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