SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/gfp.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/page-flags.h>
   8 #include <linux/module.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/swap.h>
  12 #include <linux/version.h>
  13 #include <linux/writeback.h>
  14 #include <linux/pagevec.h>
  15 #include "extent_io.h"
  16 #include "extent_map.h"
  17 #include "compat.h"
  18 #include "ctree.h"
  19 #include "btrfs_inode.h"
  20
  21 /* temporary define until extent_map moves out of btrfs */
  22 struct kmem_cache *btrfs_cache_create(const char *name, size_t size,
  23                                        unsigned long extra_flags,
  24                                        void (*ctor)(void *, struct kmem_cache *,
  25                                                     unsigned long));
  26
  27 static struct kmem_cache *extent_state_cache;
  28 static struct kmem_cache *extent_buffer_cache;
  29
  30 static LIST_HEAD(buffers);
  31 static LIST_HEAD(states);
  32
  33 #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         int err;
 481         int set = 0;
 482
 483 again:
 484         if (!prealloc && (mask & __GFP_WAIT)) {
 485                 prealloc = alloc_extent_state(mask);
 486                 if (!prealloc)
 487                         return -ENOMEM;
 488         }
 489
 490         spin_lock(&tree->lock);
 491         /*
 492          * this search will find the extents that end after
 493          * our range starts
 494          */
 495         node = tree_search(tree, start);
 496         if (!node)
 497                 goto out;
 498         state = rb_entry(node, struct extent_state, rb_node);
 499         if (state->start > end)
 500                 goto out;
 501         WARN_ON(state->end < start);
 502
 503         /*
 504          *     | ---- desired range ---- |
 505          *  | state | or
 506          *  | ------------- state -------------- |
 507          *
 508          * We need to split the extent we found, and may flip
 509          * bits on second half.
 510          *
 511          * If the extent we found extends past our range, we
 512          * just split and search again.  It'll get split again
 513          * the next time though.
 514          *
 515          * If the extent we found is inside our range, we clear
 516          * the desired bit on it.
 517          */
 518
 519         if (state->start < start) {
 520                 if (!prealloc)
 521                         prealloc = alloc_extent_state(GFP_ATOMIC);
 522                 err = split_state(tree, state, prealloc, start);
 523                 BUG_ON(err == -EEXIST);
 524                 prealloc = NULL;
 525                 if (err)
 526                         goto out;
 527                 if (state->end <= end) {
 528                         start = state->end + 1;
 529                         set |= clear_state_bit(tree, state, bits,
 530                                         wake, delete);
 531                 } else {
 532                         start = state->start;
 533                 }
 534                 goto search_again;
 535         }
 536         /*
 537          * | ---- desired range ---- |
 538          *                        | state |
 539          * We need to split the extent, and clear the bit
 540          * on the first half
 541          */
 542         if (state->start <= end && state->end > end) {
 543                 if (!prealloc)
 544                         prealloc = alloc_extent_state(GFP_ATOMIC);
 545                 err = split_state(tree, state, prealloc, end + 1);
 546                 BUG_ON(err == -EEXIST);
 547
 548                 if (wake)
 549                         wake_up(&state->wq);
 550                 set |= clear_state_bit(tree, prealloc, bits,
 551                                        wake, delete);
 552                 prealloc = NULL;
 553                 goto out;
 554         }
 555
 556         start = state->end + 1;
 557         set |= clear_state_bit(tree, state, bits, wake, delete);
 558         goto search_again;
 559
 560 out:
 561         spin_unlock(&tree->lock);
 562         if (prealloc)
 563                 free_extent_state(prealloc);
 564
 565         return set;
 566
 567 search_again:
 568         if (start > end)
 569                 goto out;
 570         spin_unlock(&tree->lock);
 571         if (mask & __GFP_WAIT)
 572                 cond_resched();
 573         goto again;
 574 }
 575 EXPORT_SYMBOL(clear_extent_bit);
 576
 577 static int wait_on_state(struct extent_io_tree *tree,
 578                          struct extent_state *state)
 579                 __releases(tree->lock)
 580                 __acquires(tree->lock)
 581 {
 582         DEFINE_WAIT(wait);
 583         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 584         spin_unlock(&tree->lock);
 585         schedule();
 586         spin_lock(&tree->lock);
 587         finish_wait(&state->wq, &wait);
 588         return 0;
 589 }
 590
 591 /*
 592  * waits for one or more bits to clear on a range in the state tree.
 593  * The range [start, end] is inclusive.
 594  * The tree lock is taken by this function
 595  */
 596 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 597 {
 598         struct extent_state *state;
 599         struct rb_node *node;
 600
 601         spin_lock(&tree->lock);
 602 again:
 603         while (1) {
 604                 /*
 605                  * this search will find all the extents that end after
 606                  * our range starts
 607                  */
 608                 node = tree_search(tree, start);
 609                 if (!node)
 610                         break;
 611
 612                 state = rb_entry(node, struct extent_state, rb_node);
 613
 614                 if (state->start > end)
 615                         goto out;
 616
 617                 if (state->state & bits) {
 618                         start = state->start;
 619                         atomic_inc(&state->refs);
 620                         wait_on_state(tree, state);
 621                         free_extent_state(state);
 622                         goto again;
 623                 }
 624                 start = state->end + 1;
 625
 626                 if (start > end)
 627                         break;
 628
 629                 if (need_resched()) {
 630                         spin_unlock(&tree->lock);
 631                         cond_resched();
 632                         spin_lock(&tree->lock);
 633                 }
 634         }
 635 out:
 636         spin_unlock(&tree->lock);
 637         return 0;
 638 }
 639 EXPORT_SYMBOL(wait_extent_bit);
 640
 641 static void set_state_bits(struct extent_io_tree *tree,
 642                            struct extent_state *state,
 643                            int bits)
 644 {
 645         if ((bits & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 646                 u64 range = state->end - state->start + 1;
 647                 tree->dirty_bytes += range;
 648         }
 649         set_state_cb(tree, state, bits);
 650         state->state |= bits;
 651 }
 652
 653 /*
 654  * set some bits on a range in the tree.  This may require allocations
 655  * or sleeping, so the gfp mask is used to indicate what is allowed.
 656  *
 657  * If 'exclusive' == 1, this will fail with -EEXIST if some part of the
 658  * range already has the desired bits set.  The start of the existing
 659  * range is returned in failed_start in this case.
 660  *
 661  * [start, end] is inclusive
 662  * This takes the tree lock.
 663  */
 664 static int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits,
 665                    int exclusive, u64 *failed_start, gfp_t mask)
 666 {
 667         struct extent_state *state;
 668         struct extent_state *prealloc = NULL;
 669         struct rb_node *node;
 670         int err = 0;
 671         int set;
 672         u64 last_start;
 673         u64 last_end;
 674 again:
 675         if (!prealloc && (mask & __GFP_WAIT)) {
 676                 prealloc = alloc_extent_state(mask);
 677                 if (!prealloc)
 678                         return -ENOMEM;
 679         }
 680
 681         spin_lock(&tree->lock);
 682         /*
 683          * this search will find all the extents that end after
 684          * our range starts.
 685          */
 686         node = tree_search(tree, start);
 687         if (!node) {
 688                 err = insert_state(tree, prealloc, start, end, bits);
 689                 prealloc = NULL;
 690                 BUG_ON(err == -EEXIST);
 691                 goto out;
 692         }
 693
 694         state = rb_entry(node, struct extent_state, rb_node);
 695         last_start = state->start;
 696         last_end = state->end;
 697
 698         /*
 699          * | ---- desired range ---- |
 700          * | state |
 701          *
 702          * Just lock what we found and keep going
 703          */
 704         if (state->start == start && state->end <= end) {
 705                 set = state->state & bits;
 706                 if (set && exclusive) {
 707                         *failed_start = state->start;
 708                         err = -EEXIST;
 709                         goto out;
 710                 }
 711                 set_state_bits(tree, state, bits);
 712                 start = state->end + 1;
 713                 merge_state(tree, state);
 714                 goto search_again;
 715         }
 716
 717         /*
 718          *     | ---- desired range ---- |
 719          * | state |
 720          *   or
 721          * | ------------- state -------------- |
 722          *
 723          * We need to split the extent we found, and may flip bits on
 724          * second half.
 725          *
 726          * If the extent we found extends past our
 727          * range, we just split and search again.  It'll get split
 728          * again the next time though.
 729          *
 730          * If the extent we found is inside our range, we set the
 731          * desired bit on it.
 732          */
 733         if (state->start < start) {
 734                 set = state->state & bits;
 735                 if (exclusive && set) {
 736                         *failed_start = start;
 737                         err = -EEXIST;
 738                         goto out;
 739                 }
 740                 err = split_state(tree, state, prealloc, start);
 741                 BUG_ON(err == -EEXIST);
 742                 prealloc = NULL;
 743                 if (err)
 744                         goto out;
 745                 if (state->end <= end) {
 746                         set_state_bits(tree, state, bits);
 747                         start = state->end + 1;
 748                         merge_state(tree, state);
 749                 } else {
 750                         start = state->start;
 751                 }
 752                 goto search_again;
 753         }
 754         /*
 755          * | ---- desired range ---- |
 756          *     | state | or               | state |
 757          *
 758          * There's a hole, we need to insert something in it and
 759          * ignore the extent we found.
 760          */
 761         if (state->start > start) {
 762                 u64 this_end;
 763                 if (end < last_start)
 764                         this_end = end;
 765                 else
 766                         this_end = last_start -1;
 767                 err = insert_state(tree, prealloc, start, this_end,
 768                                    bits);
 769                 prealloc = NULL;
 770                 BUG_ON(err == -EEXIST);
 771                 if (err)
 772                         goto out;
 773                 start = this_end + 1;
 774                 goto search_again;
 775         }
 776         /*
 777          * | ---- desired range ---- |
 778          *                        | state |
 779          * We need to split the extent, and set the bit
 780          * on the first half
 781          */
 782         if (state->start <= end && state->end > end) {
 783                 set = state->state & bits;
 784                 if (exclusive && set) {
 785                         *failed_start = start;
 786                         err = -EEXIST;
 787                         goto out;
 788                 }
 789                 err = split_state(tree, state, prealloc, end + 1);
 790                 BUG_ON(err == -EEXIST);
 791
 792                 set_state_bits(tree, prealloc, bits);
 793                 merge_state(tree, prealloc);
 794                 prealloc = NULL;
 795                 goto out;
 796         }
 797
 798         goto search_again;
 799
 800 out:
 801         spin_unlock(&tree->lock);
 802         if (prealloc)
 803                 free_extent_state(prealloc);
 804
 805         return err;
 806
 807 search_again:
 808         if (start > end)
 809                 goto out;
 810         spin_unlock(&tree->lock);
 811         if (mask & __GFP_WAIT)
 812                 cond_resched();
 813         goto again;
 814 }
 815 EXPORT_SYMBOL(set_extent_bit);
 816
 817 /* wrappers around set/clear extent bit */
 818 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 819                      gfp_t mask)
 820 {
 821         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
 822                               mask);
 823 }
 824 EXPORT_SYMBOL(set_extent_dirty);
 825
 826 int set_extent_ordered(struct extent_io_tree *tree, u64 start, u64 end,
 827                        gfp_t mask)
 828 {
 829         return set_extent_bit(tree, start, end, EXTENT_ORDERED, 0, NULL, mask);
 830 }
 831 EXPORT_SYMBOL(set_extent_ordered);
 832
 833 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 834                     int bits, gfp_t mask)
 835 {
 836         return set_extent_bit(tree, start, end, bits, 0, NULL,
 837                               mask);
 838 }
 839 EXPORT_SYMBOL(set_extent_bits);
 840
 841 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
 842                       int bits, gfp_t mask)
 843 {
 844         return clear_extent_bit(tree, start, end, bits, 0, 0, mask);
 845 }
 846 EXPORT_SYMBOL(clear_extent_bits);
 847
 848 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
 849                      gfp_t mask)
 850 {
 851         return set_extent_bit(tree, start, end,
 852                               EXTENT_DELALLOC | EXTENT_DIRTY,
 853                               0, NULL, mask);
 854 }
 855 EXPORT_SYMBOL(set_extent_delalloc);
 856
 857 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
 858                        gfp_t mask)
 859 {
 860         return clear_extent_bit(tree, start, end,
 861                                 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, mask);
 862 }
 863 EXPORT_SYMBOL(clear_extent_dirty);
 864
 865 int clear_extent_ordered(struct extent_io_tree *tree, u64 start, u64 end,
 866                          gfp_t mask)
 867 {
 868         return clear_extent_bit(tree, start, end, EXTENT_ORDERED, 1, 0, mask);
 869 }
 870 EXPORT_SYMBOL(clear_extent_ordered);
 871
 872 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 873                      gfp_t mask)
 874 {
 875         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
 876                               mask);
 877 }
 878 EXPORT_SYMBOL(set_extent_new);
 879
 880 static int clear_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
 881                        gfp_t mask)
 882 {
 883         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0, mask);
 884 }
 885
 886 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 887                         gfp_t mask)
 888 {
 889         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, NULL,
 890                               mask);
 891 }
 892 EXPORT_SYMBOL(set_extent_uptodate);
 893
 894 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
 895                           gfp_t mask)
 896 {
 897         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0, mask);
 898 }
 899
 900 static int set_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end,
 901                          gfp_t mask)
 902 {
 903         return set_extent_bit(tree, start, end, EXTENT_WRITEBACK,
 904                               0, NULL, mask);
 905 }
 906
 907 static int clear_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end,
 908                            gfp_t mask)
 909 {
 910         return clear_extent_bit(tree, start, end, EXTENT_WRITEBACK, 1, 0, mask);
 911 }
 912
 913 int wait_on_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 914 {
 915         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
 916 }
 917 EXPORT_SYMBOL(wait_on_extent_writeback);
 918
 919 /*
 920  * either insert or lock state struct between start and end use mask to tell
 921  * us if waiting is desired.
 922  */
 923 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
 924 {
 925         int err;
 926         u64 failed_start;
 927         while (1) {
 928                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED, 1,
 929                                      &failed_start, mask);
 930                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
 931                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
 932                         start = failed_start;
 933                 } else {
 934                         break;
 935                 }
 936                 WARN_ON(start > end);
 937         }
 938         return err;
 939 }
 940 EXPORT_SYMBOL(lock_extent);
 941
 942 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
 943                     gfp_t mask)
 944 {
 945         int err;
 946         u64 failed_start;
 947
 948         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, 1,
 949                              &failed_start, mask);
 950         if (err == -EEXIST) {
 951                 if (failed_start > start)
 952                         clear_extent_bit(tree, start, failed_start - 1,
 953                                          EXTENT_LOCKED, 1, 0, mask);
 954                 return 0;
 955         }
 956         return 1;
 957 }
 958 EXPORT_SYMBOL(try_lock_extent);
 959
 960 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end,
 961                   gfp_t mask)
 962 {
 963         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, mask);
 964 }
 965 EXPORT_SYMBOL(unlock_extent);
 966
 967 /*
 968  * helper function to set pages and extents in the tree dirty
 969  */
 970 int set_range_dirty(struct extent_io_tree *tree, u64 start, u64 end)
 971 {
 972         unsigned long index = start >> PAGE_CACHE_SHIFT;
 973         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 974         struct page *page;
 975
 976         while (index <= end_index) {
 977                 page = find_get_page(tree->mapping, index);
 978                 BUG_ON(!page);
 979                 __set_page_dirty_nobuffers(page);
 980                 page_cache_release(page);
 981                 index++;
 982         }
 983         set_extent_dirty(tree, start, end, GFP_NOFS);
 984         return 0;
 985 }
 986 EXPORT_SYMBOL(set_range_dirty);
 987
 988 /*
 989  * helper function to set both pages and extents in the tree writeback
 990  */
 991 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
 992 {
 993         unsigned long index = start >> PAGE_CACHE_SHIFT;
 994         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
 995         struct page *page;
 996
 997         while (index <= end_index) {
 998                 page = find_get_page(tree->mapping, index);
 999                 BUG_ON(!page);
1000                 set_page_writeback(page);
1001                 page_cache_release(page);
1002                 index++;
1003         }
1004         set_extent_writeback(tree, start, end, GFP_NOFS);
1005         return 0;
1006 }
1007
1008 /*
1009  * find the first offset in the io tree with 'bits' set. zero is
1010  * returned if we find something, and *start_ret and *end_ret are
1011  * set to reflect the state struct that was found.
1012  *
1013  * If nothing was found, 1 is returned, < 0 on error
1014  */
1015 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1016                           u64 *start_ret, u64 *end_ret, int bits)
1017 {
1018         struct rb_node *node;
1019         struct extent_state *state;
1020         int ret = 1;
1021
1022         spin_lock(&tree->lock);
1023         /*
1024          * this search will find all the extents that end after
1025          * our range starts.
1026          */
1027         node = tree_search(tree, start);
1028         if (!node) {
1029                 goto out;
1030         }
1031
1032         while(1) {
1033                 state = rb_entry(node, struct extent_state, rb_node);
1034                 if (state->end >= start && (state->state & bits)) {
1035                         *start_ret = state->start;
1036                         *end_ret = state->end;
1037                         ret = 0;
1038                         break;
1039                 }
1040                 node = rb_next(node);
1041                 if (!node)
1042                         break;
1043         }
1044 out:
1045         spin_unlock(&tree->lock);
1046         return ret;
1047 }
1048 EXPORT_SYMBOL(find_first_extent_bit);
1049
1050 /* find the first state struct with 'bits' set after 'start', and
1051  * return it.  tree->lock must be held.  NULL will returned if
1052  * nothing was found after 'start'
1053  */
1054 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1055                                                  u64 start, int bits)
1056 {
1057         struct rb_node *node;
1058         struct extent_state *state;
1059
1060         /*
1061          * this search will find all the extents that end after
1062          * our range starts.
1063          */
1064         node = tree_search(tree, start);
1065         if (!node) {
1066                 goto out;
1067         }
1068
1069         while(1) {
1070                 state = rb_entry(node, struct extent_state, rb_node);
1071                 if (state->end >= start && (state->state & bits)) {
1072                         return state;
1073                 }
1074                 node = rb_next(node);
1075                 if (!node)
1076                         break;
1077         }
1078 out:
1079         return NULL;
1080 }
1081 EXPORT_SYMBOL(find_first_extent_bit_state);
1082
1083 /*
1084  * find a contiguous range of bytes in the file marked as delalloc, not
1085  * more than 'max_bytes'.  start and end are used to return the range,
1086  *
1087  * 1 is returned if we find something, 0 if nothing was in the tree
1088  */
1089 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1090                                         u64 *start, u64 *end, u64 max_bytes)
1091 {
1092         struct rb_node *node;
1093         struct extent_state *state;
1094         u64 cur_start = *start;
1095         u64 found = 0;
1096         u64 total_bytes = 0;
1097
1098         spin_lock(&tree->lock);
1099
1100         /*
1101          * this search will find all the extents that end after
1102          * our range starts.
1103          */
1104         node = tree_search(tree, cur_start);
1105         if (!node) {
1106                 if (!found)
1107                         *end = (u64)-1;
1108                 goto out;
1109         }
1110
1111         while(1) {
1112                 state = rb_entry(node, struct extent_state, rb_node);
1113                 if (found && (state->start != cur_start ||
1114                               (state->state & EXTENT_BOUNDARY))) {
1115                         goto out;
1116                 }
1117                 if (!(state->state & EXTENT_DELALLOC)) {
1118                         if (!found)
1119                                 *end = state->end;
1120                         goto out;
1121                 }
1122                 if (!found)
1123                         *start = state->start;
1124                 found++;
1125                 *end = state->end;
1126                 cur_start = state->end + 1;
1127                 node = rb_next(node);
1128                 if (!node)
1129                         break;
1130                 total_bytes += state->end - state->start + 1;
1131                 if (total_bytes >= max_bytes)
1132                         break;
1133         }
1134 out:
1135         spin_unlock(&tree->lock);
1136         return found;
1137 }
1138
1139 static noinline int __unlock_for_delalloc(struct inode *inode,
1140                                           struct page *locked_page,
1141                                           u64 start, u64 end)
1142 {
1143         int ret;
1144         struct page *pages[16];
1145         unsigned long index = start >> PAGE_CACHE_SHIFT;
1146         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1147         unsigned long nr_pages = end_index - index + 1;
1148         int i;
1149
1150         if (index == locked_page->index && end_index == index)
1151                 return 0;
1152
1153         while(nr_pages > 0) {
1154                 ret = find_get_pages_contig(inode->i_mapping, index,
1155                                      min_t(unsigned long, nr_pages,
1156                                      ARRAY_SIZE(pages)), pages);
1157                 for (i = 0; i < ret; i++) {
1158                         if (pages[i] != locked_page)
1159                                 unlock_page(pages[i]);
1160                         page_cache_release(pages[i]);
1161                 }
1162                 nr_pages -= ret;
1163                 index += ret;
1164                 cond_resched();
1165         }
1166         return 0;
1167 }
1168
1169 static noinline int lock_delalloc_pages(struct inode *inode,
1170                                         struct page *locked_page,
1171                                         u64 delalloc_start,
1172                                         u64 delalloc_end)
1173 {
1174         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1175         unsigned long start_index = index;
1176         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1177         unsigned long pages_locked = 0;
1178         struct page *pages[16];
1179         unsigned long nrpages;
1180         int ret;
1181         int i;
1182
1183         /* the caller is responsible for locking the start index */
1184         if (index == locked_page->index && index == end_index)
1185                 return 0;
1186
1187         /* skip the page at the start index */
1188         nrpages = end_index - index + 1;
1189         while(nrpages > 0) {
1190                 ret = find_get_pages_contig(inode->i_mapping, index,
1191                                      min_t(unsigned long,
1192                                      nrpages, ARRAY_SIZE(pages)), pages);
1193                 if (ret == 0) {
1194                         ret = -EAGAIN;
1195                         goto done;
1196                 }
1197                 /* now we have an array of pages, lock them all */
1198                 for (i = 0; i < ret; i++) {
1199                         /*
1200                          * the caller is taking responsibility for
1201                          * locked_page
1202                          */
1203                         if (pages[i] != locked_page) {
1204                                 lock_page(pages[i]);
1205                                 if (!PageDirty(pages[i]) ||
1206                                     pages[i]->mapping != inode->i_mapping) {
1207                                         ret = -EAGAIN;
1208                                         unlock_page(pages[i]);
1209                                         page_cache_release(pages[i]);
1210                                         goto done;
1211                                 }
1212                         }
1213                         page_cache_release(pages[i]);
1214                         pages_locked++;
1215                 }
1216                 nrpages -= ret;
1217                 index += ret;
1218                 cond_resched();
1219         }
1220         ret = 0;
1221 done:
1222         if (ret && pages_locked) {
1223                 __unlock_for_delalloc(inode, locked_page,
1224                               delalloc_start,
1225                               ((u64)(start_index + pages_locked - 1)) <<
1226                               PAGE_CACHE_SHIFT);
1227         }
1228         return ret;
1229 }
1230
1231 /*
1232  * find a contiguous range of bytes in the file marked as delalloc, not
1233  * more than 'max_bytes'.  start and end are used to return the range,
1234  *
1235  * 1 is returned if we find something, 0 if nothing was in the tree
1236  */
1237 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1238                                              struct extent_io_tree *tree,
1239                                              struct page *locked_page,
1240                                              u64 *start, u64 *end,
1241                                              u64 max_bytes)
1242 {
1243         u64 delalloc_start;
1244         u64 delalloc_end;
1245         u64 found;
1246         int ret;
1247         int loops = 0;
1248
1249 again:
1250         /* step one, find a bunch of delalloc bytes starting at start */
1251         delalloc_start = *start;
1252         delalloc_end = 0;
1253         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1254                                     max_bytes);
1255         if (!found || delalloc_end <= *start) {
1256                 *start = delalloc_start;
1257                 *end = delalloc_end;
1258                 return found;
1259         }
1260
1261         /*
1262          * start comes from the offset of locked_page.  We have to lock
1263          * pages in order, so we can't process delalloc bytes before
1264          * locked_page
1265          */
1266         if (delalloc_start < *start) {
1267                 delalloc_start = *start;
1268         }
1269
1270         /*
1271          * make sure to limit the number of pages we try to lock down
1272          * if we're looping.
1273          */
1274         if (delalloc_end + 1 - delalloc_start > max_bytes && loops) {
1275                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1276         }
1277         /* step two, lock all the pages after the page that has start */
1278         ret = lock_delalloc_pages(inode, locked_page,
1279                                   delalloc_start, delalloc_end);
1280         if (ret == -EAGAIN) {
1281                 /* some of the pages are gone, lets avoid looping by
1282                  * shortening the size of the delalloc range we're searching
1283                  */
1284                 if (!loops) {
1285                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1286                         max_bytes = PAGE_CACHE_SIZE - offset;
1287                         loops = 1;
1288                         goto again;
1289                 } else {
1290                         found = 0;
1291                         goto out_failed;
1292                 }
1293         }
1294         BUG_ON(ret);
1295
1296         /* step three, lock the state bits for the whole range */
1297         lock_extent(tree, delalloc_start, delalloc_end, GFP_NOFS);
1298
1299         /* then test to make sure it is all still delalloc */
1300         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1301                              EXTENT_DELALLOC, 1);
1302         if (!ret) {
1303                 unlock_extent(tree, delalloc_start, delalloc_end, GFP_NOFS);
1304                 __unlock_for_delalloc(inode, locked_page,
1305                               delalloc_start, delalloc_end);
1306                 cond_resched();
1307                 goto again;
1308         }
1309         *start = delalloc_start;
1310         *end = delalloc_end;
1311 out_failed:
1312         return found;
1313 }
1314
1315 int extent_clear_unlock_delalloc(struct inode *inode,
1316                                 struct extent_io_tree *tree,
1317                                 u64 start, u64 end, struct page *locked_page,
1318                                 int unlock_pages,
1319                                 int clear_unlock,
1320                                 int clear_delalloc, int clear_dirty,
1321                                 int set_writeback,
1322                                 int end_writeback)
1323 {
1324         int ret;
1325         struct page *pages[16];
1326         unsigned long index = start >> PAGE_CACHE_SHIFT;
1327         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1328         unsigned long nr_pages = end_index - index + 1;
1329         int i;
1330         int clear_bits = 0;
1331
1332         if (clear_unlock)
1333                 clear_bits |= EXTENT_LOCKED;
1334         if (clear_dirty)
1335                 clear_bits |= EXTENT_DIRTY;
1336
1337         if (clear_delalloc)
1338                 clear_bits |= EXTENT_DELALLOC;
1339
1340         clear_extent_bit(tree, start, end, clear_bits, 1, 0, GFP_NOFS);
1341         if (!(unlock_pages || clear_dirty || set_writeback || end_writeback))
1342                 return 0;
1343
1344         while(nr_pages > 0) {
1345                 ret = find_get_pages_contig(inode->i_mapping, index,
1346                                      min_t(unsigned long,
1347                                      nr_pages, ARRAY_SIZE(pages)), pages);
1348                 for (i = 0; i < ret; i++) {
1349                         if (pages[i] == locked_page) {
1350                                 page_cache_release(pages[i]);
1351                                 continue;
1352                         }
1353                         if (clear_dirty)
1354                                 clear_page_dirty_for_io(pages[i]);
1355                         if (set_writeback)
1356                                 set_page_writeback(pages[i]);
1357                         if (end_writeback)
1358                                 end_page_writeback(pages[i]);
1359                         if (unlock_pages)
1360                                 unlock_page(pages[i]);
1361                         page_cache_release(pages[i]);
1362                 }
1363                 nr_pages -= ret;
1364                 index += ret;
1365                 cond_resched();
1366         }
1367         return 0;
1368 }
1369 EXPORT_SYMBOL(extent_clear_unlock_delalloc);
1370
1371 /*
1372  * count the number of bytes in the tree that have a given bit(s)
1373  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1374  * cached.  The total number found is returned.
1375  */
1376 u64 count_range_bits(struct extent_io_tree *tree,
1377                      u64 *start, u64 search_end, u64 max_bytes,
1378                      unsigned long bits)
1379 {
1380         struct rb_node *node;
1381         struct extent_state *state;
1382         u64 cur_start = *start;
1383         u64 total_bytes = 0;
1384         int found = 0;
1385
1386         if (search_end <= cur_start) {
1387                 printk("search_end %Lu start %Lu\n", search_end, cur_start);
1388                 WARN_ON(1);
1389                 return 0;
1390         }
1391
1392         spin_lock(&tree->lock);
1393         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1394                 total_bytes = tree->dirty_bytes;
1395                 goto out;
1396         }
1397         /*
1398          * this search will find all the extents that end after
1399          * our range starts.
1400          */
1401         node = tree_search(tree, cur_start);
1402         if (!node) {
1403                 goto out;
1404         }
1405
1406         while(1) {
1407                 state = rb_entry(node, struct extent_state, rb_node);
1408                 if (state->start > search_end)
1409                         break;
1410                 if (state->end >= cur_start && (state->state & bits)) {
1411                         total_bytes += min(search_end, state->end) + 1 -
1412                                        max(cur_start, state->start);
1413                         if (total_bytes >= max_bytes)
1414                                 break;
1415                         if (!found) {
1416                                 *start = state->start;
1417                                 found = 1;
1418                         }
1419                 }
1420                 node = rb_next(node);
1421                 if (!node)
1422                         break;
1423         }
1424 out:
1425         spin_unlock(&tree->lock);
1426         return total_bytes;
1427 }
1428
1429 #if 0
1430 /*
1431  * helper function to lock both pages and extents in the tree.
1432  * pages must be locked first.
1433  */
1434 static int lock_range(struct extent_io_tree *tree, u64 start, u64 end)
1435 {
1436         unsigned long index = start >> PAGE_CACHE_SHIFT;
1437         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1438         struct page *page;
1439         int err;
1440
1441         while (index <= end_index) {
1442                 page = grab_cache_page(tree->mapping, index);
1443                 if (!page) {
1444                         err = -ENOMEM;
1445                         goto failed;
1446                 }
1447                 if (IS_ERR(page)) {
1448                         err = PTR_ERR(page);
1449                         goto failed;
1450                 }
1451                 index++;
1452         }
1453         lock_extent(tree, start, end, GFP_NOFS);
1454         return 0;
1455
1456 failed:
1457         /*
1458          * we failed above in getting the page at 'index', so we undo here
1459          * up to but not including the page at 'index'
1460          */
1461         end_index = index;
1462         index = start >> PAGE_CACHE_SHIFT;
1463         while (index < end_index) {
1464                 page = find_get_page(tree->mapping, index);
1465                 unlock_page(page);
1466                 page_cache_release(page);
1467                 index++;
1468         }
1469         return err;
1470 }
1471
1472 /*
1473  * helper function to unlock both pages and extents in the tree.
1474  */
1475 static int unlock_range(struct extent_io_tree *tree, u64 start, u64 end)
1476 {
1477         unsigned long index = start >> PAGE_CACHE_SHIFT;
1478         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1479         struct page *page;
1480
1481         while (index <= end_index) {
1482                 page = find_get_page(tree->mapping, index);
1483                 unlock_page(page);
1484                 page_cache_release(page);
1485                 index++;
1486         }
1487         unlock_extent(tree, start, end, GFP_NOFS);
1488         return 0;
1489 }
1490 #endif
1491
1492 /*
1493  * set the private field for a given byte offset in the tree.  If there isn't
1494  * an extent_state there already, this does nothing.
1495  */
1496 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1497 {
1498         struct rb_node *node;
1499         struct extent_state *state;
1500         int ret = 0;
1501
1502         spin_lock(&tree->lock);
1503         /*
1504          * this search will find all the extents that end after
1505          * our range starts.
1506          */
1507         node = tree_search(tree, start);
1508         if (!node) {
1509                 ret = -ENOENT;
1510                 goto out;
1511         }
1512         state = rb_entry(node, struct extent_state, rb_node);
1513         if (state->start != start) {
1514                 ret = -ENOENT;
1515                 goto out;
1516         }
1517         state->private = private;
1518 out:
1519         spin_unlock(&tree->lock);
1520         return ret;
1521 }
1522
1523 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1524 {
1525         struct rb_node *node;
1526         struct extent_state *state;
1527         int ret = 0;
1528
1529         spin_lock(&tree->lock);
1530         /*
1531          * this search will find all the extents that end after
1532          * our range starts.
1533          */
1534         node = tree_search(tree, start);
1535         if (!node) {
1536                 ret = -ENOENT;
1537                 goto out;
1538         }
1539         state = rb_entry(node, struct extent_state, rb_node);
1540         if (state->start != start) {
1541                 ret = -ENOENT;
1542                 goto out;
1543         }
1544         *private = state->private;
1545 out:
1546         spin_unlock(&tree->lock);
1547         return ret;
1548 }
1549
1550 /*
1551  * searches a range in the state tree for a given mask.
1552  * If 'filled' == 1, this returns 1 only if every extent in the tree
1553  * has the bits set.  Otherwise, 1 is returned if any bit in the
1554  * range is found set.
1555  */
1556 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1557                    int bits, int filled)
1558 {
1559         struct extent_state *state = NULL;
1560         struct rb_node *node;
1561         int bitset = 0;
1562
1563         spin_lock(&tree->lock);
1564         node = tree_search(tree, start);
1565         while (node && start <= end) {
1566                 state = rb_entry(node, struct extent_state, rb_node);
1567
1568                 if (filled && state->start > start) {
1569                         bitset = 0;
1570                         break;
1571                 }
1572
1573                 if (state->start > end)
1574                         break;
1575
1576                 if (state->state & bits) {
1577                         bitset = 1;
1578                         if (!filled)
1579                                 break;
1580                 } else if (filled) {
1581                         bitset = 0;
1582                         break;
1583                 }
1584                 start = state->end + 1;
1585                 if (start > end)
1586                         break;
1587                 node = rb_next(node);
1588                 if (!node) {
1589                         if (filled)
1590                                 bitset = 0;
1591                         break;
1592                 }
1593         }
1594         spin_unlock(&tree->lock);
1595         return bitset;
1596 }
1597 EXPORT_SYMBOL(test_range_bit);
1598
1599 /*
1600  * helper function to set a given page up to date if all the
1601  * extents in the tree for that page are up to date
1602  */
1603 static int check_page_uptodate(struct extent_io_tree *tree,
1604                                struct page *page)
1605 {
1606         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1607         u64 end = start + PAGE_CACHE_SIZE - 1;
1608         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1))
1609                 SetPageUptodate(page);
1610         return 0;
1611 }
1612
1613 /*
1614  * helper function to unlock a page if all the extents in the tree
1615  * for that page are unlocked
1616  */
1617 static int check_page_locked(struct extent_io_tree *tree,
1618                              struct page *page)
1619 {
1620         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1621         u64 end = start + PAGE_CACHE_SIZE - 1;
1622         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0))
1623                 unlock_page(page);
1624         return 0;
1625 }
1626
1627 /*
1628  * helper function to end page writeback if all the extents
1629  * in the tree for that page are done with writeback
1630  */
1631 static int check_page_writeback(struct extent_io_tree *tree,
1632                              struct page *page)
1633 {
1634         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1635         u64 end = start + PAGE_CACHE_SIZE - 1;
1636         if (!test_range_bit(tree, start, end, EXTENT_WRITEBACK, 0))
1637                 end_page_writeback(page);
1638         return 0;
1639 }
1640
1641 /* lots and lots of room for performance fixes in the end_bio funcs */
1642
1643 /*
1644  * after a writepage IO is done, we need to:
1645  * clear the uptodate bits on error
1646  * clear the writeback bits in the extent tree for this IO
1647  * end_page_writeback if the page has no more pending IO
1648  *
1649  * Scheduling is not allowed, so the extent state tree is expected
1650  * to have one and only one object corresponding to this IO.
1651  */
1652 static void end_bio_extent_writepage(struct bio *bio, int err)
1653 {
1654         int uptodate = err == 0;
1655         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1656         struct extent_io_tree *tree;
1657         u64 start;
1658         u64 end;
1659         int whole_page;
1660         int ret;
1661
1662         do {
1663                 struct page *page = bvec->bv_page;
1664                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1665
1666                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1667                          bvec->bv_offset;
1668                 end = start + bvec->bv_len - 1;
1669
1670                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1671                         whole_page = 1;
1672                 else
1673                         whole_page = 0;
1674
1675                 if (--bvec >= bio->bi_io_vec)
1676                         prefetchw(&bvec->bv_page->flags);
1677                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1678                         ret = tree->ops->writepage_end_io_hook(page, start,
1679                                                        end, NULL, uptodate);
1680                         if (ret)
1681                                 uptodate = 0;
1682                 }
1683
1684                 if (!uptodate && tree->ops &&
1685                     tree->ops->writepage_io_failed_hook) {
1686                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1687                                                          start, end, NULL);
1688                         if (ret == 0) {
1689                                 uptodate = (err == 0);
1690                                 continue;
1691                         }
1692                 }
1693
1694                 if (!uptodate) {
1695                         clear_extent_uptodate(tree, start, end, GFP_ATOMIC);
1696                         ClearPageUptodate(page);
1697                         SetPageError(page);
1698                 }
1699
1700                 clear_extent_writeback(tree, start, end, GFP_ATOMIC);
1701
1702                 if (whole_page)
1703                         end_page_writeback(page);
1704                 else
1705                         check_page_writeback(tree, page);
1706         } while (bvec >= bio->bi_io_vec);
1707
1708         bio_put(bio);
1709 }
1710
1711 /*
1712  * after a readpage IO is done, we need to:
1713  * clear the uptodate bits on error
1714  * set the uptodate bits if things worked
1715  * set the page up to date if all extents in the tree are uptodate
1716  * clear the lock bit in the extent tree
1717  * unlock the page if there are no other extents locked for it
1718  *
1719  * Scheduling is not allowed, so the extent state tree is expected
1720  * to have one and only one object corresponding to this IO.
1721  */
1722 static void end_bio_extent_readpage(struct bio *bio, int err)
1723 {
1724         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1725         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1726         struct extent_io_tree *tree;
1727         u64 start;
1728         u64 end;
1729         int whole_page;
1730         int ret;
1731
1732         if (err)
1733                 uptodate = 0;
1734
1735         do {
1736                 struct page *page = bvec->bv_page;
1737                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1738
1739                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1740                         bvec->bv_offset;
1741                 end = start + bvec->bv_len - 1;
1742
1743                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1744                         whole_page = 1;
1745                 else
1746                         whole_page = 0;
1747
1748                 if (--bvec >= bio->bi_io_vec)
1749                         prefetchw(&bvec->bv_page->flags);
1750
1751                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1752                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1753                                                               NULL);
1754                         if (ret)
1755                                 uptodate = 0;
1756                 }
1757                 if (!uptodate && tree->ops &&
1758                     tree->ops->readpage_io_failed_hook) {
1759                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1760                                                          start, end, NULL);
1761                         if (ret == 0) {
1762                                 uptodate =
1763                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1764                                 if (err)
1765                                         uptodate = 0;
1766                                 continue;
1767                         }
1768                 }
1769
1770                 if (uptodate) {
1771                         set_extent_uptodate(tree, start, end,
1772                                             GFP_ATOMIC);
1773                 }
1774                 unlock_extent(tree, start, end, GFP_ATOMIC);
1775
1776                 if (whole_page) {
1777                         if (uptodate) {
1778                                 SetPageUptodate(page);
1779                         } else {
1780                                 ClearPageUptodate(page);
1781                                 SetPageError(page);
1782                         }
1783                         unlock_page(page);
1784                 } else {
1785                         if (uptodate) {
1786                                 check_page_uptodate(tree, page);
1787                         } else {
1788                                 ClearPageUptodate(page);
1789                                 SetPageError(page);
1790                         }
1791                         check_page_locked(tree, page);
1792                 }
1793         } while (bvec >= bio->bi_io_vec);
1794
1795         bio_put(bio);
1796 }
1797
1798 /*
1799  * IO done from prepare_write is pretty simple, we just unlock
1800  * the structs in the extent tree when done, and set the uptodate bits
1801  * as appropriate.
1802  */
1803 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1804 {
1805         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1806         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1807         struct extent_io_tree *tree;
1808         u64 start;
1809         u64 end;
1810
1811         do {
1812                 struct page *page = bvec->bv_page;
1813                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1814
1815                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1816                         bvec->bv_offset;
1817                 end = start + bvec->bv_len - 1;
1818
1819                 if (--bvec >= bio->bi_io_vec)
1820                         prefetchw(&bvec->bv_page->flags);
1821
1822                 if (uptodate) {
1823                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1824                 } else {
1825                         ClearPageUptodate(page);
1826                         SetPageError(page);
1827                 }
1828
1829                 unlock_extent(tree, start, end, GFP_ATOMIC);
1830
1831         } while (bvec >= bio->bi_io_vec);
1832
1833         bio_put(bio);
1834 }
1835
1836 static struct bio *
1837 extent_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1838                  gfp_t gfp_flags)
1839 {
1840         struct bio *bio;
1841
1842         bio = bio_alloc(gfp_flags, nr_vecs);
1843
1844         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1845                 while (!bio && (nr_vecs /= 2))
1846                         bio = bio_alloc(gfp_flags, nr_vecs);
1847         }
1848
1849         if (bio) {
1850                 bio->bi_size = 0;
1851                 bio->bi_bdev = bdev;
1852                 bio->bi_sector = first_sector;
1853         }
1854         return bio;
1855 }
1856
1857 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1858                           unsigned long bio_flags)
1859 {
1860         int ret = 0;
1861         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1862         struct page *page = bvec->bv_page;
1863         struct extent_io_tree *tree = bio->bi_private;
1864         u64 start;
1865         u64 end;
1866
1867         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1868         end = start + bvec->bv_len - 1;
1869
1870         bio->bi_private = NULL;
1871
1872         bio_get(bio);
1873
1874         if (tree->ops && tree->ops->submit_bio_hook)
1875                 tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1876                                            mirror_num, bio_flags);
1877         else
1878                 submit_bio(rw, bio);
1879         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1880                 ret = -EOPNOTSUPP;
1881         bio_put(bio);
1882         return ret;
1883 }
1884
1885 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1886                               struct page *page, sector_t sector,
1887                               size_t size, unsigned long offset,
1888                               struct block_device *bdev,
1889                               struct bio **bio_ret,
1890                               unsigned long max_pages,
1891                               bio_end_io_t end_io_func,
1892                               int mirror_num,
1893                               unsigned long prev_bio_flags,
1894                               unsigned long bio_flags)
1895 {
1896         int ret = 0;
1897         struct bio *bio;
1898         int nr;
1899         int contig = 0;
1900         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1901         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1902         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1903
1904         if (bio_ret && *bio_ret) {
1905                 bio = *bio_ret;
1906                 if (old_compressed)
1907                         contig = bio->bi_sector == sector;
1908                 else
1909                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
1910                                 sector;
1911
1912                 if (prev_bio_flags != bio_flags || !contig ||
1913                     (tree->ops && tree->ops->merge_bio_hook &&
1914                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
1915                                                bio_flags)) ||
1916                     bio_add_page(bio, page, page_size, offset) < page_size) {
1917                         ret = submit_one_bio(rw, bio, mirror_num,
1918                                              prev_bio_flags);
1919                         bio = NULL;
1920                 } else {
1921                         return 0;
1922                 }
1923         }
1924         if (this_compressed)
1925                 nr = BIO_MAX_PAGES;
1926         else
1927                 nr = bio_get_nr_vecs(bdev);
1928
1929         bio = extent_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1930         if (!bio) {
1931                 printk("failed to allocate bio nr %d\n", nr);
1932         }
1933
1934         bio_add_page(bio, page, page_size, offset);
1935         bio->bi_end_io = end_io_func;
1936         bio->bi_private = tree;
1937
1938         if (bio_ret) {
1939                 *bio_ret = bio;
1940         } else {
1941                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1942         }
1943
1944         return ret;
1945 }
1946
1947 void set_page_extent_mapped(struct page *page)
1948 {
1949         if (!PagePrivate(page)) {
1950                 SetPagePrivate(page);
1951                 page_cache_get(page);
1952                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1953         }
1954 }
1955 EXPORT_SYMBOL(set_page_extent_mapped);
1956
1957 static void set_page_extent_head(struct page *page, unsigned long len)
1958 {
1959         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1960 }
1961
1962 /*
1963  * basic readpage implementation.  Locked extent state structs are inserted
1964  * into the tree that are removed when the IO is done (by the end_io
1965  * handlers)
1966  */
1967 static int __extent_read_full_page(struct extent_io_tree *tree,
1968                                    struct page *page,
1969                                    get_extent_t *get_extent,
1970                                    struct bio **bio, int mirror_num,
1971                                    unsigned long *bio_flags)
1972 {
1973         struct inode *inode = page->mapping->host;
1974         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1975         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1976         u64 end;
1977         u64 cur = start;
1978         u64 extent_offset;
1979         u64 last_byte = i_size_read(inode);
1980         u64 block_start;
1981         u64 cur_end;
1982         sector_t sector;
1983         struct extent_map *em;
1984         struct block_device *bdev;
1985         int ret;
1986         int nr = 0;
1987         size_t page_offset = 0;
1988         size_t iosize;
1989         size_t disk_io_size;
1990         size_t blocksize = inode->i_sb->s_blocksize;
1991         unsigned long this_bio_flag = 0;
1992
1993         set_page_extent_mapped(page);
1994
1995         end = page_end;
1996         lock_extent(tree, start, end, GFP_NOFS);
1997
1998         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
1999                 char *userpage;
2000                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2001
2002                 if (zero_offset) {
2003                         iosize = PAGE_CACHE_SIZE - zero_offset;
2004                         userpage = kmap_atomic(page, KM_USER0);
2005                         memset(userpage + zero_offset, 0, iosize);
2006                         flush_dcache_page(page);
2007                         kunmap_atomic(userpage, KM_USER0);
2008                 }
2009         }
2010         while (cur <= end) {
2011                 if (cur >= last_byte) {
2012                         char *userpage;
2013                         iosize = PAGE_CACHE_SIZE - page_offset;
2014                         userpage = kmap_atomic(page, KM_USER0);
2015                         memset(userpage + page_offset, 0, iosize);
2016                         flush_dcache_page(page);
2017                         kunmap_atomic(userpage, KM_USER0);
2018                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2019                                             GFP_NOFS);
2020                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2021                         break;
2022                 }
2023                 em = get_extent(inode, page, page_offset, cur,
2024                                 end - cur + 1, 0);
2025                 if (IS_ERR(em) || !em) {
2026                         SetPageError(page);
2027                         unlock_extent(tree, cur, end, GFP_NOFS);
2028                         break;
2029                 }
2030                 extent_offset = cur - em->start;
2031                 if (extent_map_end(em) <= cur) {
2032 printk("bad mapping em [%Lu %Lu] cur %Lu\n", em->start, extent_map_end(em), cur);
2033                 }
2034                 BUG_ON(extent_map_end(em) <= cur);
2035                 if (end < cur) {
2036 printk("2bad mapping end %Lu cur %Lu\n", end, cur);
2037                 }
2038                 BUG_ON(end < cur);
2039
2040                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2041                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2042
2043                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2044                 cur_end = min(extent_map_end(em) - 1, end);
2045                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2046                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2047                         disk_io_size = em->block_len;
2048                         sector = em->block_start >> 9;
2049                 } else {
2050                         sector = (em->block_start + extent_offset) >> 9;
2051                         disk_io_size = iosize;
2052                 }
2053                 bdev = em->bdev;
2054                 block_start = em->block_start;
2055                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2056                         block_start = EXTENT_MAP_HOLE;
2057                 free_extent_map(em);
2058                 em = NULL;
2059
2060                 /* we've found a hole, just zero and go on */
2061                 if (block_start == EXTENT_MAP_HOLE) {
2062                         char *userpage;
2063                         userpage = kmap_atomic(page, KM_USER0);
2064                         memset(userpage + page_offset, 0, iosize);
2065                         flush_dcache_page(page);
2066                         kunmap_atomic(userpage, KM_USER0);
2067
2068                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2069                                             GFP_NOFS);
2070                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2071                         cur = cur + iosize;
2072                         page_offset += iosize;
2073                         continue;
2074                 }
2075                 /* the get_extent function already copied into the page */
2076                 if (test_range_bit(tree, cur, cur_end, EXTENT_UPTODATE, 1)) {
2077                         check_page_uptodate(tree, page);
2078                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2079                         cur = cur + iosize;
2080                         page_offset += iosize;
2081                         continue;
2082                 }
2083                 /* we have an inline extent but it didn't get marked up
2084                  * to date.  Error out
2085                  */
2086                 if (block_start == EXTENT_MAP_INLINE) {
2087                         SetPageError(page);
2088                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2089                         cur = cur + iosize;
2090                         page_offset += iosize;
2091                         continue;
2092                 }
2093
2094                 ret = 0;
2095                 if (tree->ops && tree->ops->readpage_io_hook) {
2096                         ret = tree->ops->readpage_io_hook(page, cur,
2097                                                           cur + iosize - 1);
2098                 }
2099                 if (!ret) {
2100                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2101                         pnr -= page->index;
2102                         ret = submit_extent_page(READ, tree, page,
2103                                          sector, disk_io_size, page_offset,
2104                                          bdev, bio, pnr,
2105                                          end_bio_extent_readpage, mirror_num,
2106                                          *bio_flags,
2107                                          this_bio_flag);
2108                         nr++;
2109                         *bio_flags = this_bio_flag;
2110                 }
2111                 if (ret)
2112                         SetPageError(page);
2113                 cur = cur + iosize;
2114                 page_offset += iosize;
2115         }
2116         if (!nr) {
2117                 if (!PageError(page))
2118                         SetPageUptodate(page);
2119                 unlock_page(page);
2120         }
2121         return 0;
2122 }
2123
2124 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2125                             get_extent_t *get_extent)
2126 {
2127         struct bio *bio = NULL;
2128         unsigned long bio_flags = 0;
2129         int ret;
2130
2131         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2132                                       &bio_flags);
2133         if (bio)
2134                 submit_one_bio(READ, bio, 0, bio_flags);
2135         return ret;
2136 }
2137 EXPORT_SYMBOL(extent_read_full_page);
2138
2139 /*
2140  * the writepage semantics are similar to regular writepage.  extent
2141  * records are inserted to lock ranges in the tree, and as dirty areas
2142  * are found, they are marked writeback.  Then the lock bits are removed
2143  * and the end_io handler clears the writeback ranges
2144  */
2145 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2146                               void *data)
2147 {
2148         struct inode *inode = page->mapping->host;
2149         struct extent_page_data *epd = data;
2150         struct extent_io_tree *tree = epd->tree;
2151         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2152         u64 delalloc_start;
2153         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2154         u64 end;
2155         u64 cur = start;
2156         u64 extent_offset;
2157         u64 last_byte = i_size_read(inode);
2158         u64 block_start;
2159         u64 iosize;
2160         u64 unlock_start;
2161         sector_t sector;
2162         struct extent_map *em;
2163         struct block_device *bdev;
2164         int ret;
2165         int nr = 0;
2166         size_t pg_offset = 0;
2167         size_t blocksize;
2168         loff_t i_size = i_size_read(inode);
2169         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2170         u64 nr_delalloc;
2171         u64 delalloc_end;
2172         int page_started;
2173         int compressed;
2174         unsigned long nr_written = 0;
2175
2176         WARN_ON(!PageLocked(page));
2177         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2178         if (page->index > end_index ||
2179            (page->index == end_index && !pg_offset)) {
2180                 page->mapping->a_ops->invalidatepage(page, 0);
2181                 unlock_page(page);
2182                 return 0;
2183         }
2184
2185         if (page->index == end_index) {
2186                 char *userpage;
2187
2188                 userpage = kmap_atomic(page, KM_USER0);
2189                 memset(userpage + pg_offset, 0,
2190                        PAGE_CACHE_SIZE - pg_offset);
2191                 kunmap_atomic(userpage, KM_USER0);
2192                 flush_dcache_page(page);
2193         }
2194         pg_offset = 0;
2195
2196         set_page_extent_mapped(page);
2197
2198         delalloc_start = start;
2199         delalloc_end = 0;
2200         page_started = 0;
2201         if (!epd->extent_locked) {
2202                 while(delalloc_end < page_end) {
2203                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2204                                                        page,
2205                                                        &delalloc_start,
2206                                                        &delalloc_end,
2207                                                        128 * 1024 * 1024);
2208                         if (nr_delalloc == 0) {
2209                                 delalloc_start = delalloc_end + 1;
2210                                 continue;
2211                         }
2212                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2213                                                  delalloc_end, &page_started,
2214                                                  &nr_written);
2215                         delalloc_start = delalloc_end + 1;
2216                 }
2217
2218                 /* did the fill delalloc function already unlock and start
2219                  * the IO?
2220                  */
2221                 if (page_started) {
2222                         ret = 0;
2223                         goto update_nr_written;
2224                 }
2225         }
2226         lock_extent(tree, start, page_end, GFP_NOFS);
2227
2228         unlock_start = start;
2229
2230         if (tree->ops && tree->ops->writepage_start_hook) {
2231                 ret = tree->ops->writepage_start_hook(page, start,
2232                                                       page_end);
2233                 if (ret == -EAGAIN) {
2234                         unlock_extent(tree, start, page_end, GFP_NOFS);
2235                         redirty_page_for_writepage(wbc, page);
2236                         unlock_page(page);
2237                         ret = 0;
2238                         goto update_nr_written;
2239                 }
2240         }
2241
2242         nr_written++;
2243
2244         end = page_end;
2245         if (test_range_bit(tree, start, page_end, EXTENT_DELALLOC, 0)) {
2246                 printk("found delalloc bits after lock_extent\n");
2247         }
2248
2249         if (last_byte <= start) {
2250                 clear_extent_dirty(tree, start, page_end, GFP_NOFS);
2251                 unlock_extent(tree, start, page_end, GFP_NOFS);
2252                 if (tree->ops && tree->ops->writepage_end_io_hook)
2253                         tree->ops->writepage_end_io_hook(page, start,
2254                                                          page_end, NULL, 1);
2255                 unlock_start = page_end + 1;
2256                 goto done;
2257         }
2258
2259         set_extent_uptodate(tree, start, page_end, GFP_NOFS);
2260         blocksize = inode->i_sb->s_blocksize;
2261
2262         while (cur <= end) {
2263                 if (cur >= last_byte) {
2264                         clear_extent_dirty(tree, cur, page_end, GFP_NOFS);
2265                         unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
2266                         if (tree->ops && tree->ops->writepage_end_io_hook)
2267                                 tree->ops->writepage_end_io_hook(page, cur,
2268                                                          page_end, NULL, 1);
2269                         unlock_start = page_end + 1;
2270                         break;
2271                 }
2272                 em = epd->get_extent(inode, page, pg_offset, cur,
2273                                      end - cur + 1, 1);
2274                 if (IS_ERR(em) || !em) {
2275                         SetPageError(page);
2276                         break;
2277                 }
2278
2279                 extent_offset = cur - em->start;
2280                 BUG_ON(extent_map_end(em) <= cur);
2281                 BUG_ON(end < cur);
2282                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2283                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2284                 sector = (em->block_start + extent_offset) >> 9;
2285                 bdev = em->bdev;
2286                 block_start = em->block_start;
2287                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2288                 free_extent_map(em);
2289                 em = NULL;
2290
2291                 /*
2292                  * compressed and inline extents are written through other
2293                  * paths in the FS
2294                  */
2295                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2296                     block_start == EXTENT_MAP_INLINE) {
2297                         clear_extent_dirty(tree, cur,
2298                                            cur + iosize - 1, GFP_NOFS);
2299
2300                         unlock_extent(tree, unlock_start, cur + iosize -1,
2301                                       GFP_NOFS);
2302
2303                         /*
2304                          * end_io notification does not happen here for
2305                          * compressed extents
2306                          */
2307                         if (!compressed && tree->ops &&
2308                             tree->ops->writepage_end_io_hook)
2309                                 tree->ops->writepage_end_io_hook(page, cur,
2310                                                          cur + iosize - 1,
2311                                                          NULL, 1);
2312                         else if (compressed) {
2313                                 /* we don't want to end_page_writeback on
2314                                  * a compressed extent.  this happens
2315                                  * elsewhere
2316                                  */
2317                                 nr++;
2318                         }
2319
2320                         cur += iosize;
2321                         pg_offset += iosize;
2322                         unlock_start = cur;
2323                         continue;
2324                 }
2325                 /* leave this out until we have a page_mkwrite call */
2326                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2327                                    EXTENT_DIRTY, 0)) {
2328                         cur = cur + iosize;
2329                         pg_offset += iosize;
2330                         continue;
2331                 }
2332
2333                 clear_extent_dirty(tree, cur, cur + iosize - 1, GFP_NOFS);
2334                 if (tree->ops && tree->ops->writepage_io_hook) {
2335                         ret = tree->ops->writepage_io_hook(page, cur,
2336                                                 cur + iosize - 1);
2337                 } else {
2338                         ret = 0;
2339                 }
2340                 if (ret) {
2341                         SetPageError(page);
2342                 } else {
2343                         unsigned long max_nr = end_index + 1;
2344
2345                         set_range_writeback(tree, cur, cur + iosize - 1);
2346                         if (!PageWriteback(page)) {
2347                                 printk("warning page %lu not writeback, "
2348                                        "cur %llu end %llu\n", page->index,
2349                                        (unsigned long long)cur,
2350                                        (unsigned long long)end);
2351                         }
2352
2353                         ret = submit_extent_page(WRITE, tree, page, sector,
2354                                                  iosize, pg_offset, bdev,
2355                                                  &epd->bio, max_nr,
2356                                                  end_bio_extent_writepage,
2357                                                  0, 0, 0);
2358                         if (ret)
2359                                 SetPageError(page);
2360                 }
2361                 cur = cur + iosize;
2362                 pg_offset += iosize;
2363                 nr++;
2364         }
2365 done:
2366         if (nr == 0) {
2367                 /* make sure the mapping tag for page dirty gets cleared */
2368                 set_page_writeback(page);
2369                 end_page_writeback(page);
2370         }
2371         if (unlock_start <= page_end)
2372                 unlock_extent(tree, unlock_start, page_end, GFP_NOFS);
2373         unlock_page(page);
2374
2375 update_nr_written:
2376         wbc->nr_to_write -= nr_written;
2377         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2378             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2379                 page->mapping->writeback_index = page->index + nr_written;
2380         return 0;
2381 }
2382
2383 /**
2384  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2385  * @mapping: address space structure to write
2386  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2387  * @writepage: function called for each page
2388  * @data: data passed to writepage function
2389  *
2390  * If a page is already under I/O, write_cache_pages() skips it, even
2391  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2392  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2393  * and msync() need to guarantee that all the data which was dirty at the time
2394  * the call was made get new I/O started against them.  If wbc->sync_mode is
2395  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2396  * existing IO to complete.
2397  */
2398 static int extent_write_cache_pages(struct extent_io_tree *tree,
2399                              struct address_space *mapping,
2400                              struct writeback_control *wbc,
2401                              writepage_t writepage, void *data,
2402                              void (*flush_fn)(void *))
2403 {
2404         struct backing_dev_info *bdi = mapping->backing_dev_info;
2405         int ret = 0;
2406         int done = 0;
2407         struct pagevec pvec;
2408         int nr_pages;
2409         pgoff_t index;
2410         pgoff_t end;            /* Inclusive */
2411         int scanned = 0;
2412         int range_whole = 0;
2413
2414         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2415                 wbc->encountered_congestion = 1;
2416                 return 0;
2417         }
2418
2419         pagevec_init(&pvec, 0);
2420         if (wbc->range_cyclic) {
2421                 index = mapping->writeback_index; /* Start from prev offset */
2422                 end = -1;
2423         } else {
2424                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2425                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2426                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2427                         range_whole = 1;
2428                 scanned = 1;
2429         }
2430 retry:
2431         while (!done && (index <= end) &&
2432                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2433                                               PAGECACHE_TAG_DIRTY,
2434                                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2435                 unsigned i;
2436
2437                 scanned = 1;
2438                 for (i = 0; i < nr_pages; i++) {
2439                         struct page *page = pvec.pages[i];
2440
2441                         /*
2442                          * At this point we hold neither mapping->tree_lock nor
2443                          * lock on the page itself: the page may be truncated or
2444                          * invalidated (changing page->mapping to NULL), or even
2445                          * swizzled back from swapper_space to tmpfs file
2446                          * mapping
2447                          */
2448                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2449                                 tree->ops->write_cache_pages_lock_hook(page);
2450                         else
2451                                 lock_page(page);
2452
2453                         if (unlikely(page->mapping != mapping)) {
2454                                 unlock_page(page);
2455                                 continue;
2456                         }
2457
2458                         if (!wbc->range_cyclic && page->index > end) {
2459                                 done = 1;
2460                                 unlock_page(page);
2461                                 continue;
2462                         }
2463
2464                         if (wbc->sync_mode != WB_SYNC_NONE) {
2465                                 if (PageWriteback(page))
2466                                         flush_fn(data);
2467                                 wait_on_page_writeback(page);
2468                         }
2469
2470                         if (PageWriteback(page) ||
2471                             !clear_page_dirty_for_io(page)) {
2472                                 unlock_page(page);
2473                                 continue;
2474                         }
2475
2476                         ret = (*writepage)(page, wbc, data);
2477
2478                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2479                                 unlock_page(page);
2480                                 ret = 0;
2481                         }
2482                         if (ret || wbc->nr_to_write <= 0)
2483                                 done = 1;
2484                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
2485                                 wbc->encountered_congestion = 1;
2486                                 done = 1;
2487                         }
2488                 }
2489                 pagevec_release(&pvec);
2490                 cond_resched();
2491         }
2492         if (!scanned && !done) {
2493                 /*
2494                  * We hit the last page and there is more work to be done: wrap
2495                  * back to the start of the file
2496                  */
2497                 scanned = 1;
2498                 index = 0;
2499                 goto retry;
2500         }
2501         return ret;
2502 }
2503
2504 static noinline void flush_write_bio(void *data)
2505 {
2506         struct extent_page_data *epd = data;
2507         if (epd->bio) {
2508                 submit_one_bio(WRITE, epd->bio, 0, 0);
2509                 epd->bio = NULL;
2510         }
2511 }
2512
2513 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2514                           get_extent_t *get_extent,
2515                           struct writeback_control *wbc)
2516 {
2517         int ret;
2518         struct address_space *mapping = page->mapping;
2519         struct extent_page_data epd = {
2520                 .bio = NULL,
2521                 .tree = tree,
2522                 .get_extent = get_extent,
2523                 .extent_locked = 0,
2524         };
2525         struct writeback_control wbc_writepages = {
2526                 .bdi            = wbc->bdi,
2527                 .sync_mode      = WB_SYNC_NONE,
2528                 .older_than_this = NULL,
2529                 .nr_to_write    = 64,
2530                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2531                 .range_end      = (loff_t)-1,
2532         };
2533
2534
2535         ret = __extent_writepage(page, wbc, &epd);
2536
2537         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2538                                  __extent_writepage, &epd, flush_write_bio);
2539         if (epd.bio) {
2540                 submit_one_bio(WRITE, epd.bio, 0, 0);
2541         }
2542         return ret;
2543 }
2544 EXPORT_SYMBOL(extent_write_full_page);
2545
2546 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2547                               u64 start, u64 end, get_extent_t *get_extent,
2548                               int mode)
2549 {
2550         int ret = 0;
2551         struct address_space *mapping = inode->i_mapping;
2552         struct page *page;
2553         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2554                 PAGE_CACHE_SHIFT;
2555
2556         struct extent_page_data epd = {
2557                 .bio = NULL,
2558                 .tree = tree,
2559                 .get_extent = get_extent,
2560                 .extent_locked = 1,
2561         };
2562         struct writeback_control wbc_writepages = {
2563                 .bdi            = inode->i_mapping->backing_dev_info,
2564                 .sync_mode      = mode,
2565                 .older_than_this = NULL,
2566                 .nr_to_write    = nr_pages * 2,
2567                 .range_start    = start,
2568                 .range_end      = end + 1,
2569         };
2570
2571         while(start <= end) {
2572                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2573                 if (clear_page_dirty_for_io(page))
2574                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2575                 else {
2576                         if (tree->ops && tree->ops->writepage_end_io_hook)
2577                                 tree->ops->writepage_end_io_hook(page, start,
2578                                                  start + PAGE_CACHE_SIZE - 1,
2579                                                  NULL, 1);
2580                         unlock_page(page);
2581                 }
2582                 page_cache_release(page);
2583                 start += PAGE_CACHE_SIZE;
2584         }
2585
2586         if (epd.bio)
2587                 submit_one_bio(WRITE, epd.bio, 0, 0);
2588         return ret;
2589 }
2590 EXPORT_SYMBOL(extent_write_locked_range);
2591
2592
2593 int extent_writepages(struct extent_io_tree *tree,
2594                       struct address_space *mapping,
2595                       get_extent_t *get_extent,
2596                       struct writeback_control *wbc)
2597 {
2598         int ret = 0;
2599         struct extent_page_data epd = {
2600                 .bio = NULL,
2601                 .tree = tree,
2602                 .get_extent = get_extent,
2603                 .extent_locked = 0,
2604         };
2605
2606         ret = extent_write_cache_pages(tree, mapping, wbc,
2607                                        __extent_writepage, &epd,
2608                                        flush_write_bio);
2609         if (epd.bio) {
2610                 submit_one_bio(WRITE, epd.bio, 0, 0);
2611         }
2612         return ret;
2613 }
2614 EXPORT_SYMBOL(extent_writepages);
2615
2616 int extent_readpages(struct extent_io_tree *tree,
2617                      struct address_space *mapping,
2618                      struct list_head *pages, unsigned nr_pages,
2619                      get_extent_t get_extent)
2620 {
2621         struct bio *bio = NULL;
2622         unsigned page_idx;
2623         struct pagevec pvec;
2624         unsigned long bio_flags = 0;
2625
2626         pagevec_init(&pvec, 0);
2627         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2628                 struct page *page = list_entry(pages->prev, struct page, lru);
2629
2630                 prefetchw(&page->flags);
2631                 list_del(&page->lru);
2632                 /*
2633                  * what we want to do here is call add_to_page_cache_lru,
2634                  * but that isn't exported, so we reproduce it here
2635                  */
2636                 if (!add_to_page_cache(page, mapping,
2637                                         page->index, GFP_KERNEL)) {
2638
2639                         /* open coding of lru_cache_add, also not exported */
2640                         page_cache_get(page);
2641                         if (!pagevec_add(&pvec, page))
2642                                 __pagevec_lru_add_file(&pvec);
2643                         __extent_read_full_page(tree, page, get_extent,
2644                                                 &bio, 0, &bio_flags);
2645                 }
2646                 page_cache_release(page);
2647         }
2648         if (pagevec_count(&pvec))
2649                 __pagevec_lru_add_file(&pvec);
2650         BUG_ON(!list_empty(pages));
2651         if (bio)
2652                 submit_one_bio(READ, bio, 0, bio_flags);
2653         return 0;
2654 }
2655 EXPORT_SYMBOL(extent_readpages);
2656
2657 /*
2658  * basic invalidatepage code, this waits on any locked or writeback
2659  * ranges corresponding to the page, and then deletes any extent state
2660  * records from the tree
2661  */
2662 int extent_invalidatepage(struct extent_io_tree *tree,
2663                           struct page *page, unsigned long offset)
2664 {
2665         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2666         u64 end = start + PAGE_CACHE_SIZE - 1;
2667         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2668
2669         start += (offset + blocksize -1) & ~(blocksize - 1);
2670         if (start > end)
2671                 return 0;
2672
2673         lock_extent(tree, start, end, GFP_NOFS);
2674         wait_on_extent_writeback(tree, start, end);
2675         clear_extent_bit(tree, start, end,
2676                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC,
2677                          1, 1, GFP_NOFS);
2678         return 0;
2679 }
2680 EXPORT_SYMBOL(extent_invalidatepage);
2681
2682 /*
2683  * simple commit_write call, set_range_dirty is used to mark both
2684  * the pages and the extent records as dirty
2685  */
2686 int extent_commit_write(struct extent_io_tree *tree,
2687                         struct inode *inode, struct page *page,
2688                         unsigned from, unsigned to)
2689 {
2690         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2691
2692         set_page_extent_mapped(page);
2693         set_page_dirty(page);
2694
2695         if (pos > inode->i_size) {
2696                 i_size_write(inode, pos);
2697                 mark_inode_dirty(inode);
2698         }
2699         return 0;
2700 }
2701 EXPORT_SYMBOL(extent_commit_write);
2702
2703 int extent_prepare_write(struct extent_io_tree *tree,
2704                          struct inode *inode, struct page *page,
2705                          unsigned from, unsigned to, get_extent_t *get_extent)
2706 {
2707         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2708         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2709         u64 block_start;
2710         u64 orig_block_start;
2711         u64 block_end;
2712         u64 cur_end;
2713         struct extent_map *em;
2714         unsigned blocksize = 1 << inode->i_blkbits;
2715         size_t page_offset = 0;
2716         size_t block_off_start;
2717         size_t block_off_end;
2718         int err = 0;
2719         int iocount = 0;
2720         int ret = 0;
2721         int isnew;
2722
2723         set_page_extent_mapped(page);
2724
2725         block_start = (page_start + from) & ~((u64)blocksize - 1);
2726         block_end = (page_start + to - 1) | (blocksize - 1);
2727         orig_block_start = block_start;
2728
2729         lock_extent(tree, page_start, page_end, GFP_NOFS);
2730         while(block_start <= block_end) {
2731                 em = get_extent(inode, page, page_offset, block_start,
2732                                 block_end - block_start + 1, 1);
2733                 if (IS_ERR(em) || !em) {
2734                         goto err;
2735                 }
2736                 cur_end = min(block_end, extent_map_end(em) - 1);
2737                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2738                 block_off_end = block_off_start + blocksize;
2739                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2740
2741                 if (!PageUptodate(page) && isnew &&
2742                     (block_off_end > to || block_off_start < from)) {
2743                         void *kaddr;
2744
2745                         kaddr = kmap_atomic(page, KM_USER0);
2746                         if (block_off_end > to)
2747                                 memset(kaddr + to, 0, block_off_end - to);
2748                         if (block_off_start < from)
2749                                 memset(kaddr + block_off_start, 0,
2750                                        from - block_off_start);
2751                         flush_dcache_page(page);
2752                         kunmap_atomic(kaddr, KM_USER0);
2753                 }
2754                 if ((em->block_start != EXTENT_MAP_HOLE &&
2755                      em->block_start != EXTENT_MAP_INLINE) &&
2756                     !isnew && !PageUptodate(page) &&
2757                     (block_off_end > to || block_off_start < from) &&
2758                     !test_range_bit(tree, block_start, cur_end,
2759                                     EXTENT_UPTODATE, 1)) {
2760                         u64 sector;
2761                         u64 extent_offset = block_start - em->start;
2762                         size_t iosize;
2763                         sector = (em->block_start + extent_offset) >> 9;
2764                         iosize = (cur_end - block_start + blocksize) &
2765                                 ~((u64)blocksize - 1);
2766                         /*
2767                          * we've already got the extent locked, but we
2768                          * need to split the state such that our end_bio
2769                          * handler can clear the lock.
2770                          */
2771                         set_extent_bit(tree, block_start,
2772                                        block_start + iosize - 1,
2773                                        EXTENT_LOCKED, 0, NULL, GFP_NOFS);
2774                         ret = submit_extent_page(READ, tree, page,
2775                                          sector, iosize, page_offset, em->bdev,
2776                                          NULL, 1,
2777                                          end_bio_extent_preparewrite, 0,
2778                                          0, 0);
2779                         iocount++;
2780                         block_start = block_start + iosize;
2781                 } else {
2782                         set_extent_uptodate(tree, block_start, cur_end,
2783                                             GFP_NOFS);
2784                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2785                         block_start = cur_end + 1;
2786                 }
2787                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2788                 free_extent_map(em);
2789         }
2790         if (iocount) {
2791                 wait_extent_bit(tree, orig_block_start,
2792                                 block_end, EXTENT_LOCKED);
2793         }
2794         check_page_uptodate(tree, page);
2795 err:
2796         /* FIXME, zero out newly allocated blocks on error */
2797         return err;
2798 }
2799 EXPORT_SYMBOL(extent_prepare_write);
2800
2801 /*
2802  * a helper for releasepage, this tests for areas of the page that
2803  * are locked or under IO and drops the related state bits if it is safe
2804  * to drop the page.
2805  */
2806 int try_release_extent_state(struct extent_map_tree *map,
2807                              struct extent_io_tree *tree, struct page *page,
2808                              gfp_t mask)
2809 {
2810         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2811         u64 end = start + PAGE_CACHE_SIZE - 1;
2812         int ret = 1;
2813
2814         if (test_range_bit(tree, start, end,
2815                            EXTENT_IOBITS | EXTENT_ORDERED, 0))
2816                 ret = 0;
2817         else {
2818                 if ((mask & GFP_NOFS) == GFP_NOFS)
2819                         mask = GFP_NOFS;
2820                 clear_extent_bit(tree, start, end, EXTENT_UPTODATE,
2821                                  1, 1, mask);
2822         }
2823         return ret;
2824 }
2825 EXPORT_SYMBOL(try_release_extent_state);
2826
2827 /*
2828  * a helper for releasepage.  As long as there are no locked extents
2829  * in the range corresponding to the page, both state records and extent
2830  * map records are removed
2831  */
2832 int try_release_extent_mapping(struct extent_map_tree *map,
2833                                struct extent_io_tree *tree, struct page *page,
2834                                gfp_t mask)
2835 {
2836         struct extent_map *em;
2837         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2838         u64 end = start + PAGE_CACHE_SIZE - 1;
2839
2840         if ((mask & __GFP_WAIT) &&
2841             page->mapping->host->i_size > 16 * 1024 * 1024) {
2842                 u64 len;
2843                 while (start <= end) {
2844                         len = end - start + 1;
2845                         spin_lock(&map->lock);
2846                         em = lookup_extent_mapping(map, start, len);
2847                         if (!em || IS_ERR(em)) {
2848                                 spin_unlock(&map->lock);
2849                                 break;
2850                         }
2851                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2852                             em->start != start) {
2853                                 spin_unlock(&map->lock);
2854                                 free_extent_map(em);
2855                                 break;
2856                         }
2857                         if (!test_range_bit(tree, em->start,
2858                                             extent_map_end(em) - 1,
2859                                             EXTENT_LOCKED | EXTENT_WRITEBACK |
2860                                             EXTENT_ORDERED,
2861                                             0)) {
2862                                 remove_extent_mapping(map, em);
2863                                 /* once for the rb tree */
2864                                 free_extent_map(em);
2865                         }
2866                         start = extent_map_end(em);
2867                         spin_unlock(&map->lock);
2868
2869                         /* once for us */
2870                         free_extent_map(em);
2871                 }
2872         }
2873         return try_release_extent_state(map, tree, page, mask);
2874 }
2875 EXPORT_SYMBOL(try_release_extent_mapping);
2876
2877 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2878                 get_extent_t *get_extent)
2879 {
2880         struct inode *inode = mapping->host;
2881         u64 start = iblock << inode->i_blkbits;
2882         sector_t sector = 0;
2883         size_t blksize = (1 << inode->i_blkbits);
2884         struct extent_map *em;
2885
2886         lock_extent(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2887                     GFP_NOFS);
2888         em = get_extent(inode, NULL, 0, start, blksize, 0);
2889         unlock_extent(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2890                       GFP_NOFS);
2891         if (!em || IS_ERR(em))
2892                 return 0;
2893
2894         if (em->block_start > EXTENT_MAP_LAST_BYTE)
2895                 goto out;
2896
2897         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2898 out:
2899         free_extent_map(em);
2900         return sector;
2901 }
2902
2903 static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2904                                               unsigned long i)
2905 {
2906         struct page *p;
2907         struct address_space *mapping;
2908
2909         if (i == 0)
2910                 return eb->first_page;
2911         i += eb->start >> PAGE_CACHE_SHIFT;
2912         mapping = eb->first_page->mapping;
2913         if (!mapping)
2914                 return NULL;
2915
2916         /*
2917          * extent_buffer_page is only called after pinning the page
2918          * by increasing the reference count.  So we know the page must
2919          * be in the radix tree.
2920          */
2921         rcu_read_lock();
2922         p = radix_tree_lookup(&mapping->page_tree, i);
2923         rcu_read_unlock();
2924
2925         return p;
2926 }
2927
2928 static inline unsigned long num_extent_pages(u64 start, u64 len)
2929 {
2930         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
2931                 (start >> PAGE_CACHE_SHIFT);
2932 }
2933
2934 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
2935                                                    u64 start,
2936                                                    unsigned long len,
2937                                                    gfp_t mask)
2938 {
2939         struct extent_buffer *eb = NULL;
2940 #ifdef LEAK_DEBUG
2941         unsigned long flags;
2942 #endif
2943
2944         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
2945         eb->start = start;
2946         eb->len = len;
2947         mutex_init(&eb->mutex);
2948 #ifdef LEAK_DEBUG
2949         spin_lock_irqsave(&leak_lock, flags);
2950         list_add(&eb->leak_list, &buffers);
2951         spin_unlock_irqrestore(&leak_lock, flags);
2952 #endif
2953         atomic_set(&eb->refs, 1);
2954
2955         return eb;
2956 }
2957
2958 static void __free_extent_buffer(struct extent_buffer *eb)
2959 {
2960 #ifdef LEAK_DEBUG
2961         unsigned long flags;
2962         spin_lock_irqsave(&leak_lock, flags);
2963         list_del(&eb->leak_list);
2964         spin_unlock_irqrestore(&leak_lock, flags);
2965 #endif
2966         kmem_cache_free(extent_buffer_cache, eb);
2967 }
2968
2969 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
2970                                           u64 start, unsigned long len,
2971                                           struct page *page0,
2972                                           gfp_t mask)
2973 {
2974         unsigned long num_pages = num_extent_pages(start, len);
2975         unsigned long i;
2976         unsigned long index = start >> PAGE_CACHE_SHIFT;
2977         struct extent_buffer *eb;
2978         struct extent_buffer *exists = NULL;
2979         struct page *p;
2980         struct address_space *mapping = tree->mapping;
2981         int uptodate = 1;
2982
2983         spin_lock(&tree->buffer_lock);
2984         eb = buffer_search(tree, start);
2985         if (eb) {
2986                 atomic_inc(&eb->refs);
2987                 spin_unlock(&tree->buffer_lock);
2988                 mark_page_accessed(eb->first_page);
2989                 return eb;
2990         }
2991         spin_unlock(&tree->buffer_lock);
2992
2993         eb = __alloc_extent_buffer(tree, start, len, mask);
2994         if (!eb)
2995                 return NULL;
2996
2997         if (page0) {
2998                 eb->first_page = page0;
2999                 i = 1;
3000                 index++;
3001                 page_cache_get(page0);
3002                 mark_page_accessed(page0);
3003                 set_page_extent_mapped(page0);
3004                 set_page_extent_head(page0, len);
3005                 uptodate = PageUptodate(page0);
3006         } else {
3007                 i = 0;
3008         }
3009         for (; i < num_pages; i++, index++) {
3010                 p = find_or_create_page(mapping, index, mask | __GFP_HIGHMEM);
3011                 if (!p) {
3012                         WARN_ON(1);
3013                         goto free_eb;
3014                 }
3015                 set_page_extent_mapped(p);
3016                 mark_page_accessed(p);
3017                 if (i == 0) {
3018                         eb->first_page = p;
3019                         set_page_extent_head(p, len);
3020                 } else {
3021                         set_page_private(p, EXTENT_PAGE_PRIVATE);
3022                 }
3023                 if (!PageUptodate(p))
3024                         uptodate = 0;
3025                 unlock_page(p);
3026         }
3027         if (uptodate)
3028                 eb->flags |= EXTENT_UPTODATE;
3029         eb->flags |= EXTENT_BUFFER_FILLED;
3030
3031         spin_lock(&tree->buffer_lock);
3032         exists = buffer_tree_insert(tree, start, &eb->rb_node);
3033         if (exists) {
3034                 /* add one reference for the caller */
3035                 atomic_inc(&exists->refs);
3036                 spin_unlock(&tree->buffer_lock);
3037                 goto free_eb;
3038         }
3039         spin_unlock(&tree->buffer_lock);
3040
3041         /* add one reference for the tree */
3042         atomic_inc(&eb->refs);
3043         return eb;
3044
3045 free_eb:
3046         if (!atomic_dec_and_test(&eb->refs))
3047                 return exists;
3048         for (index = 1; index < i; index++)
3049                 page_cache_release(extent_buffer_page(eb, index));
3050         page_cache_release(extent_buffer_page(eb, 0));
3051         __free_extent_buffer(eb);
3052         return exists;
3053 }
3054 EXPORT_SYMBOL(alloc_extent_buffer);
3055
3056 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3057                                          u64 start, unsigned long len,
3058                                           gfp_t mask)
3059 {
3060         struct extent_buffer *eb;
3061
3062         spin_lock(&tree->buffer_lock);
3063         eb = buffer_search(tree, start);
3064         if (eb)
3065                 atomic_inc(&eb->refs);
3066         spin_unlock(&tree->buffer_lock);
3067
3068         if (eb)
3069                 mark_page_accessed(eb->first_page);
3070
3071         return eb;
3072 }
3073 EXPORT_SYMBOL(find_extent_buffer);
3074
3075 void free_extent_buffer(struct extent_buffer *eb)
3076 {
3077         if (!eb)
3078                 return;
3079
3080         if (!atomic_dec_and_test(&eb->refs))
3081                 return;
3082
3083         WARN_ON(1);
3084 }
3085 EXPORT_SYMBOL(free_extent_buffer);
3086
3087 int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3088                               struct extent_buffer *eb)
3089 {
3090         int set;
3091         unsigned long i;
3092         unsigned long num_pages;
3093         struct page *page;
3094
3095         u64 start = eb->start;
3096         u64 end = start + eb->len - 1;
3097
3098         set = clear_extent_dirty(tree, start, end, GFP_NOFS);
3099         num_pages = num_extent_pages(eb->start, eb->len);
3100
3101         for (i = 0; i < num_pages; i++) {
3102                 page = extent_buffer_page(eb, i);
3103                 if (!set && !PageDirty(page))
3104                         continue;
3105
3106                 lock_page(page);
3107                 if (i == 0)
3108                         set_page_extent_head(page, eb->len);
3109                 else
3110                         set_page_private(page, EXTENT_PAGE_PRIVATE);
3111
3112                 /*
3113                  * if we're on the last page or the first page and the
3114                  * block isn't aligned on a page boundary, do extra checks
3115                  * to make sure we don't clean page that is partially dirty
3116                  */
3117                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3118                     ((i == num_pages - 1) &&
3119                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3120                         start = (u64)page->index << PAGE_CACHE_SHIFT;
3121                         end  = start + PAGE_CACHE_SIZE - 1;
3122                         if (test_range_bit(tree, start, end,
3123                                            EXTENT_DIRTY, 0)) {
3124                                 unlock_page(page);
3125                                 continue;
3126                         }
3127                 }
3128                 clear_page_dirty_for_io(page);
3129                 spin_lock_irq(&page->mapping->tree_lock);
3130                 if (!PageDirty(page)) {
3131                         radix_tree_tag_clear(&page->mapping->page_tree,
3132                                                 page_index(page),
3133                                                 PAGECACHE_TAG_DIRTY);
3134                 }
3135                 spin_unlock_irq(&page->mapping->tree_lock);
3136                 unlock_page(page);
3137         }
3138         return 0;
3139 }
3140 EXPORT_SYMBOL(clear_extent_buffer_dirty);
3141
3142 int wait_on_extent_buffer_writeback(struct extent_io_tree *tree,
3143                                     struct extent_buffer *eb)
3144 {
3145         return wait_on_extent_writeback(tree, eb->start,
3146                                         eb->start + eb->len - 1);
3147 }
3148 EXPORT_SYMBOL(wait_on_extent_buffer_writeback);
3149
3150 int set_extent_buffer_dirty(struct extent_io_tree *tree,
3151                              struct extent_buffer *eb)
3152 {
3153         unsigned long i;
3154         unsigned long num_pages;
3155
3156         num_pages = num_extent_pages(eb->start, eb->len);
3157         for (i = 0; i < num_pages; i++) {
3158                 struct page *page = extent_buffer_page(eb, i);
3159                 /* writepage may need to do something special for the
3160                  * first page, we have to make sure page->private is
3161                  * properly set.  releasepage may drop page->private
3162                  * on us if the page isn't already dirty.
3163                  */
3164                 lock_page(page);
3165                 if (i == 0) {
3166                         set_page_extent_head(page, eb->len);
3167                 } else if (PagePrivate(page) &&
3168                            page->private != EXTENT_PAGE_PRIVATE) {
3169                         set_page_extent_mapped(page);
3170                 }
3171                 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3172                 set_extent_dirty(tree, page_offset(page),
3173                                  page_offset(page) + PAGE_CACHE_SIZE -1,
3174                                  GFP_NOFS);
3175                 unlock_page(page);
3176         }
3177         return 0;
3178 }
3179 EXPORT_SYMBOL(set_extent_buffer_dirty);
3180
3181 int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3182                                 struct extent_buffer *eb)
3183 {
3184         unsigned long i;
3185         struct page *page;
3186         unsigned long num_pages;
3187
3188         num_pages = num_extent_pages(eb->start, eb->len);
3189         eb->flags &= ~EXTENT_UPTODATE;
3190
3191         clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3192                               GFP_NOFS);
3193         for (i = 0; i < num_pages; i++) {
3194                 page = extent_buffer_page(eb, i);
3195                 if (page)
3196                         ClearPageUptodate(page);
3197         }
3198         return 0;
3199 }
3200
3201 int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3202                                 struct extent_buffer *eb)
3203 {
3204         unsigned long i;
3205         struct page *page;
3206         unsigned long num_pages;
3207
3208         num_pages = num_extent_pages(eb->start, eb->len);
3209
3210         set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3211                             GFP_NOFS);
3212         for (i = 0; i < num_pages; i++) {
3213                 page = extent_buffer_page(eb, i);
3214                 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3215                     ((i == num_pages - 1) &&
3216                      ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3217                         check_page_uptodate(tree, page);
3218                         continue;
3219                 }
3220                 SetPageUptodate(page);
3221         }
3222         return 0;
3223 }
3224 EXPORT_SYMBOL(set_extent_buffer_uptodate);
3225
3226 int extent_range_uptodate(struct extent_io_tree *tree,
3227                           u64 start, u64 end)
3228 {
3229         struct page *page;
3230         int ret;
3231         int pg_uptodate = 1;
3232         int uptodate;
3233         unsigned long index;
3234
3235         ret = test_range_bit(tree, start, end, EXTENT_UPTODATE, 1);
3236         if (ret)
3237                 return 1;
3238         while(start <= end) {
3239                 index = start >> PAGE_CACHE_SHIFT;
3240                 page = find_get_page(tree->mapping, index);
3241                 uptodate = PageUptodate(page);
3242                 page_cache_release(page);
3243                 if (!uptodate) {
3244                         pg_uptodate = 0;
3245                         break;
3246                 }
3247                 start += PAGE_CACHE_SIZE;
3248         }
3249         return pg_uptodate;
3250 }
3251
3252 int extent_buffer_uptodate(struct extent_io_tree *tree,
3253                            struct extent_buffer *eb)
3254 {
3255         int ret = 0;
3256         unsigned long num_pages;
3257         unsigned long i;
3258         struct page *page;
3259         int pg_uptodate = 1;
3260
3261         if (eb->flags & EXTENT_UPTODATE)
3262                 return 1;
3263
3264         ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3265                            EXTENT_UPTODATE, 1);
3266         if (ret)
3267                 return ret;
3268
3269         num_pages = num_extent_pages(eb->start, eb->len);
3270         for (i = 0; i < num_pages; i++) {
3271                 page = extent_buffer_page(eb, i);
3272                 if (!PageUptodate(page)) {
3273                         pg_uptodate = 0;
3274                         break;
3275                 }
3276         }
3277         return pg_uptodate;
3278 }
3279 EXPORT_SYMBOL(extent_buffer_uptodate);
3280
3281 int read_extent_buffer_pages(struct extent_io_tree *tree,
3282                              struct extent_buffer *eb,
3283                              u64 start, int wait,
3284                              get_extent_t *get_extent, int mirror_num)
3285 {
3286         unsigned long i;
3287         unsigned long start_i;
3288         struct page *page;
3289         int err;
3290         int ret = 0;
3291         int locked_pages = 0;
3292         int all_uptodate = 1;
3293         int inc_all_pages = 0;
3294         unsigned long num_pages;
3295         struct bio *bio = NULL;
3296         unsigned long bio_flags = 0;
3297
3298         if (eb->flags & EXTENT_UPTODATE)
3299                 return 0;
3300
3301         if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3302                            EXTENT_UPTODATE, 1)) {
3303                 return 0;
3304         }
3305
3306         if (start) {
3307                 WARN_ON(start < eb->start);
3308                 start_i = (start >> PAGE_CACHE_SHIFT) -
3309                         (eb->start >> PAGE_CACHE_SHIFT);
3310         } else {
3311                 start_i = 0;
3312         }
3313
3314         num_pages = num_extent_pages(eb->start, eb->len);
3315         for (i = start_i; i < num_pages; i++) {
3316                 page = extent_buffer_page(eb, i);
3317                 if (!wait) {
3318                         if (!trylock_page(page))
3319                                 goto unlock_exit;
3320                 } else {
3321                         lock_page(page);
3322                 }
3323                 locked_pages++;
3324                 if (!PageUptodate(page)) {
3325                         all_uptodate = 0;
3326                 }
3327         }
3328         if (all_uptodate) {
3329                 if (start_i == 0)
3330                         eb->flags |= EXTENT_UPTODATE;
3331                 if (ret) {
3332                         printk("all up to date but ret is %d\n", ret);
3333                 }
3334                 goto unlock_exit;
3335         }
3336
3337         for (i = start_i; i < num_pages; i++) {
3338                 page = extent_buffer_page(eb, i);
3339                 if (inc_all_pages)
3340                         page_cache_get(page);
3341                 if (!PageUptodate(page)) {
3342                         if (start_i == 0)
3343                                 inc_all_pages = 1;
3344                         ClearPageError(page);
3345                         err = __extent_read_full_page(tree, page,
3346                                                       get_extent, &bio,
3347                                                       mirror_num, &bio_flags);
3348                         if (err) {
3349                                 ret = err;
3350                                 printk("err %d from __extent_read_full_page\n", ret);
3351                         }
3352                 } else {
3353                         unlock_page(page);
3354                 }
3355         }
3356
3357         if (bio)
3358                 submit_one_bio(READ, bio, mirror_num, bio_flags);
3359
3360         if (ret || !wait) {
3361                 if (ret)
3362                         printk("ret %d wait %d returning\n", ret, wait);
3363                 return ret;
3364         }
3365         for (i = start_i; i < num_pages; i++) {
3366                 page = extent_buffer_page(eb, i);
3367                 wait_on_page_locked(page);
3368                 if (!PageUptodate(page)) {
3369                         printk("page not uptodate after wait_on_page_locked\n");
3370                         ret = -EIO;
3371                 }
3372         }
3373         if (!ret)
3374                 eb->flags |= EXTENT_UPTODATE;
3375         return ret;
3376
3377 unlock_exit:
3378         i = start_i;
3379         while(locked_pages > 0) {
3380                 page = extent_buffer_page(eb, i);
3381                 i++;
3382                 unlock_page(page);
3383                 locked_pages--;
3384         }
3385         return ret;
3386 }
3387 EXPORT_SYMBOL(read_extent_buffer_pages);
3388
3389 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3390                         unsigned long start,
3391                         unsigned long len)
3392 {
3393         size_t cur;
3394         size_t offset;
3395         struct page *page;
3396         char *kaddr;
3397         char *dst = (char *)dstv;
3398         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3399         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3400
3401         WARN_ON(start > eb->len);
3402         WARN_ON(start + len > eb->start + eb->len);
3403
3404         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3405
3406         while(len > 0) {
3407                 page = extent_buffer_page(eb, i);
3408
3409                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3410                 kaddr = kmap_atomic(page, KM_USER1);
3411                 memcpy(dst, kaddr + offset, cur);
3412                 kunmap_atomic(kaddr, KM_USER1);
3413
3414                 dst += cur;
3415                 len -= cur;
3416                 offset = 0;
3417                 i++;
3418         }
3419 }
3420 EXPORT_SYMBOL(read_extent_buffer);
3421
3422 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3423                                unsigned long min_len, char **token, char **map,
3424                                unsigned long *map_start,
3425                                unsigned long *map_len, int km)
3426 {
3427         size_t offset = start & (PAGE_CACHE_SIZE - 1);
3428         char *kaddr;
3429         struct page *p;
3430         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3431         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3432         unsigned long end_i = (start_offset + start + min_len - 1) >>
3433                 PAGE_CACHE_SHIFT;
3434
3435         if (i != end_i)
3436                 return -EINVAL;
3437
3438         if (i == 0) {
3439                 offset = start_offset;
3440                 *map_start = 0;
3441         } else {
3442                 offset = 0;
3443                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3444         }
3445         if (start + min_len > eb->len) {
3446 printk("bad mapping eb start %Lu len %lu, wanted %lu %lu\n", eb->start, eb->len, start, min_len);
3447                 WARN_ON(1);
3448         }
3449
3450         p = extent_buffer_page(eb, i);
3451         kaddr = kmap_atomic(p, km);
3452         *token = kaddr;
3453         *map = kaddr + offset;
3454         *map_len = PAGE_CACHE_SIZE - offset;
3455         return 0;
3456 }
3457 EXPORT_SYMBOL(map_private_extent_buffer);
3458
3459 int map_extent_buffer(struct extent_buffer *eb, unsigned long start,
3460                       unsigned long min_len,
3461                       char **token, char **map,
3462                       unsigned long *map_start,
3463                       unsigned long *map_len, int km)
3464 {
3465         int err;
3466         int save = 0;
3467         if (eb->map_token) {
3468                 unmap_extent_buffer(eb, eb->map_token, km);
3469                 eb->map_token = NULL;
3470                 save = 1;
3471                 WARN_ON(!mutex_is_locked(&eb->mutex));
3472         }
3473         err = map_private_extent_buffer(eb, start, min_len, token, map,
3474                                        map_start, map_len, km);
3475         if (!err && save) {
3476                 eb->map_token = *token;
3477                 eb->kaddr = *map;
3478                 eb->map_start = *map_start;
3479                 eb->map_len = *map_len;
3480         }
3481         return err;
3482 }
3483 EXPORT_SYMBOL(map_extent_buffer);
3484
3485 void unmap_extent_buffer(struct extent_buffer *eb, char *token, int km)
3486 {
3487         kunmap_atomic(token, km);
3488 }
3489 EXPORT_SYMBOL(unmap_extent_buffer);
3490
3491 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3492                           unsigned long start,
3493                           unsigned long len)
3494 {
3495         size_t cur;
3496         size_t offset;
3497         struct page *page;
3498         char *kaddr;
3499         char *ptr = (char *)ptrv;
3500         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3501         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3502         int ret = 0;
3503
3504         WARN_ON(start > eb->len);
3505         WARN_ON(start + len > eb->start + eb->len);
3506
3507         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3508
3509         while(len > 0) {
3510                 page = extent_buffer_page(eb, i);
3511
3512                 cur = min(len, (PAGE_CACHE_SIZE - offset));
3513
3514                 kaddr = kmap_atomic(page, KM_USER0);
3515                 ret = memcmp(ptr, kaddr + offset, cur);
3516                 kunmap_atomic(kaddr, KM_USER0);
3517                 if (ret)
3518                         break;
3519
3520                 ptr += cur;
3521                 len -= cur;
3522                 offset = 0;
3523                 i++;
3524         }
3525         return ret;
3526 }
3527 EXPORT_SYMBOL(memcmp_extent_buffer);
3528
3529 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3530                          unsigned long start, unsigned long len)
3531 {
3532         size_t cur;
3533         size_t offset;
3534         struct page *page;
3535         char *kaddr;
3536         char *src = (char *)srcv;
3537         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3538         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3539
3540         WARN_ON(start > eb->len);
3541         WARN_ON(start + len > eb->start + eb->len);
3542
3543         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3544
3545         while(len > 0) {
3546                 page = extent_buffer_page(eb, i);
3547                 WARN_ON(!PageUptodate(page));
3548
3549                 cur = min(len, PAGE_CACHE_SIZE - offset);
3550                 kaddr = kmap_atomic(page, KM_USER1);
3551                 memcpy(kaddr + offset, src, cur);
3552                 kunmap_atomic(kaddr, KM_USER1);
3553
3554                 src += cur;
3555                 len -= cur;
3556                 offset = 0;
3557                 i++;
3558         }
3559 }
3560 EXPORT_SYMBOL(write_extent_buffer);
3561
3562 void memset_extent_buffer(struct extent_buffer *eb, char c,
3563                           unsigned long start, unsigned long len)
3564 {
3565         size_t cur;
3566         size_t offset;
3567         struct page *page;
3568         char *kaddr;
3569         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3570         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3571
3572         WARN_ON(start > eb->len);
3573         WARN_ON(start + len > eb->start + eb->len);
3574
3575         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3576
3577         while(len > 0) {
3578                 page = extent_buffer_page(eb, i);
3579                 WARN_ON(!PageUptodate(page));
3580
3581                 cur = min(len, PAGE_CACHE_SIZE - offset);
3582                 kaddr = kmap_atomic(page, KM_USER0);
3583                 memset(kaddr + offset, c, cur);
3584                 kunmap_atomic(kaddr, KM_USER0);
3585
3586                 len -= cur;
3587                 offset = 0;
3588                 i++;
3589         }
3590 }
3591 EXPORT_SYMBOL(memset_extent_buffer);
3592
3593 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3594                         unsigned long dst_offset, unsigned long src_offset,
3595                         unsigned long len)
3596 {
3597         u64 dst_len = dst->len;
3598         size_t cur;
3599         size_t offset;
3600         struct page *page;
3601         char *kaddr;
3602         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3603         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3604
3605         WARN_ON(src->len != dst_len);
3606
3607         offset = (start_offset + dst_offset) &
3608                 ((unsigned long)PAGE_CACHE_SIZE - 1);
3609
3610         while(len > 0) {
3611                 page = extent_buffer_page(dst, i);
3612                 WARN_ON(!PageUptodate(page));
3613
3614                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3615
3616                 kaddr = kmap_atomic(page, KM_USER0);
3617                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3618                 kunmap_atomic(kaddr, KM_USER0);
3619
3620                 src_offset += cur;
3621                 len -= cur;
3622                 offset = 0;
3623                 i++;
3624         }
3625 }
3626 EXPORT_SYMBOL(copy_extent_buffer);
3627
3628 static void move_pages(struct page *dst_page, struct page *src_page,
3629                        unsigned long dst_off, unsigned long src_off,
3630                        unsigned long len)
3631 {
3632         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3633         if (dst_page == src_page) {
3634                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3635         } else {
3636                 char *src_kaddr = kmap_atomic(src_page, KM_USER1);
3637                 char *p = dst_kaddr + dst_off + len;
3638                 char *s = src_kaddr + src_off + len;
3639
3640                 while (len--)
3641                         *--p = *--s;
3642
3643                 kunmap_atomic(src_kaddr, KM_USER1);
3644         }
3645         kunmap_atomic(dst_kaddr, KM_USER0);
3646 }
3647
3648 static void copy_pages(struct page *dst_page, struct page *src_page,
3649                        unsigned long dst_off, unsigned long src_off,
3650                        unsigned long len)
3651 {
3652         char *dst_kaddr = kmap_atomic(dst_page, KM_USER0);
3653         char *src_kaddr;
3654
3655         if (dst_page != src_page)
3656                 src_kaddr = kmap_atomic(src_page, KM_USER1);
3657         else
3658                 src_kaddr = dst_kaddr;
3659
3660         memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3661         kunmap_atomic(dst_kaddr, KM_USER0);
3662         if (dst_page != src_page)
3663                 kunmap_atomic(src_kaddr, KM_USER1);
3664 }
3665
3666 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3667                            unsigned long src_offset, unsigned long len)
3668 {
3669         size_t cur;
3670         size_t dst_off_in_page;
3671         size_t src_off_in_page;
3672         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3673         unsigned long dst_i;
3674         unsigned long src_i;
3675
3676         if (src_offset + len > dst->len) {
3677                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
3678                        src_offset, len, dst->len);
3679                 BUG_ON(1);
3680         }
3681         if (dst_offset + len > dst->len) {
3682                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
3683                        dst_offset, len, dst->len);
3684                 BUG_ON(1);
3685         }
3686
3687         while(len > 0) {
3688                 dst_off_in_page = (start_offset + dst_offset) &
3689                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3690                 src_off_in_page = (start_offset + src_offset) &
3691                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3692
3693                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3694                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3695
3696                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3697                                                src_off_in_page));
3698                 cur = min_t(unsigned long, cur,
3699                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3700
3701                 copy_pages(extent_buffer_page(dst, dst_i),
3702                            extent_buffer_page(dst, src_i),
3703                            dst_off_in_page, src_off_in_page, cur);
3704
3705                 src_offset += cur;
3706                 dst_offset += cur;
3707                 len -= cur;
3708         }
3709 }
3710 EXPORT_SYMBOL(memcpy_extent_buffer);
3711
3712 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3713                            unsigned long src_offset, unsigned long len)
3714 {
3715         size_t cur;
3716         size_t dst_off_in_page;
3717         size_t src_off_in_page;
3718         unsigned long dst_end = dst_offset + len - 1;
3719         unsigned long src_end = src_offset + len - 1;
3720         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3721         unsigned long dst_i;
3722         unsigned long src_i;
3723
3724         if (src_offset + len > dst->len) {
3725                 printk("memmove bogus src_offset %lu move len %lu len %lu\n",
3726                        src_offset, len, dst->len);
3727                 BUG_ON(1);
3728         }
3729         if (dst_offset + len > dst->len) {
3730                 printk("memmove bogus dst_offset %lu move len %lu len %lu\n",
3731                        dst_offset, len, dst->len);
3732                 BUG_ON(1);
3733         }
3734         if (dst_offset < src_offset) {
3735                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3736                 return;
3737         }
3738         while(len > 0) {
3739                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3740                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3741
3742                 dst_off_in_page = (start_offset + dst_end) &
3743                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3744                 src_off_in_page = (start_offset + src_end) &
3745                         ((unsigned long)PAGE_CACHE_SIZE - 1);
3746
3747                 cur = min_t(unsigned long, len, src_off_in_page + 1);
3748                 cur = min(cur, dst_off_in_page + 1);
3749                 move_pages(extent_buffer_page(dst, dst_i),
3750                            extent_buffer_page(dst, src_i),
3751                            dst_off_in_page - cur + 1,
3752                            src_off_in_page - cur + 1, cur);
3753
3754                 dst_end -= cur;
3755                 src_end -= cur;
3756                 len -= cur;
3757         }
3758 }
3759 EXPORT_SYMBOL(memmove_extent_buffer);
3760
3761 int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3762 {
3763         u64 start = page_offset(page);
3764         struct extent_buffer *eb;
3765         int ret = 1;
3766         unsigned long i;
3767         unsigned long num_pages;
3768
3769         spin_lock(&tree->buffer_lock);
3770         eb = buffer_search(tree, start);
3771         if (!eb)
3772                 goto out;
3773
3774         if (atomic_read(&eb->refs) > 1) {
3775                 ret = 0;
3776                 goto out;
3777         }
3778         /* at this point we can safely release the extent buffer */
3779         num_pages = num_extent_pages(eb->start, eb->len);
3780         for (i = 0; i < num_pages; i++)
3781                 page_cache_release(extent_buffer_page(eb, i));
3782         rb_erase(&eb->rb_node, &tree->buffer);
3783         __free_extent_buffer(eb);
3784 out:
3785         spin_unlock(&tree->buffer_lock);
3786         return ret;
3787 }
3788 EXPORT_SYMBOL(try_release_extent_buffer);