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