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