = cur + iosize;
1826                 page_offset += iosize;
1827         }
1828         if (!nr) {
1829                 if (!PageError(page))
1830                         SetPageUptodate(page);
1831                 unlock_page(page);
1832         }
1833         return 0;
1834 }
1835
1836 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
1837                             get_extent_t *get_extent)
1838 {
1839         struct bio *bio = NULL;
1840         int ret;
1841
1842         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0);
1843         if (bio)
1844                 submit_one_bio(READ, bio, 0);
1845         return ret;
1846 }
1847 EXPORT_SYMBOL(extent_read_full_page);
1848
1849 /*
1850  * the writepage semantics are similar to regular writepage.  extent
1851  * records are inserted to lock ranges in the tree, and as dirty areas
1852  * are found, they are marked writeback.  Then the lock bits are removed
1853  * and the end_io handler clears the writeback ranges
1854  */
1855 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
1856                               void *data)
1857 {
1858         struct inode *inode = page->mapping->host;
1859         struct extent_page_data *epd = data;
1860         struct extent_io_tree *tree = epd->tree;
1861         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1862         u64 delalloc_start;
1863         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1864         u64 end;
1865         u64 cur = start;
1866         u64 extent_offset;
1867         u64 last_byte = i_size_read(inode);
1868         u64 block_start;
1869         u64 iosize;
1870         u64 unlock_start;
1871         sector_t sector;
1872         struct extent_map *em;
1873         struct block_device *bdev;
1874         int ret;
1875         int nr = 0;
1876         size_t pg_offset = 0;
1877         size_t blocksize;
1878         loff_t i_size = i_size_read(inode);
1879         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
1880         u64 nr_delalloc;
1881         u64 delalloc_end;
1882
1883         WARN_ON(!PageLocked(page));
1884         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
1885         if (page->index > end_index ||
1886            (page->index == end_index && !pg_offset)) {
1887                 page->mapping->a_ops->invalidatepage(page, 0);
1888                 unlock_page(page);
1889                 return 0;
1890         }
1891
1892         if (page->index == end_index) {
1893                 char *userpage;
1894
1895                 userpage = kmap_atomic(page, KM_USER0);
1896                 memset(userpage + pg_offset, 0,
1897                        PAGE_CACHE_SIZE - pg_offset);
1898                 kunmap_atomic(userpage, KM_USER0);
1899                 flush_dcache_page(page);
1900         }
1901         pg_offset = 0;
1902
1903         set_page_extent_mapped(page);
1904
1905         delalloc_start = start;
1906         delalloc_end = 0;
1907         while(delalloc_end < page_end) {
1908                 nr_delalloc = find_lock_delalloc_range(tree, &delalloc_start,
1909                                                        &delalloc_end,
1910                                                        128 * 1024 * 1024);
1911                 if (nr_delalloc == 0) {
1912                         delalloc_start = delalloc_end + 1;
1913                         continue;
1914                 }
1915                 tree->ops->fill_delalloc(inode, delalloc_start,
1916                                          delalloc_end);
1917                 clear_extent_bit(tree, delalloc_start,
1918                                  delalloc_end,
1919                                  EXTENT_LOCKED | EXTENT_DELALLOC,
1920                                  1, 0, GFP_NOFS);
1921                 delalloc_start = delalloc_end + 1;
1922         }
1923         lock_extent(tree, start, page_end, GFP_NOFS);
1924         unlock_start = start;
1925
1926         if (tree->ops && tree->ops->writepage_start_hook) {
1927                 ret = tree->ops->writepage_start_hook(page, start, page_end);
1928                 if (ret == -EAGAIN) {
1929                         unlock_extent(tree, start, page_end, GFP_NOFS);
1930                         redirty_page_for_writepage(wbc, page);
1931                         unlock_page(page);
1932                         return 0;
1933                 }
1934         }
1935
1936         end = page_end;
1937         if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0)) {
1938                 printk("found delalloc bits after lock_extent\n");
1939         }
1940
1941         if (last_byte <= start) {
1942                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
1943                 unlock_extent(tree, start, page_end, GFP_NOFS);
1944                 if (tree->ops && tree->ops->writepage_end_io_hook)
1945                         tree->ops->writepage_end_io_hook(page, start,
1946                                                          page_end, NULL, 1);
1947                 unlock_start = page_end + 1;
1948                 goto done;
1949         }
1950
1951         set_extent_uptodate(tree, start, page_end, GFP_NOFS);
1952         blocksize = inode->i_sb->s_blocksize;
1953
1954         while (cur <= end) {
1955                 if (cur >= last_byte) {
1956                         clear_extent_dirty(tree, cur, page_end, GFP_NOFS);
1957                         unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
1958                         if (tree->ops && tree->ops->writepage_end_io_hook)
1959                                 tree->ops->writepage_end_io_hook(page, cur,
1960                                                          page_end, NULL, 1);
1961                         unlock_start = page_end + 1;
1962                         break;
1963                 }
1964                 em = epd->get_extent(inode, page, pg_offset, cur,
1965                                      end - cur + 1, 1);
1966                 if (IS_ERR(em) || !em) {
1967                         SetPageError(page);
1968                         break;
1969                 }
1970
1971                 extent_offset = cur - em->start;
1972                 BUG_ON(extent_map_end(em) <= cur);
1973                 BUG_ON(end < cur);
1974                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1975                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1976                 sector = (em->block_start + extent_offset) >> 9;
1977                 bdev = em->bdev;
1978                 block_start = em->block_start;
1979                 free_extent_map(em);
1980                 em = NULL;
1981
1982                 if (block_start == EXTENT_MAP_HOLE ||
1983                     block_start == EXTENT_MAP_INLINE) {
1984                         clear_extent_dirty(tree, cur,
1985                                            cur + iosize - 1, GFP_NOFS);
1986
1987                         unlock_extent(tree, unlock_start, cur + iosize -1,
1988                                       GFP_NOFS);
1989
1990                         if (tree->ops && tree->ops->writepage_end_io_hook)
1991                                 tree->ops->writepage_end_io_hook(page, cur,
1992                                                          cur + iosize - 1,
1993                                                          NULL, 1);
1994                         cur = cur + iosize;
1995                         pg_offset += iosize;
1996                         unlock_start = cur;
1997                         continue;
1998                 }
1999
2000                 /* leave this out until we have a page_mkwrite call */
2001                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2002                                    EXTENT_DIRTY, 0)) {
2003                         cur = cur + iosize;
2004                         pg_offset += iosize;
2005                         continue;
2006                 }
2007                 clear_extent_dirty(tree, cur, cur + iosize - 1, GFP_NOFS);
2008                 if (tree->ops && tree->ops->writepage_io_hook) {
2009                         ret = tree->ops->writepage_io_hook(page, cur,
2010                                                 cur + iosize - 1);
2011                 } else {
2012                         ret = 0;
2013                 }
2014                 if (ret) {
2015                         SetPageError(page);
2016                 } else {
2017                         unsigned long max_nr = end_index + 1;
2018
2019                         set_range_writeback(tree, cur, cur + iosize - 1);
2020                         if (!PageWriteback(page)) {
2021                                 printk("warning page %lu not writeback, "
2022                                        "cur %llu end %llu\n", page->index,
2023                                        (unsigned long long)cur,
2024                                        (unsigned long long)end);
2025                         }
2026
2027                         ret = submit_extent_page(WRITE, tree, page, sector,
2028                                                  iosize, pg_offset, bdev,
2029                                                  &epd->bio, max_nr,
2030                                                  end_bio_extent_writepage, 0);
2031                         if (ret)
2032                                 SetPageError(page);
2033                 }
2034                 cur = cur + iosize;
2035                 pg_offset += iosize;
2036                 nr++;
2037         }
2038 done:
2039         if (nr == 0) {
2040                 /* make sure the mapping tag for page dirty gets cleared */
2041                 set_page_writeback(page);
2042                 end_page_writeback(page);
2043         }
2044         if (unlock_start <= page_end)
2045                 unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
2046         unlock_page(page);
2047         return 0;
2048 }
2049
2050 /**
2051  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2052  * @mapping: address space structure to write
2053  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2054  * @writepage: function called for each page
2055  * @data: data passed to writepage function
2056  *
2057  * If a page is already under I/O, write_cache_pages() skips it, even
2058  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2059  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2060  * and msync() need to guarantee that all the data which was dirty at the time
2061  * the call was made get new I/O started against them.  If wbc->sync_mode is
2062  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2063  * existing IO to complete.
2064  */
2065 int extent_write_cache_pages(struct extent_io_tree *tree,
2066                              struct address_space *mapping,
2067                              struct writeback_control *wbc,
2068                              writepage_t writepage, void *data)
2069 {
2070         struct backing_dev_info *bdi = mapping->backing_dev_info;
2071         int ret = 0;
2072         int done = 0;
2073         struct pagevec pvec;
2074         int nr_pages;
2075         pgoff_t index;
2076         pgoff_t end;            /* Inclusive */
2077         int scanned = 0;
2078         int range_whole = 0;
2079
2080         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2081                 wbc->encountered_congestion = 1;
2082                 return 0;
2083         }
2084
2085         pagevec_init(&pvec, 0);
2086         if (wbc->range_cyclic) {
2087                 index = mapping->writeback_index; /* Start from prev offset */
2088                 end = -1;
2089         } else {
2090                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2091                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2092                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2093                         range_whole = 1;
2094                 scanned = 1;
2095         }
2096 retry:
2097         while (!done && (index <= end) &&
2098                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2099                                               PAGECACHE_TAG_DIRTY,
2100                                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2101                 unsigned i;
2102
2103                 scanned = 1;
2104                 for (i = 0; i < nr_pages; i++) {
2105                         struct page *page = pvec.pages[i];
2106
2107                         /*
2108                          * At this point we hold neither mapping->tree_lock nor
2109                          * lock on the page itself: the page may be truncated or
2110                          * invalidated (changing page->mapping to NULL), or even
2111                          * swizzled back from swapper_space to tmpfs file
2112                          * mapping
2113                          */
2114                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2115                                 tree->ops->write_cache_pages_lock_hook(page);
2116                         else
2117                                 lock_page(page);
2118
2119                         if (unlikely(page->mapping != mapping)) {
2120                                 unlock_page(page);
2121                                 continue;
2122                         }
2123
2124                         if (!wbc->range_cyclic && page->index > end) {
2125                                 done = 1;
2126                                 unlock_page(page);
2127                                 continue;
2128                         }
2129
2130                         if (wbc->sync_mode != WB_SYNC_NONE)
2131                                 wait_on_page_writeback(page);
2132
2133                         if (PageWriteback(page) ||
2134                             !clear_page_dirty_for_io(page)) {
2135                                 unlock_page(page);
2136                                 continue;
2137                         }
2138
2139                         ret = (*writepage)(page, wbc, data);
2140
2141                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2142                                 unlock_page(page);
2143                                 ret = 0;
2144                         }
2145                         if (ret || (--(wbc->nr_to_write) <= 0))
2146                                 done = 1;
2147                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2148                                 wbc->encountered_congestion = 1;
2149                                 done = 1;
2150                         }
2151                 }
2152                 pagevec_release(&pvec);
2153                 cond_resched();
2154         }
2155         if (!scanned && !done) {
2156                 /*
2157                  * We hit the last page and there is more work to be done: wrap
2158                  * back to the start of the file
2159                  */
2160                 scanned = 1;
2161                 index = 0;
2162                 goto retry;
2163         }
2164         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2165                 mapping->writeback_index = index;
2166
2167         if (wbc->range_cont)
2168                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2169         return ret;
2170 }
2171 EXPORT_SYMBOL(extent_write_cache_pages);
2172
2173 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2174                           get_extent_t *get_extent,
2175                           struct writeback_control *wbc)
2176 {
2177         int ret;
2178         struct address_space *mapping = page->mapping;
2179         struct extent_page_data epd = {
2180                 .bio = NULL,
2181                 .tree = tree,
2182                 .get_extent = get_extent,
2183         };
2184         struct writeback_control wbc_writepages = {
2185                 .bdi            = wbc->bdi,
2186                 .sync_mode      = WB_SYNC_NONE,
2187                 .older_than_this = NULL,
2188                 .nr_to_write    = 64,
2189                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2190                 .range_end      = (loff_t)-1,
2191         };
2192
2193
2194         ret = __extent_writepage(page, wbc, &epd);
2195
2196         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2197                                  __extent_writepage, &epd);
2198         if (epd.bio) {
2199                 submit_one_bio(WRITE, epd.bio, 0);
2200         }
2201         return ret;
2202 }
2203 EXPORT_SYMBOL(extent_write_full_page);
2204
2205
2206 int extent_writepages(struct extent_io_tree *tree,
2207                       struct address_space *mapping,
2208                       get_extent_t *get_extent,
2209                       struct writeback_control *wbc)
2210 {
2211         int ret = 0;
2212         struct extent_page_data epd = {
2213                 .bio = NULL,
2214                 .tree = tree,
2215                 .get_extent = get_extent,
2216         };
2217
2218         ret = extent_write_cache_pages(tree, mapping, wbc,
2219                                        __extent_writepage, &epd);
2220         if (epd.bio) {
2221                 submit_one_bio(WRITE, epd.bio, 0);
2222         }
2223         return ret;
2224 }
2225 EXPORT_SYMBOL(extent_writepages);
2226
2227 int extent_readpages(struct extent_io_tree *tree,
2228                      struct address_space *mapping,
2229                      struct list_head *pages, unsigned nr_pages,
2230                      get_extent_t get_extent)
2231 {
2232         struct bio *bio = NULL;
2233         unsigned page_idx;
2234         struct pagevec pvec;
2235
2236         pagevec_init(&pvec, 0);
2237         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2238                 struct page *page = list_entry(pages->prev, struct page, lru);
2239
2240                 prefetchw(&page->flags);
2241                 list_del(&page->lru);
2242                 /*
2243                  * what we want to do here is call add_to_page_cache_lru,
2244                  * but that isn't exported, so we reproduce it here
2245                  */
2246                 if (!add_to_page_cache(page, mapping,
2247                                         page->index, GFP_KERNEL)) {
2248
2249                         /* open coding of lru_cache_add, also not exported */
2250                         page_cache_get(page);
2251                         if (!pagevec_add(&pvec, page))
2252                                 __pagevec_lru_add(&pvec);
2253                         __extent_read_full_page(tree, page, get_extent,
2254                                                 &bio, 0);
2255                 }
2256                 page_cache_release(page);
2257         }
2258         if (pagevec_count(&pvec))
2259                 __pagevec_lru_add(&pvec);
2260         BUG_ON(!list_empty(pages));
2261         if (bio)
2262                 submit_one_bio(READ, bio, 0);
2263         return 0;
2264 }
2265 EXPORT_SYMBOL(extent_readpages);
2266
2267 /*
2268  * basic invalidatepage code, this waits on any locked or writeback
2269  * ranges corresponding to the page, and then deletes any extent state
2270  * records from the tree
2271  */
2272 int extent_invalidatepage(struct extent_io_tree *tree,
2273                           struct page *page, unsigned long offset)
2274 {
2275         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2276         u64 end = start + PAGE_CACHE_SIZE - 1;
2277         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2278
2279         start += (offset + blocksize -1) & ~(blocksize - 1);
2280         if (start > end)
2281                 return 0;
2282
2283         lock_extent(tree, start, end, GFP_NOFS);
2284         wait_on_extent_writeback(tree, start, end);
2285         clear_extent_bit(tree, start, end,
2286                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC,
2287                          1, 1, GFP_NOFS);
2288         return 0;
2289 }
2290 EXPORT_SYMBOL(extent_invalidatepage);
2291
2292 /*
2293  * simple commit_write call, set_range_dirty is used to mark both
2294  * the pages and the extent records as dirty
2295  */
2296 int extent_commit_write(struct extent_io_tree *tree,
2297                         struct inode *inode, struct page *page,
2298                         unsigned from, unsigned to)
2299 {
2300         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2301
2302         set_page_extent_mapped(page);
2303         set_page_dirty(page);
2304
2305         if (pos > inode->i_size) {
2306                 i_size_write(inode, pos);
2307                 mark_inode_dirty(inode);
2308         }
2309         return 0;
2310 }
2311 EXPORT_SYMBOL(extent_commit_write);
2312
2313 int extent_prepare_write(struct extent_io_tree *tree,
2314                          struct inode *inode, struct page *page,
2315                          unsigned from, unsigned to, get_extent_t *get_extent)
2316 {
2317         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2318         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2319         u64 block_start;
2320         u64 orig_block_start;
2321         u64 block_end;
2322         u64 cur_end;
2323         struct extent_map *em;
2324         unsigned blocksize = 1 << inode->i_blkbits;
2325         size_t page_offset = 0;
2326         size_t block_off_start;
2327         size_t block_off_end;
2328         int err = 0;
2329         int iocount = 0;
2330         int ret = 0;
2331         int isnew;
2332
2333         set_page_extent_mapped(page);
2334
2335         block_start = (page_start + from) & ~((u64)blocksize - 1);
2336         block_end = (page_start + to - 1) | (blocksize - 1);
2337         orig_block_start = block_start;
2338
2339         lock_extent(tree, page_start, page_end, GFP_NOFS);
2340         while(block_start <= block_end) {
2341                 em = get_extent(inode, page, page_offset, block_start,
2342                                 block_end - block_start + 1, 1);
2343                 if (IS_ERR(em) || !em) {
2344                         goto err;
2345                 }
2346                 cur_end = min(block_end, extent_map_end(em) - 1);
2347                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2348                 block_off_end = block_off_start + blocksize;
2349                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2350
2351                 if (!PageUptodate(page) && isnew &&
2352                     (block_off_end > to || block_off_start < from)) {
2353                         void *kaddr;
2354
2355                         kaddr = kmap_atomic(page, KM_USER0);
2356                         if (block_off_end > to)
2357                                 memset(kaddr + to, 0, block_off_end - to);
2358                         if (block_off_start < from)
2359                                 memset(kaddr + block_off_start, 0,
2360                                        from - block_off_start);
2361                         flush_dcache_page(page);
2362                         kunmap_atomic(kaddr, KM_USER0);
2363                 }
2364                 if ((em->block_start != EXTENT_MAP_HOLE &&
2365                      em->block_start != EXTENT_MAP_INLINE) &&
2366                     !isnew && !PageUptodate(page) &&
2367                     (block_off_end > to || block_off_start < from) &&
2368                     !test_range_bit(tree, block_start, cur_end,
2369                                     EXTENT_UPTODATE, 1)) {
2370                         u64 sector;
2371                         u64 extent_offset = block_start - em->start;
2372                         size_t iosize;
2373                         sector = (em->block_start + extent_offset) >> 9;
2374                         iosize = (cur_end - block_start + blocksize) &
2375                                 ~((u64)blocksize - 1);
2376                         /*
2377                          * we've already got the extent locked, but we
2378                          * need to split the state such that our end_bio
2379                          * handler can clear the lock.
2380                          */
2381                         set_extent_bit(tree, block_start,
2382                                        block_start + iosize - 1,
2383                                        EXTENT_LOCKED, 0, NULL, GFP_NOFS);
2384                         ret = submit_extent_page(READ, tree, page,
2385                                          sector, iosize, page_offset, em->bdev,
2386                                          NULL, 1,
2387                                          end_bio_extent_preparewrite, 0);
2388                         iocount++;
2389                         block_start = block_start + iosize;
2390                 } else {
2391                         set_extent_uptodate(tree, block_start, cur_end,
2392                                             GFP_NOFS);
2393                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2394                         block_start = cur_end + 1;
2395                 }
2396                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2397                 free_extent_map(em);
2398         }
2399         if (iocount) {
2400                 wait_extent_bit(tree, orig_block_start,
2401                                 block_end, EXTENT_LOCKED);
2402         }
2403         check_page_uptodate(tree, page);
2404 err:
2405         /* FIXME, zero out newly allocated blocks on error */
2406         return err;
2407 }
2408 EXPORT_SYMBOL(extent_prepare_write);
2409
2410 /*
2411  * a helper for releasepage, this tests for areas of the page that
2412  * are locked or under IO and drops the related state bits if it is safe
2413  * to drop the page.
2414  */
2415 int try_release_extent_state(struct extent_map_tree *map,
2416                              struct extent_io_tree *tree, struct page *page,
2417                              gfp_t mask)
2418 {
2419         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2420         u64 end = start + PAGE_CACHE_SIZE - 1;
2421         int ret = 1;
2422
2423         if (test_range_bit(tree, start, end,
2424                            EXTENT_IOBITS | EXTENT_ORDERED, 0))
2425                 ret = 0;
2426         else {
2427                 if ((mask & GFP_NOFS) == GFP_NOFS)
2428                         mask = GFP_NOFS;
2429                 clear_extent_bit(tree, start, end, EXTENT_UPTODATE,
2430                                  1, 1, mask);
2431         }
2432         return ret;
2433 }
2434 EXPORT_SYMBOL(try_release_extent_state);
2435
2436 /*
2437  * a helper for releasepage.  As long as there are no locked extents
2438  * in the range corresponding to the page, both state records and extent
2439  * map records are removed
2440  */
2441 int try_release_extent_mapping(struct extent_map_tree *map,
2442                                struct extent_io_tree *tree, struct page *page,
2443                                gfp_t mask)
2444 {
2445         struct extent_map *em;
2446         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2447         u64 end = start + PAGE_CACHE_SIZE - 1;
2448
2449         if ((mask & __GFP_WAIT) &&
2450             page->mapping->host->i_size > 16 * 1024 * 1024) {
2451                 u64 len;
2452                 while (start <= end) {
2453                         len = end - start + 1;
2454                         spin_lock(&map->lock);
2455                         em = lookup_extent_mapping(map, start, len);
2456                         if (!em || IS_ERR(em)) {
2457                                 spin_unlock(&map->lock);
2458                                 break;
2459                         }
2460                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2461                             em->start != start) {
2462                                 spin_unlock(&map->lock);
2463                                 free_extent_map(em);
2464                                 break;
2465                         }
2466                         if (!test_range_bit(tree, em->start,
2467                                             extent_map_end(em) - 1,
2468                                             EXTENT_LOCKED, 0)) {
2469                                 remove_extent_mapping(map, em);
2470                                 /* once for the rb tree */
2471                                 free_extent_map(em);
2472                         }
2473                         start = extent_map_end(em);
2474                         spin_unlock(&map->lock);
2475
2476                         /* once for us */
2477                         free_extent_map(em);
2478                 }
2479         }
2480         return try_release_extent_state(map, tree, page, mask);
2481 }
2482 EXPORT_SYMBOL(try_release_extent_mapping);
2483
2484 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2485                 get_extent_t *get_extent)
2486 {
2487         struct inode *inode = mapping->host;
2488         u64 start = iblock << inode->i_blkbits;
2489         sector_t sector = 0;
2490         struct extent_map *em;
2491
2492         em = get_extent(inode, NULL, 0, start, (1 << inode->i_blkbits), 0);
2493         if (!em || IS_ERR(em))
2494                 return 0;
2495
2496         if (em->block_start == EXTENT_MAP_INLINE ||
2497             em->block_start == EXTENT_MAP_HOLE)
2498                 goto out;
2499
2500         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2501 out:
2502         free_extent_map(em);
2503         return sector;
2504 }
2505
2506 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2507                                               unsigned long i)
2508 {
2509         struct page *p;
2510         struct address_space *mapping;
2511
2512         if (i == 0)
2513                 return eb->first_page;
2514         i += eb->start >> PAGE_CACHE_SHIFT;
2515         mapping = eb->first_page->mapping;
2516         if (!mapping)
2517                 return NULL;
2518
2519         /*
2520          * extent_buffer_page is only called after pinning the page
2521          * by increasing the reference count.  So we know the page must
2522          * be in the radix tree.
2523          */
2524         rcu_read_lock();
2525         p = radix_tree_lookup(&mapping->page_tree, i);
2526         rcu_read_unlock();
2527
2528         return p;
2529 }
2530
2531 static inline unsigned long num_extent_pages(u64 start, u64 len)
2532 {
2533         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
2534                 (start >> PAGE_CACHE_SHIFT);
2535 }
2536
2537 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
2538                                                    u64 start,
2539                                                    unsigned long len,
2540                                                    gfp_t mask)
2541 {
2542         struct extent_buffer *eb = NULL;
2543 #ifdef LEAK_DEBUG
2544         unsigned long flags;
2545 #endif
2546
2547         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
2548         eb->start = start;
2549         eb->len = len;
2550         mutex_init(&eb->mutex);
2551 #ifdef LEAK_DEBUG
2552         spin_lock_irqsave(&leak_lock, flags);
2553         list_add(&eb->leak_list, &buffers);
2554         spin_unlock_irqrestore(&leak_lock, flags);
2555 #endif
2556         atomic_set(&eb->refs, 1);
2557
2558         return eb;
2559 }
2560
2561 static void __free_extent_buffer(struct extent_buffer *eb)
2562 {
2563 #ifdef LEAK_DEBUG
2564         unsigned long flags;
2565         spin_lock_irqsave(&leak_lock, flags);
2566         list_del(&eb->leak_list);
2567         spin_unlock_irqrestore(&leak_lock, flags);
2568 #endif
2569         kmem_cache_free(extent_buffer_cache, eb);
2570 }
2571
2572 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
2573                                           u64 start, unsigned long len,
2574                                           struct page *page0,
2575                                           gfp_t mask)
2576 {
2577         unsigned long num_pages = num_extent_pages(start, len);
2578         unsigned long i;
2579         unsigned long index = start >> PAGE_CACHE_SHIFT;
2580         struct extent_buffer *eb;
2581         struct extent_buffer *exists = NULL;
2582         struct page *p;
2583         struct address_space *mapping = tree->mapping;
2584         int uptodate = 1;
2585
2586         spin_lock(&tree->buffer_lock);
2587         eb = buffer_search(tree, start);
2588         if (eb) {
2589                 atomic_inc(&eb->refs);
2590                 spin_unlock(&tree->buffer_lock);
2591                 mark_page_accessed(eb->first_page);
2592                 return eb;
2593         }
2594         spin_unlock(&tree->buffer_lock);
2595
2596         eb = __alloc_extent_buffer(tree, start, len, mask);
2597         if (!eb)
2598                 return NULL;
2599
2600         if (page0) {
2601                 eb->first_page = page0;
2602                 i = 1;
2603                 index++;
2604                 page_cache_get(page0);
2605                 mark_page_accessed(page0);
2606                 set_page_extent_mapped(page0);
2607                 set_page_extent_head(page0, len);
2608                 uptodate = PageUptodate(page0);
2609         } else {
2610                 i = 0;
2611         }
2612         for (; i < num_pages; i++, index++) {
2613                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
2614                 if (!p) {
2615                         WARN_ON(1);
2616                         goto free_eb;
2617                 }
2618                 set_page_extent_mapped(p);
2619                 mark_page_accessed(p);
2620                 if (i == 0) {
2621                         eb->first_page = p;
2622                         set_page_extent_head(p, len);
2623                 } else {
2624                         set_page_private(p, EXTENT_PAGE_PRIVATE);
2625                 }
2626                 if (!PageUptodate(p))
2627                         uptodate = 0;
2628                 unlock_page(p);
2629         }
2630         if (uptodate)
2631                 eb->flags |= EXTENT_UPTODATE;
2632         eb->flags |= EXTENT_BUFFER_FILLED;
2633
2634         spin_lock(&tree->buffer_lock);
2635         exists = buffer_tree_insert(tree, start, &eb->rb_node);
2636         if (exists) {
2637                 /* add one reference for the caller */
2638                 atomic_inc(&exists->refs);
2639                 spin_unlock(&tree->buffer_lock);
2640                 goto free_eb;
2641         }
2642         spin_unlock(&tree->buffer_lock);
2643
2644         /* add one reference for the tree */
2645         atomic_inc(&eb->refs);
2646         return eb;
2647
2648 free_eb:
2649         if (!atomic_dec_and_test(&eb->refs))
2650                 return exists;
2651         for (index = 1; index < i; index++)
2652                 page_cache_release(extent_buffer_page(eb, index));
2653         page_cache_release(extent_buffer_page(eb, 0));
2654         __free_extent_buffer(eb);
2655         return exists;
2656 }
2657 EXPORT_SYMBOL(alloc_extent_buffer);
2658
2659 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
2660                                          u64 start, unsigned long len,
2661                                           gfp_t mask)
2662 {
2663         struct extent_buffer *eb;
2664
2665         spin_lock(&tree->buffer_lock);
2666         eb = buffer_search(tree, start);
2667         if (eb)
2668                 atomic_inc(&eb->refs);
2669         spin_unlock(&tree->buffer_lock);
2670
2671         if (eb)
2672                 mark_page_accessed(eb->first_page);
2673
2674         return eb;
2675 }
2676 EXPORT_SYMBOL(find_extent_buffer);
2677
2678 void free_extent_buffer(struct extent_buffer *eb)
2679 {
2680         if (!eb)
2681                 return;
2682
2683         if (!atomic_dec_and_test(&eb->refs))
2684                 return;
2685
2686         WARN_ON(1);
2687 }
2688 EXPORT_SYMBOL(free_extent_buffer);
2689
2690 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
2691                               struct extent_buffer *eb)
2692 {
2693         int set;
2694         unsigned long i;
2695         unsigned long num_pages;
2696         struct page *page;
2697
2698         u64 start = eb->start;
2699         u64 end = start + eb->len - 1;
2700
2701         set = clear_extent_dirty(tree, start, end, GFP_NOFS);
2702         num_pages = num_extent_pages(eb->start, eb->len);
2703
2704         for (i = 0; i < num_pages; i++) {
2705                 page = extent_buffer_page(eb, i);
2706                 lock_page(page);
2707                 if (i == 0)
2708                         set_page_extent_head(page, eb->len);
2709                 else
2710                         set_page_private(page, EXTENT_PAGE_PRIVATE);
2711
2712                 /*
2713                  * if we're on the last page or the first page and the
2714                  * block isn't aligned on a page boundary, do extra checks
2715                  * to make sure we don't clean page that is partially dirty
2716                  */
2717                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
2718                     ((i == num_pages - 1) &&
2719                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
2720                         start = (u64)page->index << PAGE_CACHE_SHIFT;
2721                         end  = start + PAGE_CACHE_SIZE - 1;
2722                         if (test_range_bit(tree, start, end,
2723                                            EXTENT_DIRTY, 0)) {
2724                                 unlock_page(page);
2725                                 continue;
2726                         }
2727                 }
2728                 clear_page_dirty_for_io(page);
2729                 spin_lock_irq(&page->mapping->tree_lock);
2730                 if (!PageDirty(page)) {
2731                         radix_tree_tag_clear(&page->mapping->page_tree,
2732                                                 page_index(page),
2733                                                 PAGECACHE_TAG_DIRTY);
2734                 }
2735                 spin_unlock_irq(&page->mapping->tree_lock);
2736                 unlock_page(page);
2737         }
2738         return 0;
2739 }
2740 EXPORT_SYMBOL(clear_extent_buffer_dirty);
2741
2742 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
2743                                     struct extent_buffer *eb)
2744 {
2745         return wait_on_extent_writeback(tree, eb->start,
2746                                         eb->start + eb->len - 1);
2747 }
2748 EXPORT_SYMBOL(wait_on_extent_buffer_writeback);
2749
2750 int set_extent_buffer_dirty(struct extent_io_tree *tree,
2751                              struct extent_buffer *eb)
2752 {
2753         unsigned long i;
2754         unsigned long num_pages;
2755
2756         num_pages = num_extent_pages(eb->start, eb->len);
2757         for (i = 0; i < num_pages; i++) {
2758                 struct page *page = extent_buffer_page(eb, i);
2759                 /* writepage may need to do something special for the
2760                  * first page, we have to make sure page->private is
2761                  * properly set.  releasepage may drop page->private
2762                  * on us if the page isn't already dirty.
2763                  */
2764                 lock_page(page);
2765                 if (i == 0) {
2766                         set_page_extent_head(page, eb->len);
2767                 } else if (PagePrivate(page) &&
2768                            page->private != EXTENT_PAGE_PRIVATE) {
2769                         set_page_extent_mapped(page);
2770                 }
2771                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
2772                 set_extent_dirty(tree, page_offset(page),
2773                                  page_offset(page) + PAGE_CACHE_SIZE -1,
2774                                  GFP_NOFS);
2775                 unlock_page(page);
2776         }
2777         return 0;
2778 }
2779 EXPORT_SYMBOL(set_extent_buffer_dirty);
2780
2781 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
2782                                 struct extent_buffer *eb)
2783 {
2784         unsigned long i;
2785         struct page *page;
2786         unsigned long num_pages;
2787
2788         num_pages = num_extent_pages(eb->start, eb->len);
2789         eb->flags &= ~EXTENT_UPTODATE;
2790
2791         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
2792                               GFP_NOFS);
2793         for (i = 0; i < num_pages; i++) {
2794                 page = extent_buffer_page(eb, i);
2795                 if (page)
2796                         ClearPageUptodate(page);
2797         }
2798         return 0;
2799 }
2800
2801 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
2802                                 struct extent_buffer *eb)
2803 {
2804         unsigned long i;
2805         struct page *page;
2806         unsigned long num_pages;
2807
2808         num_pages = num_extent_pages(eb->start, eb->len);
2809
2810         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
2811                             GFP_NOFS);
2812         for (i = 0; i < num_pages; i++) {
2813                 page = extent_buffer_page(eb, i);
2814                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
2815                     ((i == num_pages - 1) &&
2816                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
2817                         check_page_uptodate(tree, page);
2818                         continue;
2819                 }
2820                 SetPageUptodate(page);
2821         }
2822         return 0;
2823 }
2824 EXPORT_SYMBOL(set_extent_buffer_uptodate);
2825
2826 int extent_range_uptodate(struct extent_io_tree *tree,
2827                           u64 start, u64 end)
2828 {
2829         struct page *page;
2830         int ret;
2831         int pg_uptodate = 1;
2832         int uptodate;
2833         unsigned long index;
2834
2835         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1);
2836         if (ret)
2837                 return 1;
2838         while(start <= end) {
2839                 index = start >> PAGE_CACHE_SHIFT;
2840                 page = find_get_page(tree->mapping, index);
2841                 uptodate = PageUptodate(page);
2842                 page_cache_release(page);
2843                 if (!uptodate) {
2844                         pg_uptodate = 0;
2845                         break;
2846                 }
2847                 start += PAGE_CACHE_SIZE;
2848         }
2849         return pg_uptodate;
2850 }
2851
2852 int extent_buffer_uptodate(struct extent_io_tree *tree,
2853                            struct extent_buffer *eb)
2854 {
2855         int ret = 0;
2856         unsigned long num_pages;
2857         unsigned long i;
2858         struct page *page;
2859         int pg_uptodate = 1;
2860
2861         if (eb->flags & EXTENT_UPTODATE)
2862                 return 1;
2863
2864         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
2865                            EXTENT_UPTODATE, 1);
2866         if (ret)
2867                 return ret;
2868
2869         num_pages = num_extent_pages(eb->start, eb->len);
2870         for (i = 0; i < num_pages; i++) {
2871                 page = extent_buffer_page(eb, i);
2872                 if (!PageUptodate(page)) {
2873                         pg_uptodate = 0;
2874                         break;
2875                 }
2876         }
2877         return pg_uptodate;
2878 }
2879 EXPORT_SYMBOL(extent_buffer_uptodate);
2880
2881 int read_extent_buffer_pages(struct extent_io_tree *tree,
2882                              struct extent_buffer *eb,
2883                              u64 start, int wait,
2884                              get_extent_t *get_extent, int mirror_num)
2885 {
2886         unsigned long i;
2887         unsigned long start_i;
2888         struct page *page;
2889         int err;
2890         int ret = 0;
2891         int locked_pages = 0;
2892         int all_uptodate = 1;
2893         int inc_all_pages = 0;
2894         unsigned long num_pages;
2895         struct bio *bio = NULL;
2896
2897         if (eb->flags & EXTENT_UPTODATE)
2898                 return 0;
2899
2900         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
2901                            EXTENT_UPTODATE, 1)) {
2902                 return 0;
2903         }
2904
2905         if (start) {
2906                 WARN_ON(start < eb->start);
2907                 start_i = (start >> PAGE_CACHE_SHIFT) -
2908                         (eb->start >> PAGE_CACHE_SHIFT);
2909         } else {
2910                 start_i = 0;
2911         }
2912
2913         num_pages = num_extent_pages(eb->start, eb->len);
2914         for (i = start_i; i < num_pages; i++) {
2915                 page = extent_buffer_page(eb, i);
2916                 if (!wait) {
2917                         if (!trylock_page(page))
2918                                 goto unlock_exit;
2919                 } else {
2920                         lock_page(page);
2921                 }
2922                 locked_pages++;
2923                 if (!PageUptodate(page)) {
2924                         all_uptodate = 0;
2925                 }
2926         }
2927         if (all_uptodate) {
2928                 if (start_i == 0)
2929                         eb->flags |= EXTENT_UPTODATE;
2930                 if (ret) {
2931                         printk("all up to date but ret is %d\n", ret);
2932                 }
2933                 goto unlock_exit;
2934         }
2935
2936         for (i = start_i; i < num_pages; i++) {
2937                 page = extent_buffer_page(eb, i);
2938                 if (inc_all_pages)
2939                         page_cache_get(page);
2940                 if (!PageUptodate(page)) {
2941                         if (start_i == 0)
2942                                 inc_all_pages = 1;
2943                         ClearPageError(page);
2944                         err = __extent_read_full_page(tree, page,
2945                                                       get_extent, &bio,
2946                                                       mirror_num);
2947                         if (err) {
2948                                 ret = err;
2949                                 printk("err %d from __extent_read_full_page\n", ret);
2950                         }
2951                 } else {
2952                         unlock_page(page);
2953                 }
2954         }
2955
2956         if (bio)
2957                 submit_one_bio(READ, bio, mirror_num);
2958
2959         if (ret || !wait) {
2960                 if (ret)
2961                         printk("ret %d wait %d returning\n", ret, wait);
2962                 return ret;
2963         }
2964         for (i = start_i; i < num_pages; i++) {
2965                 page = extent_buffer_page(eb, i);
2966                 wait_on_page_locked(page);
2967                 if (!PageUptodate(page)) {
2968                         printk("page not uptodate after wait_on_page_locked\n");
2969                         ret = -EIO;
2970                 }
2971         }
2972         if (!ret)
2973                 eb->flags |= EXTENT_UPTODATE;
2974         return ret;
2975
2976 unlock_exit:
2977         i = start_i;
2978         while(locked_pages > 0) {
2979                 page = extent_buffer_page(eb, i);
2980                 i++;
2981                 unlock_page(page);
2982                 locked_pages--;
2983         }
2984         return ret;
2985 }
2986 EXPORT_SYMBOL(read_extent_buffer_pages);
2987
2988 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
2989                         unsigned long start,
2990                         unsigned long len)
2991 {
2992         size_t cur;
2993         size_t offset;
2994         struct page *page;
2995         char *kaddr;
2996         char *dst = (char *)dstv;
2997         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
2998         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
2999
3000         WARN_ON(start > eb->len);
3001         WARN_ON(start + len > eb->start + eb->len);
3002
3003         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3004
3005         while(len > 0) {
3006                 page = extent_buffer_page(eb, i);
3007
3008                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3009                 kaddr = kmap_atomic(page, KM_USER1);
3010                 memcpy(dst, kaddr + offset, cur);
3011                 kunmap_atomic(kaddr, KM_USER1);
3012
3013                 dst += cur;
3014                 len -= cur;
3015                 offset = 0;
3016                 i++;
3017         }
3018 }
3019 EXPORT_SYMBOL(read_extent_buffer);
3020
3021 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3022                                unsigned long min_len, char **token, char **map,
3023                                unsigned long *map_start,
3024                                unsigned long *map_len, int km)
3025 {
3026         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3027         char *kaddr;
3028         struct page *p;
3029         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3030         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3031         unsigned long end_i = (start_offset + start + min_len - 1) >>
3032                 PAGE_CACHE_SHIFT;
3033
3034         if (i != end_i)
3035                 return -EINVAL;
3036
3037         if (i == 0) {
3038                 offset = start_offset;
3039                 *map_start = 0;
3040         } else {
3041                 offset = 0;
3042                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3043         }
3044         if (start + min_len > eb->len) {
3045 printk("bad mapping eb start %Lu len %lu, wanted %lu %lu\n", eb->start, eb->len, start, min_len);
3046                 WARN_ON(1);
3047         }
3048
3049         p = extent_buffer_page(eb, i);
3050         kaddr = kmap_atomic(p, km);
3051         *token = kaddr;
3052         *map = kaddr + offset;
3053         *map_len = PAGE_CACHE_SIZE - offset;
3054         return 0;
3055 }
3056 EXPORT_SYMBOL(map_private_extent_buffer);
3057
3058 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3059                       unsigned long min_len,
3060                       char **token, char **map,
3061                       unsigned long *map_start,
3062                       unsigned long *map_len, int km)
3063 {
3064         int err;
3065         int save = 0;
3066         if (eb->map_token) {
3067                 unmap_extent_buffer(eb, eb->map_token, km);
3068                 eb->map_token = NULL;
3069                 save = 1;
3070         }
3071         err = map_private_extent_buffer(eb, start, min_len, token, map,
3072                                        map_start, map_len, km);
3073         if (!err && save) {
3074                 eb->map_token = *token;
3075                 eb->kaddr = *map;
3076                 eb->map_start = *map_start;
3077                 eb->map_len = *map_len;
3078         }
3079         return err;
3080 }
3081 EXPORT_SYMBOL(map_extent_buffer);
3082
3083 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3084 {
3085         kunmap_atomic(token, km);
3086 }
3087 EXPORT_SYMBOL(unmap_extent_buffer);
3088
3089 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3090                           unsigned long start,
3091                           unsigned long len)
3092 {
3093         size_t cur;
3094         size_t offset;
3095         struct page *page;
3096         char *kaddr;
3097         char *ptr = (char *)ptrv;
3098         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3099         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3100         int ret = 0;
3101
3102         WARN_ON(start > eb->len);
3103         WARN_ON(start + len > eb->start + eb->len);
3104
3105         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3106
3107         while(len > 0) {
3108                 page = extent_buffer_page(eb, i);
3109
3110                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3111
3112                 kaddr = kmap_atomic(page, KM_USER0);
3113                 ret = memcmp(ptr, kaddr + offset, cur);
3114                 kunmap_atomic(kaddr, KM_USER0);
3115                 if (ret)
3116                         break;
3117
3118                 ptr += cur;
3119                 len -= cur;
3120                 offset = 0;
3121                 i++;
3122         }
3123         return ret;
3124 }
3125 EXPORT_SYMBOL(memcmp_extent_buffer);
3126
3127 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3128                          unsigned long start, unsigned long len)
3129 {
3130         size_t cur;
3131         size_t offset;
3132         struct page *page;
3133         char *kaddr;
3134         char *src = (char *)srcv;
3135         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3136         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3137
3138         WARN_ON(start > eb->len);
3139         WARN_ON(start + len > eb->start + eb->len);
3140
3141         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3142
3143         while(len > 0) {
3144                 page = extent_buffer_page(eb, i);
3145                 WARN_ON(!PageUptodate(page));
3146
3147                 cur = min(len, PAGE_CACHE_SIZE - offset);
3148                 kaddr = kmap_atomic(page, KM_USER1);
3149                 memcpy(kaddr + offset, src, cur);
3150                 kunmap_atomic(kaddr, KM_USER1);
3151
3152                 src += cur;
3153                 len -= cur;
3154                 offset = 0;
3155                 i++;
3156         }
3157 }
3158 EXPORT_SYMBOL(write_extent_buffer);
3159
3160 void memset_extent_buffer(struct extent_buffer *eb, char c,
3161                           unsigned long start, unsigned long len)
3162 {
3163         size_t cur;
3164         size_t offset;
3165         struct page *page;
3166         char *kaddr;
3167         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3168         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3169
3170         WARN_ON(start > eb->len);
3171         WARN_ON(start + len > eb->start + eb->len);
3172
3173         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3174
3175         while(len > 0) {
3176                 page = extent_buffer_page(eb, i);
3177                 WARN_ON(!PageUptodate(page));
3178
3179                 cur = min(len, PAGE_CACHE_SIZE - offset);
3180                 kaddr = kmap_atomic(page, KM_USER0);
3181                 memset(kaddr + offset, c, cur);
3182                 kunmap_atomic(kaddr, KM_USER0);
3183
3184                 len -= cur;
3185                 offset = 0;
3186                 i++;
3187         }
3188 }
3189 EXPORT_SYMBOL(memset_extent_buffer);
3190
3191 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3192                         unsigned long dst_offset, unsigned long src_offset,
3193                         unsigned long len)
3194 {
3195         u64 dst_len = dst->len;
3196         size_t cur;
3197         size_t offset;
3198         struct page *page;
3199         char *kaddr;
3200         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3201         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3202
3203         WARN_ON(src->len != dst_len);
3204
3205         offset = (start_offset + dst_offset) &
3206                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3207
3208         while(len > 0) {
3209                 page = extent_buffer_page(dst, i);
3210                 WARN_ON(!PageUptodate(page));
3211
3212                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3213
3214                 kaddr = kmap_atomic(page, KM_USER0);
3215                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3216                 kunmap_atomic(kaddr, KM_USER0);
3217
3218                 src_offset += cur;
3219                 len -= cur;
3220                 offset = 0;
3221                 i++;
3222         }
3223 }
3224 EXPORT_SYMBOL(copy_extent_buffer);
3225
3226 static void move_pages(struct page *dst_page, struct page *src_page,
3227                        unsigned long dst_off, unsigned long src_off,
3228                        unsigned long len)
3229 {
3230         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3231         if (dst_page == src_page) {
3232                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3233         } else {
3234                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3235                 char *p = dst_kaddr + dst_off + len;
3236                 char *s = src_kaddr + src_off + len;
3237
3238                 while (len--)
3239                         *--p = *--s;
3240
3241                 kunmap_atomic(src_kaddr, KM_USER1);
3242         }
3243         kunmap_atomic(dst_kaddr, KM_USER0);
3244 }
3245
3246 static void copy_pages(struct page *dst_page, struct page *src_page,
3247                        unsigned long dst_off, unsigned long src_off,
3248                        unsigned long len)
3249 {
3250         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3251         char *src_kaddr;
3252
3253         if (dst_page != src_page)
3254                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3255         else
3256                 src_kaddr = dst_kaddr;
3257
3258         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3259         kunmap_atomic(dst_kaddr, KM_USER0);
3260         if (dst_page != src_page)
3261                 kunmap_atomic(src_kaddr, KM_USER1);
3262 }
3263
3264 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3265                            unsigned long src_offset, unsigned long len)
3266 {
3267         size_t cur;
3268         size_t dst_off_in_page;
3269         size_t src_off_in_page;
3270         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3271         unsigned long dst_i;
3272         unsigned long src_i;
3273
3274         if (src_offset + len > dst->len) {
3275                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
3276                        src_offset, len, dst->len);
3277                 BUG_ON(1);
3278         }
3279         if (dst_offset + len > dst->len) {
3280                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
3281                        dst_offset, len, dst->len);
3282                 BUG_ON(1);
3283         }
3284
3285         while(len > 0) {
3286                 dst_off_in_page = (start_offset + dst_offset) &
3287                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3288                 src_off_in_page = (start_offset + src_offset) &
3289                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3290
3291                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3292                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3293
3294                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3295                                                src_off_in_page));
3296                 cur = min_t(unsigned long, cur,
3297                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3298
3299                 copy_pages(extent_buffer_page(dst, dst_i),
3300                            extent_buffer_page(dst, src_i),
3301                            dst_off_in_page, src_off_in_page, cur);
3302
3303                 src_offset += cur;
3304                 dst_offset += cur;
3305                 len -= cur;
3306         }
3307 }
3308 EXPORT_SYMBOL(memcpy_extent_buffer);
3309
3310 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3311                            unsigned long src_offset, unsigned long len)
3312 {
3313         size_t cur;
3314         size_t dst_off_in_page;
3315         size_t src_off_in_page;
3316         unsigned long dst_end = dst_offset + len - 1;
3317         unsigned long src_end = src_offset + len - 1;
3318         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3319         unsigned long dst_i;
3320         unsigned long src_i;
3321
3322         if (src_offset + len > dst->len) {
3323                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
3324                        src_offset, len, dst->len);
3325                 BUG_ON(1);
3326         }
3327         if (dst_offset + len > dst->len) {
3328                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
3329                        dst_offset, len, dst->len);
3330                 BUG_ON(1);
3331         }
3332         if (dst_offset < src_offset) {
3333                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3334                 return;
3335         }
3336         while(len > 0) {
3337                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3338                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3339
3340                 dst_off_in_page = (start_offset + dst_end) &
3341                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3342                 src_off_in_page = (start_offset + src_end) &
3343                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3344
3345                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3346                 cur = min(cur, dst_off_in_page + 1);
3347                 move_pages(extent_buffer_page(dst, dst_i),
3348                            extent_buffer_page(dst, src_i),
3349                            dst_off_in_page - cur + 1,
3350                            src_off_in_page - cur + 1, cur);
3351
3352                 dst_end -= cur;
3353                 src_end -= cur;
3354                 len -= cur;
3355         }
3356 }
3357 EXPORT_SYMBOL(memmove_extent_buffer);
3358
3359 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3360 {
3361         u64 start = page_offset(page);
3362         struct extent_buffer *eb;
3363         int ret = 1;
3364         unsigned long i;
3365         unsigned long num_pages;
3366
3367         spin_lock(&tree->buffer_lock);
3368         eb = buffer_search(tree, start);
3369         if (!eb)
3370                 goto out;
3371
3372         if (atomic_read(&eb->refs) > 1) {
3373                 ret = 0;
3374                 goto out;
3375         }
3376         /* at this point we can safely release the extent buffer */
3377         num_pages = num_extent_pages(eb->start, eb->len);
3378         for (i = 0; i < num_pages; i++)
3379                 page_cache_release(extent_buffer_page(eb, i));
3380         rb_erase(&eb->rb_node, &tree->buffer);
3381         __free_extent_buffer(eb);
3382 out:
3383         spin_unlock(&tree->buffer_lock);
3384         return ret;
3385 }
3386 EXPORT_SYMBOL(try_release_extent_buffer);