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 static 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 static 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 static 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 static 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
 888 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 889                         gfp_t mask)
 890 {
 891         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, NULL,
 892                               mask);
 893 }
 894 EXPORT_SYMBOL(set_extent_uptodate);
 895
 896 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 897                           gfp_t mask)
 898 {
 899         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0, mask);
 900 }
 901
 902 static int set_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end,
 903                          gfp_t mask)
 904 {
 905         return set_extent_bit(tree, start, end, EXTENT_WRITEBACK,
 906                               0, NULL, mask);
 907 }
 908
 909 static int clear_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end,
 910                            gfp_t mask)
 911 {
 912         return clear_extent_bit(tree, start, end, EXTENT_WRITEBACK, 1, 0, mask);
 913 }
 914
 915 int wait_on_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 916 {
 917         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 918 }
 919 EXPORT_SYMBOL(wait_on_extent_writeback);
 920
 921 /*
 922  * either insert or lock state struct between start and end use mask to tell
 923  * us if waiting is desired.
 924  */
 925 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
 926 {
 927         int err;
 928         u64 failed_start;
 929         while (1) {
 930                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED, 1,
 931                                      &failed_start, mask);
 932                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
 933                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
 934                         start = failed_start;
 935                 } else {
 936                         break;
 937                 }
 938                 WARN_ON(start > end);
 939         }
 940         return err;
 941 }
 942 EXPORT_SYMBOL(lock_extent);
 943
 944 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
 945                     gfp_t mask)
 946 {
 947         int err;
 948         u64 failed_start;
 949
 950         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, 1,
 951                              &failed_start, mask);
 952         if (err == -EEXIST) {
 953                 if (failed_start > start)
 954                         clear_extent_bit(tree, start, failed_start - 1,
 955                                          EXTENT_LOCKED, 1, 0, mask);
 956                 return 0;
 957         }
 958         return 1;
 959 }
 960 EXPORT_SYMBOL(try_lock_extent);
 961
 962 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end,
 963                   gfp_t mask)
 964 {
 965         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, mask);
 966 }
 967 EXPORT_SYMBOL(unlock_extent);
 968
 969 /*
 970  * helper function to set pages and extents in the tree dirty
 971  */
 972 int set_range_dirty(struct extent_io_tree *tree, u64 start, u64 end)
 973 {
 974         unsigned long index = start >> PAGE_CACHE_SHIFT;
 975         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 976         struct page *page;
 977
 978         while (index <= end_index) {
 979                 page = find_get_page(tree->mapping, index);
 980                 BUG_ON(!page);
 981                 __set_page_dirty_nobuffers(page);
 982                 page_cache_release(page);
 983                 index++;
 984         }
 985         set_extent_dirty(tree, start, end, GFP_NOFS);
 986         return 0;
 987 }
 988 EXPORT_SYMBOL(set_range_dirty);
 989
 990 /*
 991  * helper function to set both pages and extents in the tree writeback
 992  */
 993 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 994 {
 995         unsigned long index = start >> PAGE_CACHE_SHIFT;
 996         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 997         struct page *page;
 998
 999         while (index <= end_index) {
1000                 page = find_get_page(tree->mapping, index);
1001                 BUG_ON(!page);
1002                 set_page_writeback(page);
1003                 page_cache_release(page);
1004                 index++;
1005         }
1006         set_extent_writeback(tree, start, end, GFP_NOFS);
1007         return 0;
1008 }
1009
1010 /*
1011  * find the first offset in the io tree with 'bits' set. zero is
1012  * returned if we find something, and *start_ret and *end_ret are
1013  * set to reflect the state struct that was found.
1014  *
1015  * If nothing was found, 1 is returned, < 0 on error
1016  */
1017 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1018                           u64 *start_ret, u64 *end_ret, int bits)
1019 {
1020         struct rb_node *node;
1021         struct extent_state *state;
1022         int ret = 1;
1023
1024         spin_lock_irq(&tree->lock);
1025         /*
1026          * this search will find all the extents that end after
1027          * our range starts.
1028          */
1029         node = tree_search(tree, start);
1030         if (!node) {
1031                 goto out;
1032         }
1033
1034         while(1) {
1035                 state = rb_entry(node, struct extent_state, rb_node);
1036                 if (state->end >= start && (state->state & bits)) {
1037                         *start_ret = state->start;
1038                         *end_ret = state->end;
1039                         ret = 0;
1040                         break;
1041                 }
1042                 node = rb_next(node);
1043                 if (!node)
1044                         break;
1045         }
1046 out:
1047         spin_unlock_irq(&tree->lock);
1048         return ret;
1049 }
1050 EXPORT_SYMBOL(find_first_extent_bit);
1051
1052 /* find the first state struct with 'bits' set after 'start', and
1053  * return it.  tree->lock must be held.  NULL will returned if
1054  * nothing was found after 'start'
1055  */
1056 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1057                                                  u64 start, int bits)
1058 {
1059         struct rb_node *node;
1060         struct extent_state *state;
1061
1062         /*
1063          * this search will find all the extents that end after
1064          * our range starts.
1065          */
1066         node = tree_search(tree, start);
1067         if (!node) {
1068                 goto out;
1069         }
1070
1071         while(1) {
1072                 state = rb_entry(node, struct extent_state, rb_node);
1073                 if (state->end >= start && (state->state & bits)) {
1074                         return state;
1075                 }
1076                 node = rb_next(node);
1077                 if (!node)
1078                         break;
1079         }
1080 out:
1081         return NULL;
1082 }
1083 EXPORT_SYMBOL(find_first_extent_bit_state);
1084
1085 /*
1086  * find a contiguous range of bytes in the file marked as delalloc, not
1087  * more than 'max_bytes'.  start and end are used to return the range,
1088  *
1089  * 1 is returned if we find something, 0 if nothing was in the tree
1090  */
1091 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1092                                         u64 *start, u64 *end, u64 max_bytes)
1093 {
1094         struct rb_node *node;
1095         struct extent_state *state;
1096         u64 cur_start = *start;
1097         u64 found = 0;
1098         u64 total_bytes = 0;
1099
1100         spin_lock_irq(&tree->lock);
1101
1102         /*
1103          * this search will find all the extents that end after
1104          * our range starts.
1105          */
1106         node = tree_search(tree, cur_start);
1107         if (!node) {
1108                 if (!found)
1109                         *end = (u64)-1;
1110                 goto out;
1111         }
1112
1113         while(1) {
1114                 state = rb_entry(node, struct extent_state, rb_node);
1115                 if (found && (state->start != cur_start ||
1116                               (state->state & EXTENT_BOUNDARY))) {
1117                         goto out;
1118                 }
1119                 if (!(state->state & EXTENT_DELALLOC)) {
1120                         if (!found)
1121                                 *end = state->end;
1122                         goto out;
1123                 }
1124                 if (!found)
1125                         *start = state->start;
1126                 found++;
1127                 *end = state->end;
1128                 cur_start = state->end + 1;
1129                 node = rb_next(node);
1130                 if (!node)
1131                         break;
1132                 total_bytes += state->end - state->start + 1;
1133                 if (total_bytes >= max_bytes)
1134                         break;
1135         }
1136 out:
1137         spin_unlock_irq(&tree->lock);
1138         return found;
1139 }
1140
1141 static noinline int __unlock_for_delalloc(struct inode *inode,
1142                                           struct page *locked_page,
1143                                           u64 start, u64 end)
1144 {
1145         int ret;
1146         struct page *pages[16];
1147         unsigned long index = start >> PAGE_CACHE_SHIFT;
1148         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1149         unsigned long nr_pages = end_index - index + 1;
1150         int i;
1151
1152         if (index == locked_page->index && end_index == index)
1153                 return 0;
1154
1155         while(nr_pages > 0) {
1156                 ret = find_get_pages_contig(inode->i_mapping, index,
1157                                      min_t(unsigned long, nr_pages,
1158                                      ARRAY_SIZE(pages)), pages);
1159                 for (i = 0; i < ret; i++) {
1160                         if (pages[i] != locked_page)
1161                                 unlock_page(pages[i]);
1162                         page_cache_release(pages[i]);
1163                 }
1164                 nr_pages -= ret;
1165                 index += ret;
1166                 cond_resched();
1167         }
1168         return 0;
1169 }
1170
1171 static noinline int lock_delalloc_pages(struct inode *inode,
1172                                         struct page *locked_page,
1173                                         u64 delalloc_start,
1174                                         u64 delalloc_end)
1175 {
1176         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1177         unsigned long start_index = index;
1178         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1179         unsigned long pages_locked = 0;
1180         struct page *pages[16];
1181         unsigned long nrpages;
1182         int ret;
1183         int i;
1184
1185         /* the caller is responsible for locking the start index */
1186         if (index == locked_page->index && index == end_index)
1187                 return 0;
1188
1189         /* skip the page at the start index */
1190         nrpages = end_index - index + 1;
1191         while(nrpages > 0) {
1192                 ret = find_get_pages_contig(inode->i_mapping, index,
1193                                      min_t(unsigned long,
1194                                      nrpages, ARRAY_SIZE(pages)), pages);
1195                 if (ret == 0) {
1196                         ret = -EAGAIN;
1197                         goto done;
1198                 }
1199                 /* now we have an array of pages, lock them all */
1200                 for (i = 0; i < ret; i++) {
1201                         /*
1202                          * the caller is taking responsibility for
1203                          * locked_page
1204                          */
1205                         if (pages[i] != locked_page) {
1206                                 lock_page(pages[i]);
1207                                 if (!PageDirty(pages[i]) ||
1208                                     pages[i]->mapping != inode->i_mapping) {
1209                                         ret = -EAGAIN;
1210                                         unlock_page(pages[i]);
1211                                         page_cache_release(pages[i]);
1212                                         goto done;
1213                                 }
1214                         }
1215                         page_cache_release(pages[i]);
1216                         pages_locked++;
1217                 }
1218                 nrpages -= ret;
1219                 index += ret;
1220                 cond_resched();
1221         }
1222         ret = 0;
1223 done:
1224         if (ret && pages_locked) {
1225                 __unlock_for_delalloc(inode, locked_page,
1226                               delalloc_start,
1227                               ((u64)(start_index + pages_locked - 1)) <<
1228                               PAGE_CACHE_SHIFT);
1229         }
1230         return ret;
1231 }
1232
1233 /*
1234  * find a contiguous range of bytes in the file marked as delalloc, not
1235  * more than 'max_bytes'.  start and end are used to return the range,
1236  *
1237  * 1 is returned if we find something, 0 if nothing was in the tree
1238  */
1239 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1240                                              struct extent_io_tree *tree,
1241                                              struct page *locked_page,
1242                                              u64 *start, u64 *end,
1243                                              u64 max_bytes)
1244 {
1245         u64 delalloc_start;
1246         u64 delalloc_end;
1247         u64 found;
1248         int ret;
1249         int loops = 0;
1250
1251 again:
1252         /* step one, find a bunch of delalloc bytes starting at start */
1253         delalloc_start = *start;
1254         delalloc_end = 0;
1255         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1256                                     max_bytes);
1257         if (!found || delalloc_end <= *start) {
1258                 *start = delalloc_start;
1259                 *end = delalloc_end;
1260                 return found;
1261         }
1262
1263         /*
1264          * start comes from the offset of locked_page.  We have to lock
1265          * pages in order, so we can't process delalloc bytes before
1266          * locked_page
1267          */
1268         if (delalloc_start < *start) {
1269                 delalloc_start = *start;
1270         }
1271
1272         /*
1273          * make sure to limit the number of pages we try to lock down
1274          * if we're looping.
1275          */
1276         if (delalloc_end + 1 - delalloc_start > max_bytes && loops) {
1277                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1278         }
1279         /* step two, lock all the pages after the page that has start */
1280         ret = lock_delalloc_pages(inode, locked_page,
1281                                   delalloc_start, delalloc_end);
1282         if (ret == -EAGAIN) {
1283                 /* some of the pages are gone, lets avoid looping by
1284                  * shortening the size of the delalloc range we're searching
1285                  */
1286                 if (!loops) {
1287                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1288                         max_bytes = PAGE_CACHE_SIZE - offset;
1289                         loops = 1;
1290                         goto again;
1291                 } else {
1292                         found = 0;
1293                         goto out_failed;
1294                 }
1295         }
1296         BUG_ON(ret);
1297
1298         /* step three, lock the state bits for the whole range */
1299         lock_extent(tree, delalloc_start, delalloc_end, GFP_NOFS);
1300
1301         /* then test to make sure it is all still delalloc */
1302         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1303                              EXTENT_DELALLOC, 1);
1304         if (!ret) {
1305                 unlock_extent(tree, delalloc_start, delalloc_end, GFP_NOFS);
1306                 __unlock_for_delalloc(inode, locked_page,
1307                               delalloc_start, delalloc_end);
1308                 cond_resched();
1309                 goto again;
1310         }
1311         *start = delalloc_start;
1312         *end = delalloc_end;
1313 out_failed:
1314         return found;
1315 }
1316
1317 int extent_clear_unlock_delalloc(struct inode *inode,
1318                                 struct extent_io_tree *tree,
1319                                 u64 start, u64 end, struct page *locked_page,
1320                                 int unlock_pages,
1321                                 int clear_unlock,
1322                                 int clear_delalloc, int clear_dirty,
1323                                 int set_writeback,
1324                                 int end_writeback)
1325 {
1326         int ret;
1327         struct page *pages[16];
1328         unsigned long index = start >> PAGE_CACHE_SHIFT;
1329         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1330         unsigned long nr_pages = end_index - index + 1;
1331         int i;
1332         int clear_bits = 0;
1333
1334         if (clear_unlock)
1335                 clear_bits |= EXTENT_LOCKED;
1336         if (clear_dirty)
1337                 clear_bits |= EXTENT_DIRTY;
1338
1339         if (clear_delalloc)
1340                 clear_bits |= EXTENT_DELALLOC;
1341
1342         clear_extent_bit(tree, start, end, clear_bits, 1, 0, GFP_NOFS);
1343         if (!(unlock_pages || clear_dirty || set_writeback || end_writeback))
1344                 return 0;
1345
1346         while(nr_pages > 0) {
1347                 ret = find_get_pages_contig(inode->i_mapping, index,
1348                                      min_t(unsigned long,
1349                                      nr_pages, ARRAY_SIZE(pages)), pages);
1350                 for (i = 0; i < ret; i++) {
1351                         if (pages[i] == locked_page) {
1352                                 page_cache_release(pages[i]);
1353                                 continue;
1354                         }
1355                         if (clear_dirty)
1356                                 clear_page_dirty_for_io(pages[i]);
1357                         if (set_writeback)
1358                                 set_page_writeback(pages[i]);
1359                         if (end_writeback)
1360                                 end_page_writeback(pages[i]);
1361                         if (unlock_pages)
1362                                 unlock_page(pages[i]);
1363                         page_cache_release(pages[i]);
1364                 }
1365                 nr_pages -= ret;
1366                 index += ret;
1367                 cond_resched();
1368         }
1369         return 0;
1370 }
1371 EXPORT_SYMBOL(extent_clear_unlock_delalloc);
1372
1373 /*
1374  * count the number of bytes in the tree that have a given bit(s)
1375  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1376  * cached.  The total number found is returned.
1377  */
1378 u64 count_range_bits(struct extent_io_tree *tree,
1379                      u64 *start, u64 search_end, u64 max_bytes,
1380                      unsigned long bits)
1381 {
1382         struct rb_node *node;
1383         struct extent_state *state;
1384         u64 cur_start = *start;
1385         u64 total_bytes = 0;
1386         int found = 0;
1387
1388         if (search_end <= cur_start) {
1389                 printk("search_end %Lu start %Lu\n", search_end, cur_start);
1390                 WARN_ON(1);
1391                 return 0;
1392         }
1393
1394         spin_lock_irq(&tree->lock);
1395         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1396                 total_bytes = tree->dirty_bytes;
1397                 goto out;
1398         }
1399         /*
1400          * this search will find all the extents that end after
1401          * our range starts.
1402          */
1403         node = tree_search(tree, cur_start);
1404         if (!node) {
1405                 goto out;
1406         }
1407
1408         while(1) {
1409                 state = rb_entry(node, struct extent_state, rb_node);
1410                 if (state->start > search_end)
1411                         break;
1412                 if (state->end >= cur_start && (state->state & bits)) {
1413                         total_bytes += min(search_end, state->end) + 1 -
1414                                        max(cur_start, state->start);
1415                         if (total_bytes >= max_bytes)
1416                                 break;
1417                         if (!found) {
1418                                 *start = state->start;
1419                                 found = 1;
1420                         }
1421                 }
1422                 node = rb_next(node);
1423                 if (!node)
1424                         break;
1425         }
1426 out:
1427         spin_unlock_irq(&tree->lock);
1428         return total_bytes;
1429 }
1430
1431 #if 0
1432 /*
1433  * helper function to lock both pages and extents in the tree.
1434  * pages must be locked first.
1435  */
1436 static int lock_range(struct extent_io_tree *tree, u64 start, u64 end)
1437 {
1438         unsigned long index = start >> PAGE_CACHE_SHIFT;
1439         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1440         struct page *page;
1441         int err;
1442
1443         while (index <= end_index) {
1444                 page = grab_cache_page(tree->mapping, index);
1445                 if (!page) {
1446                         err = -ENOMEM;
1447                         goto failed;
1448                 }
1449                 if (IS_ERR(page)) {
1450                         err = PTR_ERR(page);
1451                         goto failed;
1452                 }
1453                 index++;
1454         }
1455         lock_extent(tree, start, end, GFP_NOFS);
1456         return 0;
1457
1458 failed:
1459         /*
1460          * we failed above in getting the page at 'index', so we undo here
1461          * up to but not including the page at 'index'
1462          */
1463         end_index = index;
1464         index = start >> PAGE_CACHE_SHIFT;
1465         while (index < end_index) {
1466                 page = find_get_page(tree->mapping, index);
1467                 unlock_page(page);
1468                 page_cache_release(page);
1469                 index++;
1470         }
1471         return err;
1472 }
1473
1474 /*
1475  * helper function to unlock both pages and extents in the tree.
1476  */
1477 static int unlock_range(struct extent_io_tree *tree, u64 start, u64 end)
1478 {
1479         unsigned long index = start >> PAGE_CACHE_SHIFT;
1480         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1481         struct page *page;
1482
1483         while (index <= end_index) {
1484                 page = find_get_page(tree->mapping, index);
1485                 unlock_page(page);
1486                 page_cache_release(page);
1487                 index++;
1488         }
1489         unlock_extent(tree, start, end, GFP_NOFS);
1490         return 0;
1491 }
1492 #endif
1493
1494 /*
1495  * set the private field for a given byte offset in the tree.  If there isn't
1496  * an extent_state there already, this does nothing.
1497  */
1498 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1499 {
1500         struct rb_node *node;
1501         struct extent_state *state;
1502         int ret = 0;
1503
1504         spin_lock_irq(&tree->lock);
1505         /*
1506          * this search will find all the extents that end after
1507          * our range starts.
1508          */
1509         node = tree_search(tree, start);
1510         if (!node) {
1511                 ret = -ENOENT;
1512                 goto out;
1513         }
1514         state = rb_entry(node, struct extent_state, rb_node);
1515         if (state->start != start) {
1516                 ret = -ENOENT;
1517                 goto out;
1518         }
1519         state->private = private;
1520 out:
1521         spin_unlock_irq(&tree->lock);
1522         return ret;
1523 }
1524
1525 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1526 {
1527         struct rb_node *node;
1528         struct extent_state *state;
1529         int ret = 0;
1530
1531         spin_lock_irq(&tree->lock);
1532         /*
1533          * this search will find all the extents that end after
1534          * our range starts.
1535          */
1536         node = tree_search(tree, start);
1537         if (!node) {
1538                 ret = -ENOENT;
1539                 goto out;
1540         }
1541         state = rb_entry(node, struct extent_state, rb_node);
1542         if (state->start != start) {
1543                 ret = -ENOENT;
1544                 goto out;
1545         }
1546         *private = state->private;
1547 out:
1548         spin_unlock_irq(&tree->lock);
1549         return ret;
1550 }
1551
1552 /*
1553  * searches a range in the state tree for a given mask.
1554  * If 'filled' == 1, this returns 1 only if every extent in the tree
1555  * has the bits set.  Otherwise, 1 is returned if any bit in the
1556  * range is found set.
1557  */
1558 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1559                    int bits, int filled)
1560 {
1561         struct extent_state *state = NULL;
1562         struct rb_node *node;
1563         int bitset = 0;
1564         unsigned long flags;
1565
1566         spin_lock_irqsave(&tree->lock, flags);
1567         node = tree_search(tree, start);
1568         while (node && start <= end) {
1569                 state = rb_entry(node, struct extent_state, rb_node);
1570
1571                 if (filled && state->start > start) {
1572                         bitset = 0;
1573                         break;
1574                 }
1575
1576                 if (state->start > end)
1577                         break;
1578
1579                 if (state->state & bits) {
1580                         bitset = 1;
1581                         if (!filled)
1582                                 break;
1583                 } else if (filled) {
1584                         bitset = 0;
1585                         break;
1586                 }
1587                 start = state->end + 1;
1588                 if (start > end)
1589                         break;
1590                 node = rb_next(node);
1591                 if (!node) {
1592                         if (filled)
1593                                 bitset = 0;
1594                         break;
1595                 }
1596         }
1597         spin_unlock_irqrestore(&tree->lock, flags);
1598         return bitset;
1599 }
1600 EXPORT_SYMBOL(test_range_bit);
1601
1602 /*
1603  * helper function to set a given page up to date if all the
1604  * extents in the tree for that page are up to date
1605  */
1606 static int check_page_uptodate(struct extent_io_tree *tree,
1607                                struct page *page)
1608 {
1609         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1610         u64 end = start + PAGE_CACHE_SIZE - 1;
1611         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1))
1612                 SetPageUptodate(page);
1613         return 0;
1614 }
1615
1616 /*
1617  * helper function to unlock a page if all the extents in the tree
1618  * for that page are unlocked
1619  */
1620 static int check_page_locked(struct extent_io_tree *tree,
1621                              struct page *page)
1622 {
1623         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1624         u64 end = start + PAGE_CACHE_SIZE - 1;
1625         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0))
1626                 unlock_page(page);
1627         return 0;
1628 }
1629
1630 /*
1631  * helper function to end page writeback if all the extents
1632  * in the tree for that page are done with writeback
1633  */
1634 static int check_page_writeback(struct extent_io_tree *tree,
1635                              struct page *page)
1636 {
1637         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1638         u64 end = start + PAGE_CACHE_SIZE - 1;
1639         if (!test_range_bit(tree, start, end, EXTENT_WRITEBACK, 0))
1640                 end_page_writeback(page);
1641         return 0;
1642 }
1643
1644 /* lots and lots of room for performance fixes in the end_bio funcs */
1645
1646 /*
1647  * after a writepage IO is done, we need to:
1648  * clear the uptodate bits on error
1649  * clear the writeback bits in the extent tree for this IO
1650  * end_page_writeback if the page has no more pending IO
1651  *
1652  * Scheduling is not allowed, so the extent state tree is expected
1653  * to have one and only one object corresponding to this IO.
1654  */
1655 static void end_bio_extent_writepage(struct bio *bio, int err)
1656 {
1657         int uptodate = err == 0;
1658         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1659         struct extent_io_tree *tree;
1660         u64 start;
1661         u64 end;
1662         int whole_page;
1663         int ret;
1664
1665         do {
1666                 struct page *page = bvec->bv_page;
1667                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1668
1669                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1670                          bvec->bv_offset;
1671                 end = start + bvec->bv_len - 1;
1672
1673                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1674                         whole_page = 1;
1675                 else
1676                         whole_page = 0;
1677
1678                 if (--bvec >= bio->bi_io_vec)
1679                         prefetchw(&bvec->bv_page->flags);
1680                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1681                         ret = tree->ops->writepage_end_io_hook(page, start,
1682                                                        end, NULL, uptodate);
1683                         if (ret)
1684                                 uptodate = 0;
1685                 }
1686
1687                 if (!uptodate && tree->ops &&
1688                     tree->ops->writepage_io_failed_hook) {
1689                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1690                                                          start, end, NULL);
1691                         if (ret == 0) {
1692                                 uptodate = (err == 0);
1693                                 continue;
1694                         }
1695                 }
1696
1697                 if (!uptodate) {
1698                         clear_extent_uptodate(tree, start, end, GFP_ATOMIC);
1699                         ClearPageUptodate(page);
1700                         SetPageError(page);
1701                 }
1702
1703                 clear_extent_writeback(tree, start, end, GFP_ATOMIC);
1704
1705                 if (whole_page)
1706                         end_page_writeback(page);
1707                 else
1708                         check_page_writeback(tree, page);
1709         } while (bvec >= bio->bi_io_vec);
1710
1711         bio_put(bio);
1712 }
1713
1714 /*
1715  * after a readpage IO is done, we need to:
1716  * clear the uptodate bits on error
1717  * set the uptodate bits if things worked
1718  * set the page up to date if all extents in the tree are uptodate
1719  * clear the lock bit in the extent tree
1720  * unlock the page if there are no other extents locked for it
1721  *
1722  * Scheduling is not allowed, so the extent state tree is expected
1723  * to have one and only one object corresponding to this IO.
1724  */
1725 static void end_bio_extent_readpage(struct bio *bio, int err)
1726 {
1727         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1728         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1729         struct extent_io_tree *tree;
1730         u64 start;
1731         u64 end;
1732         int whole_page;
1733         int ret;
1734
1735         if (err)
1736                 uptodate = 0;
1737
1738         do {
1739                 struct page *page = bvec->bv_page;
1740                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1741
1742                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1743                         bvec->bv_offset;
1744                 end = start + bvec->bv_len - 1;
1745
1746                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1747                         whole_page = 1;
1748                 else
1749                         whole_page = 0;
1750
1751                 if (--bvec >= bio->bi_io_vec)
1752                         prefetchw(&bvec->bv_page->flags);
1753
1754                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1755                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1756                                                               NULL);
1757                         if (ret)
1758                                 uptodate = 0;
1759                 }
1760                 if (!uptodate && tree->ops &&
1761                     tree->ops->readpage_io_failed_hook) {
1762                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1763                                                          start, end, NULL);
1764                         if (ret == 0) {
1765                                 uptodate =
1766                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1767                                 if (err)
1768                                         uptodate = 0;
1769                                 continue;
1770                         }
1771                 }
1772
1773                 if (uptodate) {
1774                         set_extent_uptodate(tree, start, end,
1775                                             GFP_ATOMIC);
1776                 }
1777                 unlock_extent(tree, start, end, GFP_ATOMIC);
1778
1779                 if (whole_page) {
1780                         if (uptodate) {
1781                                 SetPageUptodate(page);
1782                         } else {
1783                                 ClearPageUptodate(page);
1784                                 SetPageError(page);
1785                         }
1786                         unlock_page(page);
1787                 } else {
1788                         if (uptodate) {
1789                                 check_page_uptodate(tree, page);
1790                         } else {
1791                                 ClearPageUptodate(page);
1792                                 SetPageError(page);
1793                         }
1794                         check_page_locked(tree, page);
1795                 }
1796         } while (bvec >= bio->bi_io_vec);
1797
1798         bio_put(bio);
1799 }
1800
1801 /*
1802  * IO done from prepare_write is pretty simple, we just unlock
1803  * the structs in the extent tree when done, and set the uptodate bits
1804  * as appropriate.
1805  */
1806 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1807 {
1808         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1809         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1810         struct extent_io_tree *tree;
1811         u64 start;
1812         u64 end;
1813
1814         do {
1815                 struct page *page = bvec->bv_page;
1816                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1817
1818                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1819                         bvec->bv_offset;
1820                 end = start + bvec->bv_len - 1;
1821
1822                 if (--bvec >= bio->bi_io_vec)
1823                         prefetchw(&bvec->bv_page->flags);
1824
1825                 if (uptodate) {
1826                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1827                 } else {
1828                         ClearPageUptodate(page);
1829                         SetPageError(page);
1830                 }
1831
1832                 unlock_extent(tree, start, end, GFP_ATOMIC);
1833
1834         } while (bvec >= bio->bi_io_vec);
1835
1836         bio_put(bio);
1837 }
1838
1839 static struct bio *
1840 extent_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1841                  gfp_t gfp_flags)
1842 {
1843         struct bio *bio;
1844
1845         bio = bio_alloc(gfp_flags, nr_vecs);
1846
1847         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1848                 while (!bio && (nr_vecs /= 2))
1849                         bio = bio_alloc(gfp_flags, nr_vecs);
1850         }
1851
1852         if (bio) {
1853                 bio->bi_size = 0;
1854                 bio->bi_bdev = bdev;
1855                 bio->bi_sector = first_sector;
1856         }
1857         return bio;
1858 }
1859
1860 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1861                           unsigned long bio_flags)
1862 {
1863         int ret = 0;
1864         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1865         struct page *page = bvec->bv_page;
1866         struct extent_io_tree *tree = bio->bi_private;
1867         u64 start;
1868         u64 end;
1869
1870         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1871         end = start + bvec->bv_len - 1;
1872
1873         bio->bi_private = NULL;
1874
1875         bio_get(bio);
1876
1877         if (tree->ops && tree->ops->submit_bio_hook)
1878                 tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1879                                            mirror_num, bio_flags);
1880         else
1881                 submit_bio(rw, bio);
1882         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1883                 ret = -EOPNOTSUPP;
1884         bio_put(bio);
1885         return ret;
1886 }
1887
1888 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1889                               struct page *page, sector_t sector,
1890                               size_t size, unsigned long offset,
1891                               struct block_device *bdev,
1892                               struct bio **bio_ret,
1893                               unsigned long max_pages,
1894                               bio_end_io_t end_io_func,
1895                               int mirror_num,
1896                               unsigned long prev_bio_flags,
1897                               unsigned long bio_flags)
1898 {
1899         int ret = 0;
1900         struct bio *bio;
1901         int nr;
1902         int contig = 0;
1903         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1904         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1905         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1906
1907         if (bio_ret && *bio_ret) {
1908                 bio = *bio_ret;
1909                 if (old_compressed)
1910                         contig = bio->bi_sector == sector;
1911                 else
1912                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
1913                                 sector;
1914
1915                 if (prev_bio_flags != bio_flags || !contig ||
1916                     (tree->ops && tree->ops->merge_bio_hook &&
1917                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
1918                                                bio_flags)) ||
1919                     bio_add_page(bio, page, page_size, offset) < page_size) {
1920                         ret = submit_one_bio(rw, bio, mirror_num,
1921                                              prev_bio_flags);
1922                         bio = NULL;
1923                 } else {
1924                         return 0;
1925                 }
1926         }
1927         if (this_compressed)
1928                 nr = BIO_MAX_PAGES;
1929         else
1930                 nr = bio_get_nr_vecs(bdev);
1931
1932         bio = extent_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1933         if (!bio) {
1934                 printk("failed to allocate bio nr %d\n", nr);
1935         }
1936
1937         bio_add_page(bio, page, page_size, offset);
1938         bio->bi_end_io = end_io_func;
1939         bio->bi_private = tree;
1940
1941         if (bio_ret) {
1942                 *bio_ret = bio;
1943         } else {
1944                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1945         }
1946
1947         return ret;
1948 }
1949
1950 void set_page_extent_mapped(struct page *page)
1951 {
1952         if (!PagePrivate(page)) {
1953                 SetPagePrivate(page);
1954                 page_cache_get(page);
1955                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1956         }
1957 }
1958 EXPORT_SYMBOL(set_page_extent_mapped);
1959
1960 static void set_page_extent_head(struct page *page, unsigned long len)
1961 {
1962         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1963 }
1964
1965 /*
1966  * basic readpage implementation.  Locked extent state structs are inserted
1967  * into the tree that are removed when the IO is done (by the end_io
1968  * handlers)
1969  */
1970 static int __extent_read_full_page(struct extent_io_tree *tree,
1971                                    struct page *page,
1972                                    get_extent_t *get_extent,
1973                                    struct bio **bio, int mirror_num,
1974                                    unsigned long *bio_flags)
1975 {
1976         struct inode *inode = page->mapping->host;
1977         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1978         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1979         u64 end;
1980         u64 cur = start;
1981         u64 extent_offset;
1982         u64 last_byte = i_size_read(inode);
1983         u64 block_start;
1984         u64 cur_end;
1985         sector_t sector;
1986         struct extent_map *em;
1987         struct block_device *bdev;
1988         int ret;
1989         int nr = 0;
1990         size_t page_offset = 0;
1991         size_t iosize;
1992         size_t disk_io_size;
1993         size_t blocksize = inode->i_sb->s_blocksize;
1994         unsigned long this_bio_flag = 0;
1995
1996         set_page_extent_mapped(page);
1997
1998         end = page_end;
1999         lock_extent(tree, start, end, GFP_NOFS);
2000
2001         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2002                 char *userpage;
2003                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2004
2005                 if (zero_offset) {
2006                         iosize = PAGE_CACHE_SIZE - zero_offset;
2007                         userpage = kmap_atomic(page, KM_USER0);
2008                         memset(userpage + zero_offset, 0, iosize);
2009                         flush_dcache_page(page);
2010                         kunmap_atomic(userpage, KM_USER0);
2011                 }
2012         }
2013         while (cur <= end) {
2014                 if (cur >= last_byte) {
2015                         char *userpage;
2016                         iosize = PAGE_CACHE_SIZE - page_offset;
2017                         userpage = kmap_atomic(page, KM_USER0);
2018                         memset(userpage + page_offset, 0, iosize);
2019                         flush_dcache_page(page);
2020                         kunmap_atomic(userpage, KM_USER0);
2021                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2022                                             GFP_NOFS);
2023                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2024                         break;
2025                 }
2026                 em = get_extent(inode, page, page_offset, cur,
2027                                 end - cur + 1, 0);
2028                 if (IS_ERR(em) || !em) {
2029                         SetPageError(page);
2030                         unlock_extent(tree, cur, end, GFP_NOFS);
2031                         break;
2032                 }
2033                 extent_offset = cur - em->start;
2034                 if (extent_map_end(em) <= cur) {
2035 printk("bad mapping em [%Lu %Lu] cur %Lu\n", em->start, extent_map_end(em), cur);
2036                 }
2037                 BUG_ON(extent_map_end(em) <= cur);
2038                 if (end < cur) {
2039 printk("2bad mapping end %Lu cur %Lu\n", end, cur);
2040                 }
2041                 BUG_ON(end < cur);
2042
2043                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2044                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2045
2046                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2047                 cur_end = min(extent_map_end(em) - 1, end);
2048                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2049                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2050                         disk_io_size = em->block_len;
2051                         sector = em->block_start >> 9;
2052                 } else {
2053                         sector = (em->block_start + extent_offset) >> 9;
2054                         disk_io_size = iosize;
2055                 }
2056                 bdev = em->bdev;
2057                 block_start = em->block_start;
2058                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2059                         block_start = EXTENT_MAP_HOLE;
2060                 free_extent_map(em);
2061                 em = NULL;
2062
2063                 /* we've found a hole, just zero and go on */
2064                 if (block_start == EXTENT_MAP_HOLE) {
2065                         char *userpage;
2066                         userpage = kmap_atomic(page, KM_USER0);
2067                         memset(userpage + page_offset, 0, iosize);
2068                         flush_dcache_page(page);
2069                         kunmap_atomic(userpage, KM_USER0);
2070
2071                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2072                                             GFP_NOFS);
2073                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2074                         cur = cur + iosize;
2075                         page_offset += iosize;
2076                         continue;
2077                 }
2078                 /* the get_extent function already copied into the page */
2079                 if (test_range_bit(tree, cur, cur_end, EXTENT_UPTODATE, 1)) {
2080                         check_page_uptodate(tree, page);
2081                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2082                         cur = cur + iosize;
2083                         page_offset += iosize;
2084                         continue;
2085                 }
2086                 /* we have an inline extent but it didn't get marked up
2087                  * to date.  Error out
2088                  */
2089                 if (block_start == EXTENT_MAP_INLINE) {
2090                         SetPageError(page);
2091                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2092                         cur = cur + iosize;
2093                         page_offset += iosize;
2094                         continue;
2095                 }
2096
2097                 ret = 0;
2098                 if (tree->ops && tree->ops->readpage_io_hook) {
2099                         ret = tree->ops->readpage_io_hook(page, cur,
2100                                                           cur + iosize - 1);
2101                 }
2102                 if (!ret) {
2103                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2104                         pnr -= page->index;
2105                         ret = submit_extent_page(READ, tree, page,
2106                                          sector, disk_io_size, page_offset,
2107                                          bdev, bio, pnr,
2108                                          end_bio_extent_readpage, mirror_num,
2109                                          *bio_flags,
2110                                          this_bio_flag);
2111                         nr++;
2112                         *bio_flags = this_bio_flag;
2113                 }
2114                 if (ret)
2115                         SetPageError(page);
2116                 cur = cur + iosize;
2117                 page_offset += iosize;
2118         }
2119         if (!nr) {
2120                 if (!PageError(page))
2121                         SetPageUptodate(page);
2122                 unlock_page(page);
2123         }
2124         return 0;
2125 }
2126
2127 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2128                             get_extent_t *get_extent)
2129 {
2130         struct bio *bio = NULL;
2131         unsigned long bio_flags = 0;
2132         int ret;
2133
2134         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2135                                       &bio_flags);
2136         if (bio)
2137                 submit_one_bio(READ, bio, 0, bio_flags);
2138         return ret;
2139 }
2140 EXPORT_SYMBOL(extent_read_full_page);
2141
2142 /*
2143  * the writepage semantics are similar to regular writepage.  extent
2144  * records are inserted to lock ranges in the tree, and as dirty areas
2145  * are found, they are marked writeback.  Then the lock bits are removed
2146  * and the end_io handler clears the writeback ranges
2147  */
2148 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2149                               void *data)
2150 {
2151         struct inode *inode = page->mapping->host;
2152         struct extent_page_data *epd = data;
2153         struct extent_io_tree *tree = epd->tree;
2154         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2155         u64 delalloc_start;
2156         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2157         u64 end;
2158         u64 cur = start;
2159         u64 extent_offset;
2160         u64 last_byte = i_size_read(inode);
2161         u64 block_start;
2162         u64 iosize;
2163         u64 unlock_start;
2164         sector_t sector;
2165         struct extent_map *em;
2166         struct block_device *bdev;
2167         int ret;
2168         int nr = 0;
2169         size_t pg_offset = 0;
2170         size_t blocksize;
2171         loff_t i_size = i_size_read(inode);
2172         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2173         u64 nr_delalloc;
2174         u64 delalloc_end;
2175         int page_started;
2176         int compressed;
2177         unsigned long nr_written = 0;
2178
2179         WARN_ON(!PageLocked(page));
2180         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2181         if (page->index > end_index ||
2182            (page->index == end_index && !pg_offset)) {
2183                 page->mapping->a_ops->invalidatepage(page, 0);
2184                 unlock_page(page);
2185                 return 0;
2186         }
2187
2188         if (page->index == end_index) {
2189                 char *userpage;
2190
2191                 userpage = kmap_atomic(page, KM_USER0);
2192                 memset(userpage + pg_offset, 0,
2193                        PAGE_CACHE_SIZE - pg_offset);
2194                 kunmap_atomic(userpage, KM_USER0);
2195                 flush_dcache_page(page);
2196         }
2197         pg_offset = 0;
2198
2199         set_page_extent_mapped(page);
2200
2201         delalloc_start = start;
2202         delalloc_end = 0;
2203         page_started = 0;
2204         if (!epd->extent_locked) {
2205                 while(delalloc_end < page_end) {
2206                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2207                                                        page,
2208                                                        &delalloc_start,
2209                                                        &delalloc_end,
2210                                                        128 * 1024 * 1024);
2211                         if (nr_delalloc == 0) {
2212                                 delalloc_start = delalloc_end + 1;
2213                                 continue;
2214                         }
2215                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2216                                                  delalloc_end, &page_started,
2217                                                  &nr_written);
2218                         delalloc_start = delalloc_end + 1;
2219                 }
2220
2221                 /* did the fill delalloc function already unlock and start
2222                  * the IO?
2223                  */
2224                 if (page_started) {
2225                         ret = 0;
2226                         goto update_nr_written;
2227                 }
2228         }
2229         lock_extent(tree, start, page_end, GFP_NOFS);
2230
2231         unlock_start = start;
2232
2233         if (tree->ops && tree->ops->writepage_start_hook) {
2234                 ret = tree->ops->writepage_start_hook(page, start,
2235                                                       page_end);
2236                 if (ret == -EAGAIN) {
2237                         unlock_extent(tree, start, page_end, GFP_NOFS);
2238                         redirty_page_for_writepage(wbc, page);
2239                         unlock_page(page);
2240                         ret = 0;
2241                         goto update_nr_written;
2242                 }
2243         }
2244
2245         nr_written++;
2246
2247         end = page_end;
2248         if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0)) {
2249                 printk("found delalloc bits after lock_extent\n");
2250         }
2251
2252         if (last_byte <= start) {
2253                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
2254                 unlock_extent(tree, start, page_end, GFP_NOFS);
2255                 if (tree->ops && tree->ops->writepage_end_io_hook)
2256                         tree->ops->writepage_end_io_hook(page, start,
2257                                                          page_end, NULL, 1);
2258                 unlock_start = page_end + 1;
2259                 goto done;
2260         }
2261
2262         set_extent_uptodate(tree, start, page_end, GFP_NOFS);
2263         blocksize = inode->i_sb->s_blocksize;
2264
2265         while (cur <= end) {
2266                 if (cur >= last_byte) {
2267                         clear_extent_dirty(tree, cur, page_end, GFP_NOFS);
2268                         unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
2269                         if (tree->ops && tree->ops->writepage_end_io_hook)
2270                                 tree->ops->writepage_end_io_hook(page, cur,
2271                                                          page_end, NULL, 1);
2272                         unlock_start = page_end + 1;
2273                         break;
2274                 }
2275                 em = epd->get_extent(inode, page, pg_offset, cur,
2276                                      end - cur + 1, 1);
2277                 if (IS_ERR(em) || !em) {
2278                         SetPageError(page);
2279                         break;
2280                 }
2281
2282                 extent_offset = cur - em->start;
2283                 BUG_ON(extent_map_end(em) <= cur);
2284                 BUG_ON(end < cur);
2285                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2286                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2287                 sector = (em->block_start + extent_offset) >> 9;
2288                 bdev = em->bdev;
2289                 block_start = em->block_start;
2290                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2291                 free_extent_map(em);
2292                 em = NULL;
2293
2294                 /*
2295                  * compressed and inline extents are written through other
2296                  * paths in the FS
2297                  */
2298                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2299                     block_start == EXTENT_MAP_INLINE) {
2300                         clear_extent_dirty(tree, cur,
2301                                            cur + iosize - 1, GFP_NOFS);
2302
2303                         unlock_extent(tree, unlock_start, cur + iosize -1,
2304                                       GFP_NOFS);
2305
2306                         /*
2307                          * end_io notification does not happen here for
2308                          * compressed extents
2309                          */
2310                         if (!compressed && tree->ops &&
2311                             tree->ops->writepage_end_io_hook)
2312                                 tree->ops->writepage_end_io_hook(page, cur,
2313                                                          cur + iosize - 1,
2314                                                          NULL, 1);
2315                         else if (compressed) {
2316                                 /* we don't want to end_page_writeback on
2317                                  * a compressed extent.  this happens
2318                                  * elsewhere
2319                                  */
2320                                 nr++;
2321                         }
2322
2323                         cur += iosize;
2324                         pg_offset += iosize;
2325                         unlock_start = cur;
2326                         continue;
2327                 }
2328                 /* leave this out until we have a page_mkwrite call */
2329                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2330                                    EXTENT_DIRTY, 0)) {
2331                         cur = cur + iosize;
2332                         pg_offset += iosize;
2333                         continue;
2334                 }
2335
2336                 clear_extent_dirty(tree, cur, cur + iosize - 1, GFP_NOFS);
2337                 if (tree->ops && tree->ops->writepage_io_hook) {
2338                         ret = tree->ops->writepage_io_hook(page, cur,
2339                                                 cur + iosize - 1);
2340                 } else {
2341                         ret = 0;
2342                 }
2343                 if (ret) {
2344                         SetPageError(page);
2345                 } else {
2346                         unsigned long max_nr = end_index + 1;
2347
2348                         set_range_writeback(tree, cur, cur + iosize - 1);
2349                         if (!PageWriteback(page)) {
2350                                 printk("warning page %lu not writeback, "
2351                                        "cur %llu end %llu\n", page->index,
2352                                        (unsigned long long)cur,
2353                                        (unsigned long long)end);
2354                         }
2355
2356                         ret = submit_extent_page(WRITE, tree, page, sector,
2357                                                  iosize, pg_offset, bdev,
2358                                                  &epd->bio, max_nr,
2359                                                  end_bio_extent_writepage,
2360                                                  0, 0, 0);
2361                         if (ret)
2362                                 SetPageError(page);
2363                 }
2364                 cur = cur + iosize;
2365                 pg_offset += iosize;
2366                 nr++;
2367         }
2368 done:
2369         if (nr == 0) {
2370                 /* make sure the mapping tag for page dirty gets cleared */
2371                 set_page_writeback(page);
2372                 end_page_writeback(page);
2373         }
2374         if (unlock_start <= page_end)
2375                 unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
2376         unlock_page(page);
2377
2378 update_nr_written:
2379         wbc->nr_to_write -= nr_written;
2380         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2381             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2382                 page->mapping->writeback_index = page->index + nr_written;
2383         return 0;
2384 }
2385
2386 /**
2387  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2388  * @mapping: address space structure to write
2389  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2390  * @writepage: function called for each page
2391  * @data: data passed to writepage function
2392  *
2393  * If a page is already under I/O, write_cache_pages() skips it, even
2394  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2395  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2396  * and msync() need to guarantee that all the data which was dirty at the time
2397  * the call was made get new I/O started against them.  If wbc->sync_mode is
2398  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2399  * existing IO to complete.
2400  */
2401 static int extent_write_cache_pages(struct extent_io_tree *tree,
2402                              struct address_space *mapping,
2403                              struct writeback_control *wbc,
2404                              writepage_t writepage, void *data,
2405                              void (*flush_fn)(void *))
2406 {
2407         struct backing_dev_info *bdi = mapping->backing_dev_info;
2408         int ret = 0;
2409         int done = 0;
2410         struct pagevec pvec;
2411         int nr_pages;
2412         pgoff_t index;
2413         pgoff_t end;            /* Inclusive */
2414         int scanned = 0;
2415         int range_whole = 0;
2416
2417         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2418                 wbc->encountered_congestion = 1;
2419                 return 0;
2420         }
2421
2422         pagevec_init(&pvec, 0);
2423         if (wbc->range_cyclic) {
2424                 index = mapping->writeback_index; /* Start from prev offset */
2425                 end = -1;
2426         } else {
2427                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2428                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2429                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2430                         range_whole = 1;
2431                 scanned = 1;
2432         }
2433 retry:
2434         while (!done && (index <= end) &&
2435                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2436                                               PAGECACHE_TAG_DIRTY,
2437                                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2438                 unsigned i;
2439
2440                 scanned = 1;
2441                 for (i = 0; i < nr_pages; i++) {
2442                         struct page *page = pvec.pages[i];
2443
2444                         /*
2445                          * At this point we hold neither mapping->tree_lock nor
2446                          * lock on the page itself: the page may be truncated or
2447                          * invalidated (changing page->mapping to NULL), or even
2448                          * swizzled back from swapper_space to tmpfs file
2449                          * mapping
2450                          */
2451                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2452                                 tree->ops->write_cache_pages_lock_hook(page);
2453                         else
2454                                 lock_page(page);
2455
2456                         if (unlikely(page->mapping != mapping)) {
2457                                 unlock_page(page);
2458                                 continue;
2459                         }
2460
2461                         if (!wbc->range_cyclic && page->index > end) {
2462                                 done = 1;
2463                                 unlock_page(page);
2464                                 continue;
2465                         }
2466
2467                         if (wbc->sync_mode != WB_SYNC_NONE) {
2468                                 if (PageWriteback(page))
2469                                         flush_fn(data);
2470                                 wait_on_page_writeback(page);
2471                         }
2472
2473                         if (PageWriteback(page) ||
2474                             !clear_page_dirty_for_io(page)) {
2475                                 unlock_page(page);
2476                                 continue;
2477                         }
2478
2479                         ret = (*writepage)(page, wbc, data);
2480
2481                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2482                                 unlock_page(page);
2483                                 ret = 0;
2484                         }
2485                         if (ret || wbc->nr_to_write <= 0)
2486                                 done = 1;
2487                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2488                                 wbc->encountered_congestion = 1;
2489                                 done = 1;
2490                         }
2491                 }
2492                 pagevec_release(&pvec);
2493                 cond_resched();
2494         }
2495         if (!scanned && !done) {
2496                 /*
2497                  * We hit the last page and there is more work to be done: wrap
2498                  * back to the start of the file
2499                  */
2500                 scanned = 1;
2501                 index = 0;
2502                 goto retry;
2503         }
2504         return ret;
2505 }
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);