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/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/module.h>
   8 #include <linux/spinlock.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/swap.h>
  11 #include <linux/writeback.h>
  12 #include <linux/pagevec.h>
  13 #include <linux/prefetch.h>
  14 #include <linux/cleancache.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 #include "volumes.h"
  21 #include "check-integrity.h"
  22 #include "locking.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26
  27 static LIST_HEAD(buffers);
  28 static LIST_HEAD(states);
  29
  30 #define LEAK_DEBUG 0
  31 #if LEAK_DEBUG
  32 static DEFINE_SPINLOCK(leak_lock);
  33 #endif
  34
  35 #define BUFFER_LRU_MAX 64
  36
  37 struct tree_entry {
  38         u64 start;
  39         u64 end;
  40         struct rb_node rb_node;
  41 };
  42
  43 struct extent_page_data {
  44         struct bio *bio;
  45         struct extent_io_tree *tree;
  46         get_extent_t *get_extent;
  47
  48         /* tells writepage not to lock the state bits for this range
  49          * it still does the unlocking
  50          */
  51         unsigned int extent_locked:1;
  52
  53         /* tells the submit_bio code to use a WRITE_SYNC */
  54         unsigned int sync_io:1;
  55 };
  56
  57 static noinline void flush_write_bio(void *data);
  58
  59 int __init extent_io_init(void)
  60 {
  61         extent_state_cache = kmem_cache_create("extent_state",
  62                         sizeof(struct extent_state), 0,
  63                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  64         if (!extent_state_cache)
  65                 return -ENOMEM;
  66
  67         extent_buffer_cache = kmem_cache_create("extent_buffers",
  68                         sizeof(struct extent_buffer), 0,
  69                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  70         if (!extent_buffer_cache)
  71                 goto free_state_cache;
  72         return 0;
  73
  74 free_state_cache:
  75         kmem_cache_destroy(extent_state_cache);
  76         return -ENOMEM;
  77 }
  78
  79 void extent_io_exit(void)
  80 {
  81         struct extent_state *state;
  82         struct extent_buffer *eb;
  83
  84         while (!list_empty(&states)) {
  85                 state = list_entry(states.next, struct extent_state, leak_list);
  86                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  87                        "state %lu in tree %p refs %d\n",
  88                        (unsigned long long)state->start,
  89                        (unsigned long long)state->end,
  90                        state->state, state->tree, atomic_read(&state->refs));
  91                 list_del(&state->leak_list);
  92                 kmem_cache_free(extent_state_cache, state);
  93
  94         }
  95
  96         while (!list_empty(&buffers)) {
  97                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  98                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
  99                        "refs %d\n", (unsigned long long)eb->start,
 100                        eb->len, atomic_read(&eb->refs));
 101                 list_del(&eb->leak_list);
 102                 kmem_cache_free(extent_buffer_cache, eb);
 103         }
 104         if (extent_state_cache)
 105                 kmem_cache_destroy(extent_state_cache);
 106         if (extent_buffer_cache)
 107                 kmem_cache_destroy(extent_buffer_cache);
 108 }
 109
 110 void extent_io_tree_init(struct extent_io_tree *tree,
 111                          struct address_space *mapping)
 112 {
 113         tree->state = RB_ROOT;
 114         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 115         tree->ops = NULL;
 116         tree->dirty_bytes = 0;
 117         spin_lock_init(&tree->lock);
 118         spin_lock_init(&tree->buffer_lock);
 119         tree->mapping = mapping;
 120 }
 121
 122 static struct extent_state *alloc_extent_state(gfp_t mask)
 123 {
 124         struct extent_state *state;
 125 #if LEAK_DEBUG
 126         unsigned long flags;
 127 #endif
 128
 129         state = kmem_cache_alloc(extent_state_cache, mask);
 130         if (!state)
 131                 return state;
 132         state->state = 0;
 133         state->private = 0;
 134         state->tree = NULL;
 135 #if LEAK_DEBUG
 136         spin_lock_irqsave(&leak_lock, flags);
 137         list_add(&state->leak_list, &states);
 138         spin_unlock_irqrestore(&leak_lock, flags);
 139 #endif
 140         atomic_set(&state->refs, 1);
 141         init_waitqueue_head(&state->wq);
 142         return state;
 143 }
 144
 145 void free_extent_state(struct extent_state *state)
 146 {
 147         if (!state)
 148                 return;
 149         if (atomic_dec_and_test(&state->refs)) {
 150 #if LEAK_DEBUG
 151                 unsigned long flags;
 152 #endif
 153                 WARN_ON(state->tree);
 154 #if LEAK_DEBUG
 155                 spin_lock_irqsave(&leak_lock, flags);
 156                 list_del(&state->leak_list);
 157                 spin_unlock_irqrestore(&leak_lock, flags);
 158 #endif
 159                 kmem_cache_free(extent_state_cache, state);
 160         }
 161 }
 162
 163 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 164                                    struct rb_node *node)
 165 {
 166         struct rb_node **p = &root->rb_node;
 167         struct rb_node *parent = NULL;
 168         struct tree_entry *entry;
 169
 170         while (*p) {
 171                 parent = *p;
 172                 entry = rb_entry(parent, struct tree_entry, rb_node);
 173
 174                 if (offset < entry->start)
 175                         p = &(*p)->rb_left;
 176                 else if (offset > entry->end)
 177                         p = &(*p)->rb_right;
 178                 else
 179                         return parent;
 180         }
 181
 182         entry = rb_entry(node, struct tree_entry, rb_node);
 183         rb_link_node(node, parent, p);
 184         rb_insert_color(node, root);
 185         return NULL;
 186 }
 187
 188 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 189                                      struct rb_node **prev_ret,
 190                                      struct rb_node **next_ret)
 191 {
 192         struct rb_root *root = &tree->state;
 193         struct rb_node *n = root->rb_node;
 194         struct rb_node *prev = NULL;
 195         struct rb_node *orig_prev = NULL;
 196         struct tree_entry *entry;
 197         struct tree_entry *prev_entry = NULL;
 198
 199         while (n) {
 200                 entry = rb_entry(n, struct tree_entry, rb_node);
 201                 prev = n;
 202                 prev_entry = entry;
 203
 204                 if (offset < entry->start)
 205                         n = n->rb_left;
 206                 else if (offset > entry->end)
 207                         n = n->rb_right;
 208                 else
 209                         return n;
 210         }
 211
 212         if (prev_ret) {
 213                 orig_prev = prev;
 214                 while (prev && offset > prev_entry->end) {
 215                         prev = rb_next(prev);
 216                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 217                 }
 218                 *prev_ret = prev;
 219                 prev = orig_prev;
 220         }
 221
 222         if (next_ret) {
 223                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 224                 while (prev && offset < prev_entry->start) {
 225                         prev = rb_prev(prev);
 226                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 227                 }
 228                 *next_ret = prev;
 229         }
 230         return NULL;
 231 }
 232
 233 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 234                                           u64 offset)
 235 {
 236         struct rb_node *prev = NULL;
 237         struct rb_node *ret;
 238
 239         ret = __etree_search(tree, offset, &prev, NULL);
 240         if (!ret)
 241                 return prev;
 242         return ret;
 243 }
 244
 245 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 246                      struct extent_state *other)
 247 {
 248         if (tree->ops && tree->ops->merge_extent_hook)
 249                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 250                                              other);
 251 }
 252
 253 /*
 254  * utility function to look for merge candidates inside a given range.
 255  * Any extents with matching state are merged together into a single
 256  * extent in the tree.  Extents with EXTENT_IO in their state field
 257  * are not merged because the end_io handlers need to be able to do
 258  * operations on them without sleeping (or doing allocations/splits).
 259  *
 260  * This should be called with the tree lock held.
 261  */
 262 static void merge_state(struct extent_io_tree *tree,
 263                         struct extent_state *state)
 264 {
 265         struct extent_state *other;
 266         struct rb_node *other_node;
 267
 268         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 269                 return;
 270
 271         other_node = rb_prev(&state->rb_node);
 272         if (other_node) {
 273                 other = rb_entry(other_node, struct extent_state, rb_node);
 274                 if (other->end == state->start - 1 &&
 275                     other->state == state->state) {
 276                         merge_cb(tree, state, other);
 277                         state->start = other->start;
 278                         other->tree = NULL;
 279                         rb_erase(&other->rb_node, &tree->state);
 280                         free_extent_state(other);
 281                 }
 282         }
 283         other_node = rb_next(&state->rb_node);
 284         if (other_node) {
 285                 other = rb_entry(other_node, struct extent_state, rb_node);
 286                 if (other->start == state->end + 1 &&
 287                     other->state == state->state) {
 288                         merge_cb(tree, state, other);
 289                         state->end = other->end;
 290                         other->tree = NULL;
 291                         rb_erase(&other->rb_node, &tree->state);
 292                         free_extent_state(other);
 293                 }
 294         }
 295 }
 296
 297 static void set_state_cb(struct extent_io_tree *tree,
 298                          struct extent_state *state, int *bits)
 299 {
 300         if (tree->ops && tree->ops->set_bit_hook)
 301                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 302 }
 303
 304 static void clear_state_cb(struct extent_io_tree *tree,
 305                            struct extent_state *state, int *bits)
 306 {
 307         if (tree->ops && tree->ops->clear_bit_hook)
 308                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 309 }
 310
 311 static void set_state_bits(struct extent_io_tree *tree,
 312                            struct extent_state *state, int *bits);
 313
 314 /*
 315  * insert an extent_state struct into the tree.  'bits' are set on the
 316  * struct before it is inserted.
 317  *
 318  * This may return -EEXIST if the extent is already there, in which case the
 319  * state struct is freed.
 320  *
 321  * The tree lock is not taken internally.  This is a utility function and
 322  * probably isn't what you want to call (see set/clear_extent_bit).
 323  */
 324 static int insert_state(struct extent_io_tree *tree,
 325                         struct extent_state *state, u64 start, u64 end,
 326                         int *bits)
 327 {
 328         struct rb_node *node;
 329
 330         if (end < start) {
 331                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
 332                        (unsigned long long)end,
 333                        (unsigned long long)start);
 334                 WARN_ON(1);
 335         }
 336         state->start = start;
 337         state->end = end;
 338
 339         set_state_bits(tree, state, bits);
 340
 341         node = tree_insert(&tree->state, end, &state->rb_node);
 342         if (node) {
 343                 struct extent_state *found;
 344                 found = rb_entry(node, struct extent_state, rb_node);
 345                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 346                        "%llu %llu\n", (unsigned long long)found->start,
 347                        (unsigned long long)found->end,
 348                        (unsigned long long)start, (unsigned long long)end);
 349                 return -EEXIST;
 350         }
 351         state->tree = tree;
 352         merge_state(tree, state);
 353         return 0;
 354 }
 355
 356 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 357                      u64 split)
 358 {
 359         if (tree->ops && tree->ops->split_extent_hook)
 360                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 361 }
 362
 363 /*
 364  * split a given extent state struct in two, inserting the preallocated
 365  * struct 'prealloc' as the newly created second half.  'split' indicates an
 366  * offset inside 'orig' where it should be split.
 367  *
 368  * Before calling,
 369  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 370  * are two extent state structs in the tree:
 371  * prealloc: [orig->start, split - 1]
 372  * orig: [ split, orig->end ]
 373  *
 374  * The tree locks are not taken by this function. They need to be held
 375  * by the caller.
 376  */
 377 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 378                        struct extent_state *prealloc, u64 split)
 379 {
 380         struct rb_node *node;
 381
 382         split_cb(tree, orig, split);
 383
 384         prealloc->start = orig->start;
 385         prealloc->end = split - 1;
 386         prealloc->state = orig->state;
 387         orig->start = split;
 388
 389         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 390         if (node) {
 391                 free_extent_state(prealloc);
 392                 return -EEXIST;
 393         }
 394         prealloc->tree = tree;
 395         return 0;
 396 }
 397
 398 /*
 399  * utility function to clear some bits in an extent state struct.
 400  * it will optionally wake up any one waiting on this state (wake == 1), or
 401  * forcibly remove the state from the tree (delete == 1).
 402  *
 403  * If no bits are set on the state struct after clearing things, the
 404  * struct is freed and removed from the tree
 405  */
 406 static int clear_state_bit(struct extent_io_tree *tree,
 407                             struct extent_state *state,
 408                             int *bits, int wake)
 409 {
 410         int bits_to_clear = *bits & ~EXTENT_CTLBITS;
 411         int ret = state->state & bits_to_clear;
 412
 413         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 414                 u64 range = state->end - state->start + 1;
 415                 WARN_ON(range > tree->dirty_bytes);
 416                 tree->dirty_bytes -= range;
 417         }
 418         clear_state_cb(tree, state, bits);
 419         state->state &= ~bits_to_clear;
 420         if (wake)
 421                 wake_up(&state->wq);
 422         if (state->state == 0) {
 423                 if (state->tree) {
 424                         rb_erase(&state->rb_node, &tree->state);
 425                         state->tree = NULL;
 426                         free_extent_state(state);
 427                 } else {
 428                         WARN_ON(1);
 429                 }
 430         } else {
 431                 merge_state(tree, state);
 432         }
 433         return ret;
 434 }
 435
 436 static struct extent_state *
 437 alloc_extent_state_atomic(struct extent_state *prealloc)
 438 {
 439         if (!prealloc)
 440                 prealloc = alloc_extent_state(GFP_ATOMIC);
 441
 442         return prealloc;
 443 }
 444
 445 /*
 446  * clear some bits on a range in the tree.  This may require splitting
 447  * or inserting elements in the tree, so the gfp mask is used to
 448  * indicate which allocations or sleeping are allowed.
 449  *
 450  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 451  * the given range from the tree regardless of state (ie for truncate).
 452  *
 453  * the range [start, end] is inclusive.
 454  *
 455  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
 456  * bits were already set, or zero if none of the bits were already set.
 457  */
 458 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 459                      int bits, int wake, int delete,
 460                      struct extent_state **cached_state,
 461                      gfp_t mask)
 462 {
 463         struct extent_state *state;
 464         struct extent_state *cached;
 465         struct extent_state *prealloc = NULL;
 466         struct rb_node *next_node;
 467         struct rb_node *node;
 468         u64 last_end;
 469         int err;
 470         int set = 0;
 471         int clear = 0;
 472
 473         if (delete)
 474                 bits |= ~EXTENT_CTLBITS;
 475         bits |= EXTENT_FIRST_DELALLOC;
 476
 477         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 478                 clear = 1;
 479 again:
 480         if (!prealloc && (mask & __GFP_WAIT)) {
 481                 prealloc = alloc_extent_state(mask);
 482                 if (!prealloc)
 483                         return -ENOMEM;
 484         }
 485
 486         spin_lock(&tree->lock);
 487         if (cached_state) {
 488                 cached = *cached_state;
 489
 490                 if (clear) {
 491                         *cached_state = NULL;
 492                         cached_state = NULL;
 493                 }
 494
 495                 if (cached && cached->tree && cached->start <= start &&
 496                     cached->end > start) {
 497                         if (clear)
 498                                 atomic_dec(&cached->refs);
 499                         state = cached;
 500                         goto hit_next;
 501                 }
 502                 if (clear)
 503                         free_extent_state(cached);
 504         }
 505         /*
 506          * this search will find the extents that end after
 507          * our range starts
 508          */
 509         node = tree_search(tree, start);
 510         if (!node)
 511                 goto out;
 512         state = rb_entry(node, struct extent_state, rb_node);
 513 hit_next:
 514         if (state->start > end)
 515                 goto out;
 516         WARN_ON(state->end < start);
 517         last_end = state->end;
 518
 519         if (state->end < end && !need_resched())
 520                 next_node = rb_next(&state->rb_node);
 521         else
 522                 next_node = NULL;
 523
 524         /* the state doesn't have the wanted bits, go ahead */
 525         if (!(state->state & bits))
 526                 goto next;
 527
 528         /*
 529          *     | ---- desired range ---- |
 530          *  | state | or
 531          *  | ------------- state -------------- |
 532          *
 533          * We need to split the extent we found, and may flip
 534          * bits on second half.
 535          *
 536          * If the extent we found extends past our range, we
 537          * just split and search again.  It'll get split again
 538          * the next time though.
 539          *
 540          * If the extent we found is inside our range, we clear
 541          * the desired bit on it.
 542          */
 543
 544         if (state->start < start) {
 545                 prealloc = alloc_extent_state_atomic(prealloc);
 546                 BUG_ON(!prealloc);
 547                 err = split_state(tree, state, prealloc, start);
 548                 BUG_ON(err == -EEXIST);
 549                 prealloc = NULL;
 550                 if (err)
 551                         goto out;
 552                 if (state->end <= end) {
 553                         set |= clear_state_bit(tree, state, &bits, wake);
 554                         if (last_end == (u64)-1)
 555                                 goto out;
 556                         start = last_end + 1;
 557                 }
 558                 goto search_again;
 559         }
 560         /*
 561          * | ---- desired range ---- |
 562          *                        | state |
 563          * We need to split the extent, and clear the bit
 564          * on the first half
 565          */
 566         if (state->start <= end && state->end > end) {
 567                 prealloc = alloc_extent_state_atomic(prealloc);
 568                 BUG_ON(!prealloc);
 569                 err = split_state(tree, state, prealloc, end + 1);
 570                 BUG_ON(err == -EEXIST);
 571                 if (wake)
 572                         wake_up(&state->wq);
 573
 574                 set |= clear_state_bit(tree, prealloc, &bits, wake);
 575
 576                 prealloc = NULL;
 577                 goto out;
 578         }
 579
 580         set |= clear_state_bit(tree, state, &bits, wake);
 581 next:
 582         if (last_end == (u64)-1)
 583                 goto out;
 584         start = last_end + 1;
 585         if (start <= end && next_node) {
 586                 state = rb_entry(next_node, struct extent_state,
 587                                  rb_node);
 588                 goto hit_next;
 589         }
 590         goto search_again;
 591
 592 out:
 593         spin_unlock(&tree->lock);
 594         if (prealloc)
 595                 free_extent_state(prealloc);
 596
 597         return set;
 598
 599 search_again:
 600         if (start > end)
 601                 goto out;
 602         spin_unlock(&tree->lock);
 603         if (mask & __GFP_WAIT)
 604                 cond_resched();
 605         goto again;
 606 }
 607
 608 static int wait_on_state(struct extent_io_tree *tree,
 609                          struct extent_state *state)
 610                 __releases(tree->lock)
 611                 __acquires(tree->lock)
 612 {
 613         DEFINE_WAIT(wait);
 614         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 615         spin_unlock(&tree->lock);
 616         schedule();
 617         spin_lock(&tree->lock);
 618         finish_wait(&state->wq, &wait);
 619         return 0;
 620 }
 621
 622 /*
 623  * waits for one or more bits to clear on a range in the state tree.
 624  * The range [start, end] is inclusive.
 625  * The tree lock is taken by this function
 626  */
 627 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 628 {
 629         struct extent_state *state;
 630         struct rb_node *node;
 631
 632         spin_lock(&tree->lock);
 633 again:
 634         while (1) {
 635                 /*
 636                  * this search will find all the extents that end after
 637                  * our range starts
 638                  */
 639                 node = tree_search(tree, start);
 640                 if (!node)
 641                         break;
 642
 643                 state = rb_entry(node, struct extent_state, rb_node);
 644
 645                 if (state->start > end)
 646                         goto out;
 647
 648                 if (state->state & bits) {
 649                         start = state->start;
 650                         atomic_inc(&state->refs);
 651                         wait_on_state(tree, state);
 652                         free_extent_state(state);
 653                         goto again;
 654                 }
 655                 start = state->end + 1;
 656
 657                 if (start > end)
 658                         break;
 659
 660                 cond_resched_lock(&tree->lock);
 661         }
 662 out:
 663         spin_unlock(&tree->lock);
 664         return 0;
 665 }
 666
 667 static void set_state_bits(struct extent_io_tree *tree,
 668                            struct extent_state *state,
 669                            int *bits)
 670 {
 671         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 672
 673         set_state_cb(tree, state, bits);
 674         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 675                 u64 range = state->end - state->start + 1;
 676                 tree->dirty_bytes += range;
 677         }
 678         state->state |= bits_to_set;
 679 }
 680
 681 static void cache_state(struct extent_state *state,
 682                         struct extent_state **cached_ptr)
 683 {
 684         if (cached_ptr && !(*cached_ptr)) {
 685                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 686                         *cached_ptr = state;
 687                         atomic_inc(&state->refs);
 688                 }
 689         }
 690 }
 691
 692 static void uncache_state(struct extent_state **cached_ptr)
 693 {
 694         if (cached_ptr && (*cached_ptr)) {
 695                 struct extent_state *state = *cached_ptr;
 696                 *cached_ptr = NULL;
 697                 free_extent_state(state);
 698         }
 699 }
 700
 701 /*
 702  * set some bits on a range in the tree.  This may require allocations or
 703  * sleeping, so the gfp mask is used to indicate what is allowed.
 704  *
 705  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 706  * part of the range already has the desired bits set.  The start of the
 707  * existing range is returned in failed_start in this case.
 708  *
 709  * [start, end] is inclusive This takes the tree lock.
 710  */
 711
 712 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 713                    int bits, int exclusive_bits, u64 *failed_start,
 714                    struct extent_state **cached_state, gfp_t mask)
 715 {
 716         struct extent_state *state;
 717         struct extent_state *prealloc = NULL;
 718         struct rb_node *node;
 719         int err = 0;
 720         u64 last_start;
 721         u64 last_end;
 722
 723         bits |= EXTENT_FIRST_DELALLOC;
 724 again:
 725         if (!prealloc && (mask & __GFP_WAIT)) {
 726                 prealloc = alloc_extent_state(mask);
 727                 BUG_ON(!prealloc);
 728         }
 729
 730         spin_lock(&tree->lock);
 731         if (cached_state && *cached_state) {
 732                 state = *cached_state;
 733                 if (state->start <= start && state->end > start &&
 734                     state->tree) {
 735                         node = &state->rb_node;
 736                         goto hit_next;
 737                 }
 738         }
 739         /*
 740          * this search will find all the extents that end after
 741          * our range starts.
 742          */
 743         node = tree_search(tree, start);
 744         if (!node) {
 745                 prealloc = alloc_extent_state_atomic(prealloc);
 746                 BUG_ON(!prealloc);
 747                 err = insert_state(tree, prealloc, start, end, &bits);
 748                 prealloc = NULL;
 749                 BUG_ON(err == -EEXIST);
 750                 goto out;
 751         }
 752         state = rb_entry(node, struct extent_state, rb_node);
 753 hit_next:
 754         last_start = state->start;
 755         last_end = state->end;
 756
 757         /*
 758          * | ---- desired range ---- |
 759          * | state |
 760          *
 761          * Just lock what we found and keep going
 762          */
 763         if (state->start == start && state->end <= end) {
 764                 struct rb_node *next_node;
 765                 if (state->state & exclusive_bits) {
 766                         *failed_start = state->start;
 767                         err = -EEXIST;
 768                         goto out;
 769                 }
 770
 771                 set_state_bits(tree, state, &bits);
 772
 773                 cache_state(state, cached_state);
 774                 merge_state(tree, state);
 775                 if (last_end == (u64)-1)
 776                         goto out;
 777
 778                 start = last_end + 1;
 779                 next_node = rb_next(&state->rb_node);
 780                 if (next_node && start < end && prealloc && !need_resched()) {
 781                         state = rb_entry(next_node, struct extent_state,
 782                                          rb_node);
 783                         if (state->start == start)
 784                                 goto hit_next;
 785                 }
 786                 goto search_again;
 787         }
 788
 789         /*
 790          *     | ---- desired range ---- |
 791          * | state |
 792          *   or
 793          * | ------------- state -------------- |
 794          *
 795          * We need to split the extent we found, and may flip bits on
 796          * second half.
 797          *
 798          * If the extent we found extends past our
 799          * range, we just split and search again.  It'll get split
 800          * again the next time though.
 801          *
 802          * If the extent we found is inside our range, we set the
 803          * desired bit on it.
 804          */
 805         if (state->start < start) {
 806                 if (state->state & exclusive_bits) {
 807                         *failed_start = start;
 808                         err = -EEXIST;
 809                         goto out;
 810                 }
 811
 812                 prealloc = alloc_extent_state_atomic(prealloc);
 813                 BUG_ON(!prealloc);
 814                 err = split_state(tree, state, prealloc, start);
 815                 BUG_ON(err == -EEXIST);
 816                 prealloc = NULL;
 817                 if (err)
 818                         goto out;
 819                 if (state->end <= end) {
 820                         set_state_bits(tree, state, &bits);
 821                         cache_state(state, cached_state);
 822                         merge_state(tree, state);
 823                         if (last_end == (u64)-1)
 824                                 goto out;
 825                         start = last_end + 1;
 826                 }
 827                 goto search_again;
 828         }
 829         /*
 830          * | ---- desired range ---- |
 831          *     | state | or               | state |
 832          *
 833          * There's a hole, we need to insert something in it and
 834          * ignore the extent we found.
 835          */
 836         if (state->start > start) {
 837                 u64 this_end;
 838                 if (end < last_start)
 839                         this_end = end;
 840                 else
 841                         this_end = last_start - 1;
 842
 843                 prealloc = alloc_extent_state_atomic(prealloc);
 844                 BUG_ON(!prealloc);
 845
 846                 /*
 847                  * Avoid to free 'prealloc' if it can be merged with
 848                  * the later extent.
 849                  */
 850                 err = insert_state(tree, prealloc, start, this_end,
 851                                    &bits);
 852                 BUG_ON(err == -EEXIST);
 853                 if (err) {
 854                         free_extent_state(prealloc);
 855                         prealloc = NULL;
 856                         goto out;
 857                 }
 858                 cache_state(prealloc, cached_state);
 859                 prealloc = NULL;
 860                 start = this_end + 1;
 861                 goto search_again;
 862         }
 863         /*
 864          * | ---- desired range ---- |
 865          *                        | state |
 866          * We need to split the extent, and set the bit
 867          * on the first half
 868          */
 869         if (state->start <= end && state->end > end) {
 870                 if (state->state & exclusive_bits) {
 871                         *failed_start = start;
 872                         err = -EEXIST;
 873                         goto out;
 874                 }
 875
 876                 prealloc = alloc_extent_state_atomic(prealloc);
 877                 BUG_ON(!prealloc);
 878                 err = split_state(tree, state, prealloc, end + 1);
 879                 BUG_ON(err == -EEXIST);
 880
 881                 set_state_bits(tree, prealloc, &bits);
 882                 cache_state(prealloc, cached_state);
 883                 merge_state(tree, prealloc);
 884                 prealloc = NULL;
 885                 goto out;
 886         }
 887
 888         goto search_again;
 889
 890 out:
 891         spin_unlock(&tree->lock);
 892         if (prealloc)
 893                 free_extent_state(prealloc);
 894
 895         return err;
 896
 897 search_again:
 898         if (start > end)
 899                 goto out;
 900         spin_unlock(&tree->lock);
 901         if (mask & __GFP_WAIT)
 902                 cond_resched();
 903         goto again;
 904 }
 905
 906 /**
 907  * convert_extent - convert all bits in a given range from one bit to another
 908  * @tree:       the io tree to search
 909  * @start:      the start offset in bytes
 910  * @end:        the end offset in bytes (inclusive)
 911  * @bits:       the bits to set in this range
 912  * @clear_bits: the bits to clear in this range
 913  * @mask:       the allocation mask
 914  *
 915  * This will go through and set bits for the given range.  If any states exist
 916  * already in this range they are set with the given bit and cleared of the
 917  * clear_bits.  This is only meant to be used by things that are mergeable, ie
 918  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
 919  * boundary bits like LOCK.
 920  */
 921 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 922                        int bits, int clear_bits, gfp_t mask)
 923 {
 924         struct extent_state *state;
 925         struct extent_state *prealloc = NULL;
 926         struct rb_node *node;
 927         int err = 0;
 928         u64 last_start;
 929         u64 last_end;
 930
 931 again:
 932         if (!prealloc && (mask & __GFP_WAIT)) {
 933                 prealloc = alloc_extent_state(mask);
 934                 if (!prealloc)
 935                         return -ENOMEM;
 936         }
 937
 938         spin_lock(&tree->lock);
 939         /*
 940          * this search will find all the extents that end after
 941          * our range starts.
 942          */
 943         node = tree_search(tree, start);
 944         if (!node) {
 945                 prealloc = alloc_extent_state_atomic(prealloc);
 946                 if (!prealloc) {
 947                         err = -ENOMEM;
 948                         goto out;
 949                 }
 950                 err = insert_state(tree, prealloc, start, end, &bits);
 951                 prealloc = NULL;
 952                 BUG_ON(err == -EEXIST);
 953                 goto out;
 954         }
 955         state = rb_entry(node, struct extent_state, rb_node);
 956 hit_next:
 957         last_start = state->start;
 958         last_end = state->end;
 959
 960         /*
 961          * | ---- desired range ---- |
 962          * | state |
 963          *
 964          * Just lock what we found and keep going
 965          */
 966         if (state->start == start && state->end <= end) {
 967                 struct rb_node *next_node;
 968
 969                 set_state_bits(tree, state, &bits);
 970                 clear_state_bit(tree, state, &clear_bits, 0);
 971                 if (last_end == (u64)-1)
 972                         goto out;
 973
 974                 start = last_end + 1;
 975                 next_node = rb_next(&state->rb_node);
 976                 if (next_node && start < end && prealloc && !need_resched()) {
 977                         state = rb_entry(next_node, struct extent_state,
 978                                          rb_node);
 979                         if (state->start == start)
 980                                 goto hit_next;
 981                 }
 982                 goto search_again;
 983         }
 984
 985         /*
 986          *     | ---- desired range ---- |
 987          * | state |
 988          *   or
 989          * | ------------- state -------------- |
 990          *
 991          * We need to split the extent we found, and may flip bits on
 992          * second half.
 993          *
 994          * If the extent we found extends past our
 995          * range, we just split and search again.  It'll get split
 996          * again the next time though.
 997          *
 998          * If the extent we found is inside our range, we set the
 999          * desired bit on it.
1000          */
1001         if (state->start < start) {
1002                 prealloc = alloc_extent_state_atomic(prealloc);
1003                 if (!prealloc) {
1004                         err = -ENOMEM;
1005                         goto out;
1006                 }
1007                 err = split_state(tree, state, prealloc, start);
1008                 BUG_ON(err == -EEXIST);
1009                 prealloc = NULL;
1010                 if (err)
1011                         goto out;
1012                 if (state->end <= end) {
1013                         set_state_bits(tree, state, &bits);
1014                         clear_state_bit(tree, state, &clear_bits, 0);
1015                         if (last_end == (u64)-1)
1016                                 goto out;
1017                         start = last_end + 1;
1018                 }
1019                 goto search_again;
1020         }
1021         /*
1022          * | ---- desired range ---- |
1023          *     | state | or               | state |
1024          *
1025          * There's a hole, we need to insert something in it and
1026          * ignore the extent we found.
1027          */
1028         if (state->start > start) {
1029                 u64 this_end;
1030                 if (end < last_start)
1031                         this_end = end;
1032                 else
1033                         this_end = last_start - 1;
1034
1035                 prealloc = alloc_extent_state_atomic(prealloc);
1036                 if (!prealloc) {
1037                         err = -ENOMEM;
1038                         goto out;
1039                 }
1040
1041                 /*
1042                  * Avoid to free 'prealloc' if it can be merged with
1043                  * the later extent.
1044                  */
1045                 err = insert_state(tree, prealloc, start, this_end,
1046                                    &bits);
1047                 BUG_ON(err == -EEXIST);
1048                 if (err) {
1049                         free_extent_state(prealloc);
1050                         prealloc = NULL;
1051                         goto out;
1052                 }
1053                 prealloc = NULL;
1054                 start = this_end + 1;
1055                 goto search_again;
1056         }
1057         /*
1058          * | ---- desired range ---- |
1059          *                        | state |
1060          * We need to split the extent, and set the bit
1061          * on the first half
1062          */
1063         if (state->start <= end && state->end > end) {
1064                 prealloc = alloc_extent_state_atomic(prealloc);
1065                 if (!prealloc) {
1066                         err = -ENOMEM;
1067                         goto out;
1068                 }
1069
1070                 err = split_state(tree, state, prealloc, end + 1);
1071                 BUG_ON(err == -EEXIST);
1072
1073                 set_state_bits(tree, prealloc, &bits);
1074                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1075                 prealloc = NULL;
1076                 goto out;
1077         }
1078
1079         goto search_again;
1080
1081 out:
1082         spin_unlock(&tree->lock);
1083         if (prealloc)
1084                 free_extent_state(prealloc);
1085
1086         return err;
1087
1088 search_again:
1089         if (start > end)
1090                 goto out;
1091         spin_unlock(&tree->lock);
1092         if (mask & __GFP_WAIT)
1093                 cond_resched();
1094         goto again;
1095 }
1096
1097 /* wrappers around set/clear extent bit */
1098 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1099                      gfp_t mask)
1100 {
1101         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
1102                               NULL, mask);
1103 }
1104
1105 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1106                     int bits, gfp_t mask)
1107 {
1108         return set_extent_bit(tree, start, end, bits, 0, NULL,
1109                               NULL, mask);
1110 }
1111
1112 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1113                       int bits, gfp_t mask)
1114 {
1115         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1116 }
1117
1118 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1119                         struct extent_state **cached_state, gfp_t mask)
1120 {
1121         return set_extent_bit(tree, start, end,
1122                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1123                               0, NULL, cached_state, mask);
1124 }
1125
1126 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1127                        gfp_t mask)
1128 {
1129         return clear_extent_bit(tree, start, end,
1130                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1131                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1132 }
1133
1134 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1135                      gfp_t mask)
1136 {
1137         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
1138                               NULL, mask);
1139 }
1140
1141 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1142                         struct extent_state **cached_state, gfp_t mask)
1143 {
1144         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
1145                               NULL, cached_state, mask);
1146 }
1147
1148 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
1149                                  u64 end, struct extent_state **cached_state,
1150                                  gfp_t mask)
1151 {
1152         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1153                                 cached_state, mask);
1154 }
1155
1156 /*
1157  * either insert or lock state struct between start and end use mask to tell
1158  * us if waiting is desired.
1159  */
1160 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1161                      int bits, struct extent_state **cached_state, gfp_t mask)
1162 {
1163         int err;
1164         u64 failed_start;
1165         while (1) {
1166                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1167                                      EXTENT_LOCKED, &failed_start,
1168                                      cached_state, mask);
1169                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
1170                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1171                         start = failed_start;
1172                 } else {
1173                         break;
1174                 }
1175                 WARN_ON(start > end);
1176         }
1177         return err;
1178 }
1179
1180 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1181 {
1182         return lock_extent_bits(tree, start, end, 0, NULL, mask);
1183 }
1184
1185 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1186                     gfp_t mask)
1187 {
1188         int err;
1189         u64 failed_start;
1190
1191         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1192                              &failed_start, NULL, mask);
1193         if (err == -EEXIST) {
1194                 if (failed_start > start)
1195                         clear_extent_bit(tree, start, failed_start - 1,
1196                                          EXTENT_LOCKED, 1, 0, NULL, mask);
1197                 return 0;
1198         }
1199         return 1;
1200 }
1201
1202 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1203                          struct extent_state **cached, gfp_t mask)
1204 {
1205         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1206                                 mask);
1207 }
1208
1209 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1210 {
1211         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1212                                 mask);
1213 }
1214
1215 /*
1216  * helper function to set both pages and extents in the tree writeback
1217  */
1218 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1219 {
1220         unsigned long index = start >> PAGE_CACHE_SHIFT;
1221         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1222         struct page *page;
1223
1224         while (index <= end_index) {
1225                 page = find_get_page(tree->mapping, index);
1226                 BUG_ON(!page);
1227                 set_page_writeback(page);
1228                 page_cache_release(page);
1229                 index++;
1230         }
1231         return 0;
1232 }
1233
1234 /* find the first state struct with 'bits' set after 'start', and
1235  * return it.  tree->lock must be held.  NULL will returned if
1236  * nothing was found after 'start'
1237  */
1238 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1239                                                  u64 start, int bits)
1240 {
1241         struct rb_node *node;
1242         struct extent_state *state;
1243
1244         /*
1245          * this search will find all the extents that end after
1246          * our range starts.
1247          */
1248         node = tree_search(tree, start);
1249         if (!node)
1250                 goto out;
1251
1252         while (1) {
1253                 state = rb_entry(node, struct extent_state, rb_node);
1254                 if (state->end >= start && (state->state & bits))
1255                         return state;
1256
1257                 node = rb_next(node);
1258                 if (!node)
1259                         break;
1260         }
1261 out:
1262         return NULL;
1263 }
1264
1265 /*
1266  * find the first offset in the io tree with 'bits' set. zero is
1267  * returned if we find something, and *start_ret and *end_ret are
1268  * set to reflect the state struct that was found.
1269  *
1270  * If nothing was found, 1 is returned, < 0 on error
1271  */
1272 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1273                           u64 *start_ret, u64 *end_ret, int bits)
1274 {
1275         struct extent_state *state;
1276         int ret = 1;
1277
1278         spin_lock(&tree->lock);
1279         state = find_first_extent_bit_state(tree, start, bits);
1280         if (state) {
1281                 *start_ret = state->start;
1282                 *end_ret = state->end;
1283                 ret = 0;
1284         }
1285         spin_unlock(&tree->lock);
1286         return ret;
1287 }
1288
1289 /*
1290  * find a contiguous range of bytes in the file marked as delalloc, not
1291  * more than 'max_bytes'.  start and end are used to return the range,
1292  *
1293  * 1 is returned if we find something, 0 if nothing was in the tree
1294  */
1295 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1296                                         u64 *start, u64 *end, u64 max_bytes,
1297                                         struct extent_state **cached_state)
1298 {
1299         struct rb_node *node;
1300         struct extent_state *state;
1301         u64 cur_start = *start;
1302         u64 found = 0;
1303         u64 total_bytes = 0;
1304
1305         spin_lock(&tree->lock);
1306
1307         /*
1308          * this search will find all the extents that end after
1309          * our range starts.
1310          */
1311         node = tree_search(tree, cur_start);
1312         if (!node) {
1313                 if (!found)
1314                         *end = (u64)-1;
1315                 goto out;
1316         }
1317
1318         while (1) {
1319                 state = rb_entry(node, struct extent_state, rb_node);
1320                 if (found && (state->start != cur_start ||
1321                               (state->state & EXTENT_BOUNDARY))) {
1322                         goto out;
1323                 }
1324                 if (!(state->state & EXTENT_DELALLOC)) {
1325                         if (!found)
1326                                 *end = state->end;
1327                         goto out;
1328                 }
1329                 if (!found) {
1330                         *start = state->start;
1331                         *cached_state = state;
1332                         atomic_inc(&state->refs);
1333                 }
1334                 found++;
1335                 *end = state->end;
1336                 cur_start = state->end + 1;
1337                 node = rb_next(node);
1338                 if (!node)
1339                         break;
1340                 total_bytes += state->end - state->start + 1;
1341                 if (total_bytes >= max_bytes)
1342                         break;
1343         }
1344 out:
1345         spin_unlock(&tree->lock);
1346         return found;
1347 }
1348
1349 static noinline int __unlock_for_delalloc(struct inode *inode,
1350                                           struct page *locked_page,
1351                                           u64 start, u64 end)
1352 {
1353         int ret;
1354         struct page *pages[16];
1355         unsigned long index = start >> PAGE_CACHE_SHIFT;
1356         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1357         unsigned long nr_pages = end_index - index + 1;
1358         int i;
1359
1360         if (index == locked_page->index && end_index == index)
1361                 return 0;
1362
1363         while (nr_pages > 0) {
1364                 ret = find_get_pages_contig(inode->i_mapping, index,
1365                                      min_t(unsigned long, nr_pages,
1366                                      ARRAY_SIZE(pages)), pages);
1367                 for (i = 0; i < ret; i++) {
1368                         if (pages[i] != locked_page)
1369                                 unlock_page(pages[i]);
1370                         page_cache_release(pages[i]);
1371                 }
1372                 nr_pages -= ret;
1373                 index += ret;
1374                 cond_resched();
1375         }
1376         return 0;
1377 }
1378
1379 static noinline int lock_delalloc_pages(struct inode *inode,
1380                                         struct page *locked_page,
1381                                         u64 delalloc_start,
1382                                         u64 delalloc_end)
1383 {
1384         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1385         unsigned long start_index = index;
1386         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1387         unsigned long pages_locked = 0;
1388         struct page *pages[16];
1389         unsigned long nrpages;
1390         int ret;
1391         int i;
1392
1393         /* the caller is responsible for locking the start index */
1394         if (index == locked_page->index && index == end_index)
1395                 return 0;
1396
1397         /* skip the page at the start index */
1398         nrpages = end_index - index + 1;
1399         while (nrpages > 0) {
1400                 ret = find_get_pages_contig(inode->i_mapping, index,
1401                                      min_t(unsigned long,
1402                                      nrpages, ARRAY_SIZE(pages)), pages);
1403                 if (ret == 0) {
1404                         ret = -EAGAIN;
1405                         goto done;
1406                 }
1407                 /* now we have an array of pages, lock them all */
1408                 for (i = 0; i < ret; i++) {
1409                         /*
1410                          * the caller is taking responsibility for
1411                          * locked_page
1412                          */
1413                         if (pages[i] != locked_page) {
1414                                 lock_page(pages[i]);
1415                                 if (!PageDirty(pages[i]) ||
1416                                     pages[i]->mapping != inode->i_mapping) {
1417                                         ret = -EAGAIN;
1418                                         unlock_page(pages[i]);
1419                                         page_cache_release(pages[i]);
1420                                         goto done;
1421                                 }
1422                         }
1423                         page_cache_release(pages[i]);
1424                         pages_locked++;
1425                 }
1426                 nrpages -= ret;
1427                 index += ret;
1428                 cond_resched();
1429         }
1430         ret = 0;
1431 done:
1432         if (ret && pages_locked) {
1433                 __unlock_for_delalloc(inode, locked_page,
1434                               delalloc_start,
1435                               ((u64)(start_index + pages_locked - 1)) <<
1436                               PAGE_CACHE_SHIFT);
1437         }
1438         return ret;
1439 }
1440
1441 /*
1442  * find a contiguous range of bytes in the file marked as delalloc, not
1443  * more than 'max_bytes'.  start and end are used to return the range,
1444  *
1445  * 1 is returned if we find something, 0 if nothing was in the tree
1446  */
1447 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1448                                              struct extent_io_tree *tree,
1449                                              struct page *locked_page,
1450                                              u64 *start, u64 *end,
1451                                              u64 max_bytes)
1452 {
1453         u64 delalloc_start;
1454         u64 delalloc_end;
1455         u64 found;
1456         struct extent_state *cached_state = NULL;
1457         int ret;
1458         int loops = 0;
1459
1460 again:
1461         /* step one, find a bunch of delalloc bytes starting at start */
1462         delalloc_start = *start;
1463         delalloc_end = 0;
1464         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1465                                     max_bytes, &cached_state);
1466         if (!found || delalloc_end <= *start) {
1467                 *start = delalloc_start;
1468                 *end = delalloc_end;
1469                 free_extent_state(cached_state);
1470                 return found;
1471         }
1472
1473         /*
1474          * start comes from the offset of locked_page.  We have to lock
1475          * pages in order, so we can't process delalloc bytes before
1476          * locked_page
1477          */
1478         if (delalloc_start < *start)
1479                 delalloc_start = *start;
1480
1481         /*
1482          * make sure to limit the number of pages we try to lock down
1483          * if we're looping.
1484          */
1485         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1486                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1487
1488         /* step two, lock all the pages after the page that has start */
1489         ret = lock_delalloc_pages(inode, locked_page,
1490                                   delalloc_start, delalloc_end);
1491         if (ret == -EAGAIN) {
1492                 /* some of the pages are gone, lets avoid looping by
1493                  * shortening the size of the delalloc range we're searching
1494                  */
1495                 free_extent_state(cached_state);
1496                 if (!loops) {
1497                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1498                         max_bytes = PAGE_CACHE_SIZE - offset;
1499                         loops = 1;
1500                         goto again;
1501                 } else {
1502                         found = 0;
1503                         goto out_failed;
1504                 }
1505         }
1506         BUG_ON(ret);
1507
1508         /* step three, lock the state bits for the whole range */
1509         lock_extent_bits(tree, delalloc_start, delalloc_end,
1510                          0, &cached_state, GFP_NOFS);
1511
1512         /* then test to make sure it is all still delalloc */
1513         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1514                              EXTENT_DELALLOC, 1, cached_state);
1515         if (!ret) {
1516                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1517                                      &cached_state, GFP_NOFS);
1518                 __unlock_for_delalloc(inode, locked_page,
1519                               delalloc_start, delalloc_end);
1520                 cond_resched();
1521                 goto again;
1522         }
1523         free_extent_state(cached_state);
1524         *start = delalloc_start;
1525         *end = delalloc_end;
1526 out_failed:
1527         return found;
1528 }
1529
1530 int extent_clear_unlock_delalloc(struct inode *inode,
1531                                 struct extent_io_tree *tree,
1532                                 u64 start, u64 end, struct page *locked_page,
1533                                 unsigned long op)
1534 {
1535         int ret;
1536         struct page *pages[16];
1537         unsigned long index = start >> PAGE_CACHE_SHIFT;
1538         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1539         unsigned long nr_pages = end_index - index + 1;
1540         int i;
1541         int clear_bits = 0;
1542
1543         if (op & EXTENT_CLEAR_UNLOCK)
1544                 clear_bits |= EXTENT_LOCKED;
1545         if (op & EXTENT_CLEAR_DIRTY)
1546                 clear_bits |= EXTENT_DIRTY;
1547
1548         if (op & EXTENT_CLEAR_DELALLOC)
1549                 clear_bits |= EXTENT_DELALLOC;
1550
1551         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1552         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1553                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1554                     EXTENT_SET_PRIVATE2)))
1555                 return 0;
1556
1557         while (nr_pages > 0) {
1558                 ret = find_get_pages_contig(inode->i_mapping, index,
1559                                      min_t(unsigned long,
1560                                      nr_pages, ARRAY_SIZE(pages)), pages);
1561                 for (i = 0; i < ret; i++) {
1562
1563                         if (op & EXTENT_SET_PRIVATE2)
1564                                 SetPagePrivate2(pages[i]);
1565
1566                         if (pages[i] == locked_page) {
1567                                 page_cache_release(pages[i]);
1568                                 continue;
1569                         }
1570                         if (op & EXTENT_CLEAR_DIRTY)
1571                                 clear_page_dirty_for_io(pages[i]);
1572                         if (op & EXTENT_SET_WRITEBACK)
1573                                 set_page_writeback(pages[i]);
1574                         if (op & EXTENT_END_WRITEBACK)
1575                                 end_page_writeback(pages[i]);
1576                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1577                                 unlock_page(pages[i]);
1578                         page_cache_release(pages[i]);
1579                 }
1580                 nr_pages -= ret;
1581                 index += ret;
1582                 cond_resched();
1583         }
1584         return 0;
1585 }
1586
1587 /*
1588  * count the number of bytes in the tree that have a given bit(s)
1589  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1590  * cached.  The total number found is returned.
1591  */
1592 u64 count_range_bits(struct extent_io_tree *tree,
1593                      u64 *start, u64 search_end, u64 max_bytes,
1594                      unsigned long bits, int contig)
1595 {
1596         struct rb_node *node;
1597         struct extent_state *state;
1598         u64 cur_start = *start;
1599         u64 total_bytes = 0;
1600         u64 last = 0;
1601         int found = 0;
1602
1603         if (search_end <= cur_start) {
1604                 WARN_ON(1);
1605                 return 0;
1606         }
1607
1608         spin_lock(&tree->lock);
1609         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1610                 total_bytes = tree->dirty_bytes;
1611                 goto out;
1612         }
1613         /*
1614          * this search will find all the extents that end after
1615          * our range starts.
1616          */
1617         node = tree_search(tree, cur_start);
1618         if (!node)
1619                 goto out;
1620
1621         while (1) {
1622                 state = rb_entry(node, struct extent_state, rb_node);
1623                 if (state->start > search_end)
1624                         break;
1625                 if (contig && found && state->start > last + 1)
1626                         break;
1627                 if (state->end >= cur_start && (state->state & bits) == bits) {
1628                         total_bytes += min(search_end, state->end) + 1 -
1629                                        max(cur_start, state->start);
1630                         if (total_bytes >= max_bytes)
1631                                 break;
1632                         if (!found) {
1633                                 *start = max(cur_start, state->start);
1634                                 found = 1;
1635                         }
1636                         last = state->end;
1637                 } else if (contig && found) {
1638                         break;
1639                 }
1640                 node = rb_next(node);
1641                 if (!node)
1642                         break;
1643         }
1644 out:
1645         spin_unlock(&tree->lock);
1646         return total_bytes;
1647 }
1648
1649 /*
1650  * set the private field for a given byte offset in the tree.  If there isn't
1651  * an extent_state there already, this does nothing.
1652  */
1653 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1654 {
1655         struct rb_node *node;
1656         struct extent_state *state;
1657         int ret = 0;
1658
1659         spin_lock(&tree->lock);
1660         /*
1661          * this search will find all the extents that end after
1662          * our range starts.
1663          */
1664         node = tree_search(tree, start);
1665         if (!node) {
1666                 ret = -ENOENT;
1667                 goto out;
1668         }
1669         state = rb_entry(node, struct extent_state, rb_node);
1670         if (state->start != start) {
1671                 ret = -ENOENT;
1672                 goto out;
1673         }
1674         state->private = private;
1675 out:
1676         spin_unlock(&tree->lock);
1677         return ret;
1678 }
1679
1680 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1681 {
1682         struct rb_node *node;
1683         struct extent_state *state;
1684         int ret = 0;
1685
1686         spin_lock(&tree->lock);
1687         /*
1688          * this search will find all the extents that end after
1689          * our range starts.
1690          */
1691         node = tree_search(tree, start);
1692         if (!node) {
1693                 ret = -ENOENT;
1694                 goto out;
1695         }
1696         state = rb_entry(node, struct extent_state, rb_node);
1697         if (state->start != start) {
1698                 ret = -ENOENT;
1699                 goto out;
1700         }
1701         *private = state->private;
1702 out:
1703         spin_unlock(&tree->lock);
1704         return ret;
1705 }
1706
1707 /*
1708  * searches a range in the state tree for a given mask.
1709  * If 'filled' == 1, this returns 1 only if every extent in the tree
1710  * has the bits set.  Otherwise, 1 is returned if any bit in the
1711  * range is found set.
1712  */
1713 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1714                    int bits, int filled, struct extent_state *cached)
1715 {
1716         struct extent_state *state = NULL;
1717         struct rb_node *node;
1718         int bitset = 0;
1719
1720         spin_lock(&tree->lock);
1721         if (cached && cached->tree && cached->start <= start &&
1722             cached->end > start)
1723                 node = &cached->rb_node;
1724         else
1725                 node = tree_search(tree, start);
1726         while (node && start <= end) {
1727                 state = rb_entry(node, struct extent_state, rb_node);
1728
1729                 if (filled && state->start > start) {
1730                         bitset = 0;
1731                         break;
1732                 }
1733
1734                 if (state->start > end)
1735                         break;
1736
1737                 if (state->state & bits) {
1738                         bitset = 1;
1739                         if (!filled)
1740                                 break;
1741                 } else if (filled) {
1742                         bitset = 0;
1743                         break;
1744                 }
1745
1746                 if (state->end == (u64)-1)
1747                         break;
1748
1749                 start = state->end + 1;
1750                 if (start > end)
1751                         break;
1752                 node = rb_next(node);
1753                 if (!node) {
1754                         if (filled)
1755                                 bitset = 0;
1756                         break;
1757                 }
1758         }
1759         spin_unlock(&tree->lock);
1760         return bitset;
1761 }
1762
1763 /*
1764  * helper function to set a given page up to date if all the
1765  * extents in the tree for that page are up to date
1766  */
1767 static int check_page_uptodate(struct extent_io_tree *tree,
1768                                struct page *page)
1769 {
1770         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1771         u64 end = start + PAGE_CACHE_SIZE - 1;
1772         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1773                 SetPageUptodate(page);
1774         return 0;
1775 }
1776
1777 /*
1778  * helper function to unlock a page if all the extents in the tree
1779  * for that page are unlocked
1780  */
1781 static int check_page_locked(struct extent_io_tree *tree,
1782                              struct page *page)
1783 {
1784         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1785         u64 end = start + PAGE_CACHE_SIZE - 1;
1786         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1787                 unlock_page(page);
1788         return 0;
1789 }
1790
1791 /*
1792  * helper function to end page writeback if all the extents
1793  * in the tree for that page are done with writeback
1794  */
1795 static int check_page_writeback(struct extent_io_tree *tree,
1796                              struct page *page)
1797 {
1798         end_page_writeback(page);
1799         return 0;
1800 }
1801
1802 /*
1803  * When IO fails, either with EIO or csum verification fails, we
1804  * try other mirrors that might have a good copy of the data.  This
1805  * io_failure_record is used to record state as we go through all the
1806  * mirrors.  If another mirror has good data, the page is set up to date
1807  * and things continue.  If a good mirror can't be found, the original
1808  * bio end_io callback is called to indicate things have failed.
1809  */
1810 struct io_failure_record {
1811         struct page *page;
1812         u64 start;
1813         u64 len;
1814         u64 logical;
1815         unsigned long bio_flags;
1816         int this_mirror;
1817         int failed_mirror;
1818         int in_validation;
1819 };
1820
1821 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1822                                 int did_repair)
1823 {
1824         int ret;
1825         int err = 0;
1826         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1827
1828         set_state_private(failure_tree, rec->start, 0);
1829         ret = clear_extent_bits(failure_tree, rec->start,
1830                                 rec->start + rec->len - 1,
1831                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1832         if (ret)
1833                 err = ret;
1834
1835         if (did_repair) {
1836                 ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1837                                         rec->start + rec->len - 1,
1838                                         EXTENT_DAMAGED, GFP_NOFS);
1839                 if (ret && !err)
1840                         err = ret;
1841         }
1842
1843         kfree(rec);
1844         return err;
1845 }
1846
1847 static void repair_io_failure_callback(struct bio *bio, int err)
1848 {
1849         complete(bio->bi_private);
1850 }
1851
1852 /*
1853  * this bypasses the standard btrfs submit functions deliberately, as
1854  * the standard behavior is to write all copies in a raid setup. here we only
1855  * want to write the one bad copy. so we do the mapping for ourselves and issue
1856  * submit_bio directly.
1857  * to avoid any synchonization issues, wait for the data after writing, which
1858  * actually prevents the read that triggered the error from finishing.
1859  * currently, there can be no more than two copies of every data bit. thus,
1860  * exactly one rewrite is required.
1861  */
1862 int repair_io_failure(struct btrfs_mapping_tree *map_tree, u64 start,
1863                         u64 length, u64 logical, struct page *page,
1864                         int mirror_num)
1865 {
1866         struct bio *bio;
1867         struct btrfs_device *dev;
1868         DECLARE_COMPLETION_ONSTACK(compl);
1869         u64 map_length = 0;
1870         u64 sector;
1871         struct btrfs_bio *bbio = NULL;
1872         int ret;
1873
1874         BUG_ON(!mirror_num);
1875
1876         bio = bio_alloc(GFP_NOFS, 1);
1877         if (!bio)
1878                 return -EIO;
1879         bio->bi_private = &compl;
1880         bio->bi_end_io = repair_io_failure_callback;
1881         bio->bi_size = 0;
1882         map_length = length;
1883
1884         ret = btrfs_map_block(map_tree, WRITE, logical,
1885                               &map_length, &bbio, mirror_num);
1886         if (ret) {
1887                 bio_put(bio);
1888                 return -EIO;
1889         }
1890         BUG_ON(mirror_num != bbio->mirror_num);
1891         sector = bbio->stripes[mirror_num-1].physical >> 9;
1892         bio->bi_sector = sector;
1893         dev = bbio->stripes[mirror_num-1].dev;
1894         kfree(bbio);
1895         if (!dev || !dev->bdev || !dev->writeable) {
1896                 bio_put(bio);
1897                 return -EIO;
1898         }
1899         bio->bi_bdev = dev->bdev;
1900         bio_add_page(bio, page, length, start-page_offset(page));
1901         btrfsic_submit_bio(WRITE_SYNC, bio);
1902         wait_for_completion(&compl);
1903
1904         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
1905                 /* try to remap that extent elsewhere? */
1906                 bio_put(bio);
1907                 return -EIO;
1908         }
1909
1910         printk(KERN_INFO "btrfs read error corrected: ino %lu off %llu (dev %s "
1911                         "sector %llu)\n", page->mapping->host->i_ino, start,
1912                         dev->name, sector);
1913
1914         bio_put(bio);
1915         return 0;
1916 }
1917
1918 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
1919                          int mirror_num)
1920 {
1921         struct btrfs_mapping_tree *map_tree = &root->fs_info->mapping_tree;
1922         u64 start = eb->start;
1923         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
1924         int ret;
1925
1926         for (i = 0; i < num_pages; i++) {
1927                 struct page *p = extent_buffer_page(eb, i);
1928                 ret = repair_io_failure(map_tree, start, PAGE_CACHE_SIZE,
1929                                         start, p, mirror_num);
1930                 if (ret)
1931                         break;
1932                 start += PAGE_CACHE_SIZE;
1933         }
1934
1935         return ret;
1936 }
1937
1938 /*
1939  * each time an IO finishes, we do a fast check in the IO failure tree
1940  * to see if we need to process or clean up an io_failure_record
1941  */
1942 static int clean_io_failure(u64 start, struct page *page)
1943 {
1944         u64 private;
1945         u64 private_failure;
1946         struct io_failure_record *failrec;
1947         struct btrfs_mapping_tree *map_tree;
1948         struct extent_state *state;
1949         int num_copies;
1950         int did_repair = 0;
1951         int ret;
1952         struct inode *inode = page->mapping->host;
1953
1954         private = 0;
1955         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
1956                                 (u64)-1, 1, EXTENT_DIRTY, 0);
1957         if (!ret)
1958                 return 0;
1959
1960         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
1961                                 &private_failure);
1962         if (ret)
1963                 return 0;
1964
1965         failrec = (struct io_failure_record *)(unsigned long) private_failure;
1966         BUG_ON(!failrec->this_mirror);
1967
1968         if (failrec->in_validation) {
1969                 /* there was no real error, just free the record */
1970                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
1971                          failrec->start);
1972                 did_repair = 1;
1973                 goto out;
1974         }
1975
1976         spin_lock(&BTRFS_I(inode)->io_tree.lock);
1977         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
1978                                             failrec->start,
1979                                             EXTENT_LOCKED);
1980         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
1981
1982         if (state && state->start == failrec->start) {
1983                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
1984                 num_copies = btrfs_num_copies(map_tree, failrec->logical,
1985                                                 failrec->len);
1986                 if (num_copies > 1)  {
1987                         ret = repair_io_failure(map_tree, start, failrec->len,
1988                                                 failrec->logical, page,
1989                                                 failrec->failed_mirror);
1990                         did_repair = !ret;
1991                 }
1992         }
1993
1994 out:
1995         if (!ret)
1996                 ret = free_io_failure(inode, failrec, did_repair);
1997
1998         return ret;
1999 }
2000
2001 /*
2002  * this is a generic handler for readpage errors (default
2003  * readpage_io_failed_hook). if other copies exist, read those and write back
2004  * good data to the failed position. does not investigate in remapping the
2005  * failed extent elsewhere, hoping the device will be smart enough to do this as
2006  * needed
2007  */
2008
2009 static int bio_readpage_error(struct bio *failed_bio, struct page *page,
2010                                 u64 start, u64 end, int failed_mirror,
2011                                 struct extent_state *state)
2012 {
2013         struct io_failure_record *failrec = NULL;
2014         u64 private;
2015         struct extent_map *em;
2016         struct inode *inode = page->mapping->host;
2017         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2018         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2019         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2020         struct bio *bio;
2021         int num_copies;
2022         int ret;
2023         int read_mode;
2024         u64 logical;
2025
2026         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2027
2028         ret = get_state_private(failure_tree, start, &private);
2029         if (ret) {
2030                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2031                 if (!failrec)
2032                         return -ENOMEM;
2033                 failrec->start = start;
2034                 failrec->len = end - start + 1;
2035                 failrec->this_mirror = 0;
2036                 failrec->bio_flags = 0;
2037                 failrec->in_validation = 0;
2038
2039                 read_lock(&em_tree->lock);
2040                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2041                 if (!em) {
2042                         read_unlock(&em_tree->lock);
2043                         kfree(failrec);
2044                         return -EIO;
2045                 }
2046
2047                 if (em->start > start || em->start + em->len < start) {
2048                         free_extent_map(em);
2049                         em = NULL;
2050                 }
2051                 read_unlock(&em_tree->lock);
2052
2053                 if (!em || IS_ERR(em)) {
2054                         kfree(failrec);
2055                         return -EIO;
2056                 }
2057                 logical = start - em->start;
2058                 logical = em->block_start + logical;
2059                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2060                         logical = em->block_start;
2061                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2062                         extent_set_compress_type(&failrec->bio_flags,
2063                                                  em->compress_type);
2064                 }
2065                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2066                          "len=%llu\n", logical, start, failrec->len);
2067                 failrec->logical = logical;
2068                 free_extent_map(em);
2069
2070                 /* set the bits in the private failure tree */
2071                 ret = set_extent_bits(failure_tree, start, end,
2072                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2073                 if (ret >= 0)
2074                         ret = set_state_private(failure_tree, start,
2075                                                 (u64)(unsigned long)failrec);
2076                 /* set the bits in the inode's tree */
2077                 if (ret >= 0)
2078                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2079                                                 GFP_NOFS);
2080                 if (ret < 0) {
2081                         kfree(failrec);
2082                         return ret;
2083                 }
2084         } else {
2085                 failrec = (struct io_failure_record *)(unsigned long)private;
2086                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2087                          "start=%llu, len=%llu, validation=%d\n",
2088                          failrec->logical, failrec->start, failrec->len,
2089                          failrec->in_validation);
2090                 /*
2091                  * when data can be on disk more than twice, add to failrec here
2092                  * (e.g. with a list for failed_mirror) to make
2093                  * clean_io_failure() clean all those errors at once.
2094                  */
2095         }
2096         num_copies = btrfs_num_copies(
2097                               &BTRFS_I(inode)->root->fs_info->mapping_tree,
2098                               failrec->logical, failrec->len);
2099         if (num_copies == 1) {
2100                 /*
2101                  * we only have a single copy of the data, so don't bother with
2102                  * all the retry and error correction code that follows. no
2103                  * matter what the error is, it is very likely to persist.
2104                  */
2105                 pr_debug("bio_readpage_error: cannot repair, num_copies == 1. "
2106                          "state=%p, num_copies=%d, next_mirror %d, "
2107                          "failed_mirror %d\n", state, num_copies,
2108                          failrec->this_mirror, failed_mirror);
2109                 free_io_failure(inode, failrec, 0);
2110                 return -EIO;
2111         }
2112
2113         if (!state) {
2114                 spin_lock(&tree->lock);
2115                 state = find_first_extent_bit_state(tree, failrec->start,
2116                                                     EXTENT_LOCKED);
2117                 if (state && state->start != failrec->start)
2118                         state = NULL;
2119                 spin_unlock(&tree->lock);
2120         }
2121
2122         /*
2123          * there are two premises:
2124          *      a) deliver good data to the caller
2125          *      b) correct the bad sectors on disk
2126          */
2127         if (failed_bio->bi_vcnt > 1) {
2128                 /*
2129                  * to fulfill b), we need to know the exact failing sectors, as
2130                  * we don't want to rewrite any more than the failed ones. thus,
2131                  * we need separate read requests for the failed bio
2132                  *
2133                  * if the following BUG_ON triggers, our validation request got
2134                  * merged. we need separate requests for our algorithm to work.
2135                  */
2136                 BUG_ON(failrec->in_validation);
2137                 failrec->in_validation = 1;
2138                 failrec->this_mirror = failed_mirror;
2139                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2140         } else {
2141                 /*
2142                  * we're ready to fulfill a) and b) alongside. get a good copy
2143                  * of the failed sector and if we succeed, we have setup
2144                  * everything for repair_io_failure to do the rest for us.
2145                  */
2146                 if (failrec->in_validation) {
2147                         BUG_ON(failrec->this_mirror != failed_mirror);
2148                         failrec->in_validation = 0;
2149                         failrec->this_mirror = 0;
2150                 }
2151                 failrec->failed_mirror = failed_mirror;
2152                 failrec->this_mirror++;
2153                 if (failrec->this_mirror == failed_mirror)
2154                         failrec->this_mirror++;
2155                 read_mode = READ_SYNC;
2156         }
2157
2158         if (!state || failrec->this_mirror > num_copies) {
2159                 pr_debug("bio_readpage_error: (fail) state=%p, num_copies=%d, "
2160                          "next_mirror %d, failed_mirror %d\n", state,
2161                          num_copies, failrec->this_mirror, failed_mirror);
2162                 free_io_failure(inode, failrec, 0);
2163                 return -EIO;
2164         }
2165
2166         bio = bio_alloc(GFP_NOFS, 1);
2167         bio->bi_private = state;
2168         bio->bi_end_io = failed_bio->bi_end_io;
2169         bio->bi_sector = failrec->logical >> 9;
2170         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2171         bio->bi_size = 0;
2172
2173         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2174
2175         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2176                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2177                  failrec->this_mirror, num_copies, failrec->in_validation);
2178
2179         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2180                                          failrec->this_mirror,
2181                                          failrec->bio_flags, 0);
2182         return ret;
2183 }
2184
2185 /* lots and lots of room for performance fixes in the end_bio funcs */
2186
2187 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2188 {
2189         int uptodate = (err == 0);
2190         struct extent_io_tree *tree;
2191         int ret;
2192
2193         tree = &BTRFS_I(page->mapping->host)->io_tree;
2194
2195         if (tree->ops && tree->ops->writepage_end_io_hook) {
2196                 ret = tree->ops->writepage_end_io_hook(page, start,
2197                                                end, NULL, uptodate);
2198                 if (ret)
2199                         uptodate = 0;
2200         }
2201
2202         if (!uptodate && tree->ops &&
2203             tree->ops->writepage_io_failed_hook) {
2204                 ret = tree->ops->writepage_io_failed_hook(NULL, page,
2205                                                  start, end, NULL);
2206                 /* Writeback already completed */
2207                 if (ret == 0)
2208                         return 1;
2209         }
2210
2211         if (!uptodate) {
2212                 clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
2213                 ClearPageUptodate(page);
2214                 SetPageError(page);
2215         }
2216         return 0;
2217 }
2218
2219 /*
2220  * after a writepage IO is done, we need to:
2221  * clear the uptodate bits on error
2222  * clear the writeback bits in the extent tree for this IO
2223  * end_page_writeback if the page has no more pending IO
2224  *
2225  * Scheduling is not allowed, so the extent state tree is expected
2226  * to have one and only one object corresponding to this IO.
2227  */
2228 static void end_bio_extent_writepage(struct bio *bio, int err)
2229 {
2230         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2231         struct extent_io_tree *tree;
2232         u64 start;
2233         u64 end;
2234         int whole_page;
2235
2236         do {
2237                 struct page *page = bvec->bv_page;
2238                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2239
2240                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2241                          bvec->bv_offset;
2242                 end = start + bvec->bv_len - 1;
2243
2244                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2245                         whole_page = 1;
2246                 else
2247                         whole_page = 0;
2248
2249                 if (--bvec >= bio->bi_io_vec)
2250                         prefetchw(&bvec->bv_page->flags);
2251
2252                 if (end_extent_writepage(page, err, start, end))
2253                         continue;
2254
2255                 if (whole_page)
2256                         end_page_writeback(page);
2257                 else
2258                         check_page_writeback(tree, page);
2259         } while (bvec >= bio->bi_io_vec);
2260
2261         bio_put(bio);
2262 }
2263
2264 /*
2265  * after a readpage IO is done, we need to:
2266  * clear the uptodate bits on error
2267  * set the uptodate bits if things worked
2268  * set the page up to date if all extents in the tree are uptodate
2269  * clear the lock bit in the extent tree
2270  * unlock the page if there are no other extents locked for it
2271  *
2272  * Scheduling is not allowed, so the extent state tree is expected
2273  * to have one and only one object corresponding to this IO.
2274  */
2275 static void end_bio_extent_readpage(struct bio *bio, int err)
2276 {
2277         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2278         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2279         struct bio_vec *bvec = bio->bi_io_vec;
2280         struct extent_io_tree *tree;
2281         u64 start;
2282         u64 end;
2283         int whole_page;
2284         int failed_mirror;
2285         int ret;
2286
2287         if (err)
2288                 uptodate = 0;
2289
2290         do {
2291                 struct page *page = bvec->bv_page;
2292                 struct extent_state *cached = NULL;
2293                 struct extent_state *state;
2294
2295                 pr_debug("end_bio_extent_readpage: bi_vcnt=%d, idx=%d, err=%d, "
2296                          "mirror=%ld\n", bio->bi_vcnt, bio->bi_idx, err,
2297                          (long int)bio->bi_bdev);
2298                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2299
2300                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2301                         bvec->bv_offset;
2302                 end = start + bvec->bv_len - 1;
2303
2304                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2305                         whole_page = 1;
2306                 else
2307                         whole_page = 0;
2308
2309                 if (++bvec <= bvec_end)
2310                         prefetchw(&bvec->bv_page->flags);
2311
2312                 spin_lock(&tree->lock);
2313                 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
2314                 if (state && state->start == start) {
2315                         /*
2316                          * take a reference on the state, unlock will drop
2317                          * the ref
2318                          */
2319                         cache_state(state, &cached);
2320                 }
2321                 spin_unlock(&tree->lock);
2322
2323                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
2324                         ret = tree->ops->readpage_end_io_hook(page, start, end,
2325                                                               state);
2326                         if (ret)
2327                                 uptodate = 0;
2328                         else
2329                                 clean_io_failure(start, page);
2330                 }
2331
2332                 if (!uptodate)
2333                         failed_mirror = (int)(unsigned long)bio->bi_bdev;
2334
2335                 if (!uptodate && tree->ops && tree->ops->readpage_io_failed_hook) {
2336                         ret = tree->ops->readpage_io_failed_hook(page, failed_mirror);
2337                         if (!ret && !err &&
2338                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2339                                 uptodate = 1;
2340                 } else if (!uptodate) {
2341                         /*
2342                          * The generic bio_readpage_error handles errors the
2343                          * following way: If possible, new read requests are
2344                          * created and submitted and will end up in
2345                          * end_bio_extent_readpage as well (if we're lucky, not
2346                          * in the !uptodate case). In that case it returns 0 and
2347                          * we just go on with the next page in our bio. If it
2348                          * can't handle the error it will return -EIO and we
2349                          * remain responsible for that page.
2350                          */
2351                         ret = bio_readpage_error(bio, page, start, end,
2352                                                         failed_mirror, NULL);
2353                         if (ret == 0) {
2354                                 uptodate =
2355                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2356                                 if (err)
2357                                         uptodate = 0;
2358                                 uncache_state(&cached);
2359                                 continue;
2360                         }
2361                 }
2362
2363                 if (uptodate && tree->track_uptodate) {
2364                         set_extent_uptodate(tree, start, end, &cached,
2365                                             GFP_ATOMIC);
2366                 }
2367                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2368
2369                 if (whole_page) {
2370                         if (uptodate) {
2371                                 SetPageUptodate(page);
2372                         } else {
2373                                 ClearPageUptodate(page);
2374                                 SetPageError(page);
2375                         }
2376                         unlock_page(page);
2377                 } else {
2378                         if (uptodate) {
2379                                 check_page_uptodate(tree, page);
2380                         } else {
2381                                 ClearPageUptodate(page);
2382                                 SetPageError(page);
2383                         }
2384                         check_page_locked(tree, page);
2385                 }
2386         } while (bvec <= bvec_end);
2387
2388         bio_put(bio);
2389 }
2390
2391 struct bio *
2392 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2393                 gfp_t gfp_flags)
2394 {
2395         struct bio *bio;
2396
2397         bio = bio_alloc(gfp_flags, nr_vecs);
2398
2399         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2400                 while (!bio && (nr_vecs /= 2))
2401                         bio = bio_alloc(gfp_flags, nr_vecs);
2402         }
2403
2404         if (bio) {
2405                 bio->bi_size = 0;
2406                 bio->bi_bdev = bdev;
2407                 bio->bi_sector = first_sector;
2408         }
2409         return bio;
2410 }
2411
2412 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
2413                           unsigned long bio_flags)
2414 {
2415         int ret = 0;
2416         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2417         struct page *page = bvec->bv_page;
2418         struct extent_io_tree *tree = bio->bi_private;
2419         u64 start;
2420
2421         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
2422
2423         bio->bi_private = NULL;
2424
2425         bio_get(bio);
2426
2427         if (tree->ops && tree->ops->submit_bio_hook)
2428                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2429                                            mirror_num, bio_flags, start);
2430         else
2431                 btrfsic_submit_bio(rw, bio);
2432
2433         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2434                 ret = -EOPNOTSUPP;
2435         bio_put(bio);
2436         return ret;
2437 }
2438
2439 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2440                               struct page *page, sector_t sector,
2441                               size_t size, unsigned long offset,
2442                               struct block_device *bdev,
2443                               struct bio **bio_ret,
2444                               unsigned long max_pages,
2445                               bio_end_io_t end_io_func,
2446                               int mirror_num,
2447                               unsigned long prev_bio_flags,
2448                               unsigned long bio_flags)
2449 {
2450         int ret = 0;
2451         struct bio *bio;
2452         int nr;
2453         int contig = 0;
2454         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2455         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2456         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2457
2458         if (bio_ret && *bio_ret) {
2459                 bio = *bio_ret;
2460                 if (old_compressed)
2461                         contig = bio->bi_sector == sector;
2462                 else
2463                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
2464                                 sector;
2465
2466                 if (prev_bio_flags != bio_flags || !contig ||
2467                     (tree->ops && tree->ops->merge_bio_hook &&
2468                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
2469                                                bio_flags)) ||
2470                     bio_add_page(bio, page, page_size, offset) < page_size) {
2471                         ret = submit_one_bio(rw, bio, mirror_num,
2472                                              prev_bio_flags);
2473                         bio = NULL;
2474                 } else {
2475                         return 0;
2476                 }
2477         }
2478         if (this_compressed)
2479                 nr = BIO_MAX_PAGES;
2480         else
2481                 nr = bio_get_nr_vecs(bdev);
2482
2483         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2484         if (!bio)
2485                 return -ENOMEM;
2486
2487         bio_add_page(bio, page, page_size, offset);
2488         bio->bi_end_io = end_io_func;
2489         bio->bi_private = tree;
2490
2491         if (bio_ret)
2492                 *bio_ret = bio;
2493         else
2494                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2495
2496         return ret;
2497 }
2498
2499 void attach_extent_buffer_page(struct extent_buffer *eb, struct page *page)
2500 {
2501         if (!PagePrivate(page)) {
2502                 SetPagePrivate(page);
2503                 page_cache_get(page);
2504                 set_page_private(page, (unsigned long)eb);
2505         } else {
2506                 WARN_ON(page->private != (unsigned long)eb);
2507         }
2508 }
2509
2510 void set_page_extent_mapped(struct page *page)
2511 {
2512         if (!PagePrivate(page)) {
2513                 SetPagePrivate(page);
2514                 page_cache_get(page);
2515                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2516         }
2517 }
2518
2519 /*
2520  * basic readpage implementation.  Locked extent state structs are inserted
2521  * into the tree that are removed when the IO is done (by the end_io
2522  * handlers)
2523  */
2524 static int __extent_read_full_page(struct extent_io_tree *tree,
2525                                    struct page *page,
2526                                    get_extent_t *get_extent,
2527                                    struct bio **bio, int mirror_num,
2528                                    unsigned long *bio_flags)
2529 {
2530         struct inode *inode = page->mapping->host;
2531         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2532         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2533         u64 end;
2534         u64 cur = start;
2535         u64 extent_offset;
2536         u64 last_byte = i_size_read(inode);
2537         u64 block_start;
2538         u64 cur_end;
2539         sector_t sector;
2540         struct extent_map *em;
2541         struct block_device *bdev;
2542         struct btrfs_ordered_extent *ordered;
2543         int ret;
2544         int nr = 0;
2545         size_t pg_offset = 0;
2546         size_t iosize;
2547         size_t disk_io_size;
2548         size_t blocksize = inode->i_sb->s_blocksize;
2549         unsigned long this_bio_flag = 0;
2550
2551         set_page_extent_mapped(page);
2552
2553         if (!PageUptodate(page)) {
2554                 if (cleancache_get_page(page) == 0) {
2555                         BUG_ON(blocksize != PAGE_SIZE);
2556                         goto out;
2557                 }
2558         }
2559
2560         end = page_end;
2561         while (1) {
2562                 lock_extent(tree, start, end, GFP_NOFS);
2563                 ordered = btrfs_lookup_ordered_extent(inode, start);
2564                 if (!ordered)
2565                         break;
2566                 unlock_extent(tree, start, end, GFP_NOFS);
2567                 btrfs_start_ordered_extent(inode, ordered, 1);
2568                 btrfs_put_ordered_extent(ordered);
2569         }
2570
2571         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2572                 char *userpage;
2573                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2574
2575                 if (zero_offset) {
2576                         iosize = PAGE_CACHE_SIZE - zero_offset;
2577                         userpage = kmap_atomic(page, KM_USER0);
2578                         memset(userpage + zero_offset, 0, iosize);
2579                         flush_dcache_page(page);
2580                         kunmap_atomic(userpage, KM_USER0);
2581                 }
2582         }
2583         while (cur <= end) {
2584                 if (cur >= last_byte) {
2585                         char *userpage;
2586                         struct extent_state *cached = NULL;
2587
2588                         iosize = PAGE_CACHE_SIZE - pg_offset;
2589                         userpage = kmap_atomic(page, KM_USER0);
2590                         memset(userpage + pg_offset, 0, iosize);
2591                         flush_dcache_page(page);
2592                         kunmap_atomic(userpage, KM_USER0);
2593                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2594                                             &cached, GFP_NOFS);
2595                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2596                                              &cached, GFP_NOFS);
2597                         break;
2598                 }
2599                 em = get_extent(inode, page, pg_offset, cur,
2600                                 end - cur + 1, 0);
2601                 if (IS_ERR_OR_NULL(em)) {
2602                         SetPageError(page);
2603                         unlock_extent(tree, cur, end, GFP_NOFS);
2604                         break;
2605                 }
2606                 extent_offset = cur - em->start;
2607                 BUG_ON(extent_map_end(em) <= cur);
2608                 BUG_ON(end < cur);
2609
2610                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2611                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2612                         extent_set_compress_type(&this_bio_flag,
2613                                                  em->compress_type);
2614                 }
2615
2616                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2617                 cur_end = min(extent_map_end(em) - 1, end);
2618                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2619                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2620                         disk_io_size = em->block_len;
2621                         sector = em->block_start >> 9;
2622                 } else {
2623                         sector = (em->block_start + extent_offset) >> 9;
2624                         disk_io_size = iosize;
2625                 }
2626                 bdev = em->bdev;
2627                 block_start = em->block_start;
2628                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2629                         block_start = EXTENT_MAP_HOLE;
2630                 free_extent_map(em);
2631                 em = NULL;
2632
2633                 /* we've found a hole, just zero and go on */
2634                 if (block_start == EXTENT_MAP_HOLE) {
2635                         char *userpage;
2636                         struct extent_state *cached = NULL;
2637
2638                         userpage = kmap_atomic(page, KM_USER0);
2639                         memset(userpage + pg_offset, 0, iosize);
2640                         flush_dcache_page(page);
2641                         kunmap_atomic(userpage, KM_USER0);
2642
2643                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2644                                             &cached, GFP_NOFS);
2645                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2646                                              &cached, GFP_NOFS);
2647                         cur = cur + iosize;
2648                         pg_offset += iosize;
2649                         continue;
2650                 }
2651                 /* the get_extent function already copied into the page */
2652                 if (test_range_bit(tree, cur, cur_end,
2653                                    EXTENT_UPTODATE, 1, NULL)) {
2654                         check_page_uptodate(tree, page);
2655                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2656                         cur = cur + iosize;
2657                         pg_offset += iosize;
2658                         continue;
2659                 }
2660                 /* we have an inline extent but it didn't get marked up
2661                  * to date.  Error out
2662                  */
2663                 if (block_start == EXTENT_MAP_INLINE) {
2664                         SetPageError(page);
2665                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2666                         cur = cur + iosize;
2667                         pg_offset += iosize;
2668                         continue;
2669                 }
2670
2671                 ret = 0;
2672                 if (tree->ops && tree->ops->readpage_io_hook) {
2673                         ret = tree->ops->readpage_io_hook(page, cur,
2674                                                           cur + iosize - 1);
2675                 }
2676                 if (!ret) {
2677                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2678                         pnr -= page->index;
2679                         ret = submit_extent_page(READ, tree, page,
2680                                          sector, disk_io_size, pg_offset,
2681                                          bdev, bio, pnr,
2682                                          end_bio_extent_readpage, mirror_num,
2683                                          *bio_flags,
2684                                          this_bio_flag);
2685                         nr++;
2686                         *bio_flags = this_bio_flag;
2687                 }
2688                 if (ret)
2689                         SetPageError(page);
2690                 cur = cur + iosize;
2691                 pg_offset += iosize;
2692         }
2693 out:
2694         if (!nr) {
2695                 if (!PageError(page))
2696                         SetPageUptodate(page);
2697                 unlock_page(page);
2698         }
2699         return 0;
2700 }
2701
2702 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2703                             get_extent_t *get_extent, int mirror_num)
2704 {
2705         struct bio *bio = NULL;
2706         unsigned long bio_flags = 0;
2707         int ret;
2708
2709         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
2710                                       &bio_flags);
2711         if (bio)
2712                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
2713         return ret;
2714 }
2715
2716 static noinline void update_nr_written(struct page *page,
2717                                       struct writeback_control *wbc,
2718                                       unsigned long nr_written)
2719 {
2720         wbc->nr_to_write -= nr_written;
2721         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2722             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2723                 page->mapping->writeback_index = page->index + nr_written;
2724 }
2725
2726 /*
2727  * the writepage semantics are similar to regular writepage.  extent
2728  * records are inserted to lock ranges in the tree, and as dirty areas
2729  * are found, they are marked writeback.  Then the lock bits are removed
2730  * and the end_io handler clears the writeback ranges
2731  */
2732 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2733                               void *data)
2734 {
2735         struct inode *inode = page->mapping->host;
2736         struct extent_page_data *epd = data;
2737         struct extent_io_tree *tree = epd->tree;
2738         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2739         u64 delalloc_start;
2740         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2741         u64 end;
2742         u64 cur = start;
2743         u64 extent_offset;
2744         u64 last_byte = i_size_read(inode);
2745         u64 block_start;
2746         u64 iosize;
2747         sector_t sector;
2748         struct extent_state *cached_state = NULL;
2749         struct extent_map *em;
2750         struct block_device *bdev;
2751         int ret;
2752         int nr = 0;
2753         size_t pg_offset = 0;
2754         size_t blocksize;
2755         loff_t i_size = i_size_read(inode);
2756         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2757         u64 nr_delalloc;
2758         u64 delalloc_end;
2759         int page_started;
2760         int compressed;
2761         int write_flags;
2762         unsigned long nr_written = 0;
2763         bool fill_delalloc = true;
2764
2765         if (wbc->sync_mode == WB_SYNC_ALL)
2766                 write_flags = WRITE_SYNC;
2767         else
2768                 write_flags = WRITE;
2769
2770         trace___extent_writepage(page, inode, wbc);
2771
2772         WARN_ON(!PageLocked(page));
2773
2774         ClearPageError(page);
2775
2776         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2777         if (page->index > end_index ||
2778            (page->index == end_index && !pg_offset)) {
2779                 page->mapping->a_ops->invalidatepage(page, 0);
2780                 unlock_page(page);
2781                 return 0;
2782         }
2783
2784         if (page->index == end_index) {
2785                 char *userpage;
2786
2787                 userpage = kmap_atomic(page, KM_USER0);
2788                 memset(userpage + pg_offset, 0,
2789                        PAGE_CACHE_SIZE - pg_offset);
2790                 kunmap_atomic(userpage, KM_USER0);
2791                 flush_dcache_page(page);
2792         }
2793         pg_offset = 0;
2794
2795         set_page_extent_mapped(page);
2796
2797         if (!tree->ops || !tree->ops->fill_delalloc)
2798                 fill_delalloc = false;
2799
2800         delalloc_start = start;
2801         delalloc_end = 0;
2802         page_started = 0;
2803         if (!epd->extent_locked && fill_delalloc) {
2804                 u64 delalloc_to_write = 0;
2805                 /*
2806                  * make sure the wbc mapping index is at least updated
2807                  * to this page.
2808                  */
2809                 update_nr_written(page, wbc, 0);
2810
2811                 while (delalloc_end < page_end) {
2812                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2813                                                        page,
2814                                                        &delalloc_start,
2815                                                        &delalloc_end,
2816                                                        128 * 1024 * 1024);
2817                         if (nr_delalloc == 0) {
2818                                 delalloc_start = delalloc_end + 1;
2819                                 continue;
2820                         }
2821                         ret = tree->ops->fill_delalloc(inode, page,
2822                                                        delalloc_start,
2823                                                        delalloc_end,
2824                                                        &page_started,
2825                                                        &nr_written);
2826                         BUG_ON(ret);
2827                         /*
2828                          * delalloc_end is already one less than the total
2829                          * length, so we don't subtract one from
2830                          * PAGE_CACHE_SIZE
2831                          */
2832                         delalloc_to_write += (delalloc_end - delalloc_start +
2833                                               PAGE_CACHE_SIZE) >>
2834                                               PAGE_CACHE_SHIFT;
2835                         delalloc_start = delalloc_end + 1;
2836                 }
2837                 if (wbc->nr_to_write < delalloc_to_write) {
2838                         int thresh = 8192;
2839
2840                         if (delalloc_to_write < thresh * 2)
2841                                 thresh = delalloc_to_write;
2842                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2843                                                  thresh);
2844                 }
2845
2846                 /* did the fill delalloc function already unlock and start
2847                  * the IO?
2848                  */
2849                 if (page_started) {
2850                         ret = 0;
2851                         /*
2852                          * we've unlocked the page, so we can't update
2853                          * the mapping's writeback index, just update
2854                          * nr_to_write.
2855                          */
2856                         wbc->nr_to_write -= nr_written;
2857                         goto done_unlocked;
2858                 }
2859         }
2860         if (tree->ops && tree->ops->writepage_start_hook) {
2861                 ret = tree->ops->writepage_start_hook(page, start,
2862                                                       page_end);
2863                 if (ret) {
2864                         /* Fixup worker will requeue */
2865                         if (ret == -EBUSY)
2866                                 wbc->pages_skipped++;
2867                         else
2868                                 redirty_page_for_writepage(wbc, page);
2869                         update_nr_written(page, wbc, nr_written);
2870                         unlock_page(page);
2871                         ret = 0;
2872                         goto done_unlocked;
2873                 }
2874         }
2875
2876         /*
2877          * we don't want to touch the inode after unlocking the page,
2878          * so we update the mapping writeback index now
2879          */
2880         update_nr_written(page, wbc, nr_written + 1);
2881
2882         end = page_end;
2883         if (last_byte <= start) {
2884                 if (tree->ops && tree->ops->writepage_end_io_hook)
2885                         tree->ops->writepage_end_io_hook(page, start,
2886                                                          page_end, NULL, 1);
2887                 goto done;
2888         }
2889
2890         blocksize = inode->i_sb->s_blocksize;
2891
2892         while (cur <= end) {
2893                 if (cur >= last_byte) {
2894                         if (tree->ops && tree->ops->writepage_end_io_hook)
2895                                 tree->ops->writepage_end_io_hook(page, cur,
2896                                                          page_end, NULL, 1);
2897                         break;
2898                 }
2899                 em = epd->get_extent(inode, page, pg_offset, cur,
2900                                      end - cur + 1, 1);
2901                 if (IS_ERR_OR_NULL(em)) {
2902                         SetPageError(page);
2903                         break;
2904                 }
2905
2906                 extent_offset = cur - em->start;
2907                 BUG_ON(extent_map_end(em) <= cur);
2908                 BUG_ON(end < cur);
2909                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2910                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2911                 sector = (em->block_start + extent_offset) >> 9;
2912                 bdev = em->bdev;
2913                 block_start = em->block_start;
2914                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2915                 free_extent_map(em);
2916                 em = NULL;
2917
2918                 /*
2919                  * compressed and inline extents are written through other
2920                  * paths in the FS
2921                  */
2922                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2923                     block_start == EXTENT_MAP_INLINE) {
2924                         /*
2925                          * end_io notification does not happen here for
2926                          * compressed extents
2927                          */
2928                         if (!compressed && tree->ops &&
2929                             tree->ops->writepage_end_io_hook)
2930                                 tree->ops->writepage_end_io_hook(page, cur,
2931                                                          cur + iosize - 1,
2932                                                          NULL, 1);
2933                         else if (compressed) {
2934                                 /* we don't want to end_page_writeback on
2935                                  * a compressed extent.  this happens
2936                                  * elsewhere
2937                                  */
2938                                 nr++;
2939                         }
2940
2941                         cur += iosize;
2942                         pg_offset += iosize;
2943                         continue;
2944                 }
2945                 /* leave this out until we have a page_mkwrite call */
2946                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2947                                    EXTENT_DIRTY, 0, NULL)) {
2948                         cur = cur + iosize;
2949                         pg_offset += iosize;
2950                         continue;
2951                 }
2952
2953                 if (tree->ops && tree->ops->writepage_io_hook) {
2954                         ret = tree->ops->writepage_io_hook(page, cur,
2955                                                 cur + iosize - 1);
2956                 } else {
2957                         ret = 0;
2958                 }
2959                 if (ret) {
2960                         SetPageError(page);
2961                 } else {
2962                         unsigned long max_nr = end_index + 1;
2963
2964                         set_range_writeback(tree, cur, cur + iosize - 1);
2965                         if (!PageWriteback(page)) {
2966                                 printk(KERN_ERR "btrfs warning page %lu not "
2967                                        "writeback, cur %llu end %llu\n",
2968                                        page->index, (unsigned long long)cur,
2969                                        (unsigned long long)end);
2970                         }
2971
2972                         ret = submit_extent_page(write_flags, tree, page,
2973                                                  sector, iosize, pg_offset,
2974                                                  bdev, &epd->bio, max_nr,
2975                                                  end_bio_extent_writepage,
2976                                                  0, 0, 0);
2977                         if (ret)
2978                                 SetPageError(page);
2979                 }
2980                 cur = cur + iosize;
2981                 pg_offset += iosize;
2982                 nr++;
2983         }
2984 done:
2985         if (nr == 0) {
2986                 /* make sure the mapping tag for page dirty gets cleared */
2987                 set_page_writeback(page);
2988                 end_page_writeback(page);
2989         }
2990         unlock_page(page);
2991
2992 done_unlocked:
2993
2994         /* drop our reference on any cached states */
2995         free_extent_state(cached_state);
2996         return 0;
2997 }
2998
2999 static int eb_wait(void *word)
3000 {
3001         io_schedule();
3002         return 0;
3003 }
3004
3005 static void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3006 {
3007         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3008                     TASK_UNINTERRUPTIBLE);
3009 }
3010
3011 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3012                                      struct btrfs_fs_info *fs_info,
3013                                      struct extent_page_data *epd)
3014 {
3015         unsigned long i, num_pages;
3016         int flush = 0;
3017         int ret = 0;
3018
3019         if (!btrfs_try_tree_write_lock(eb)) {
3020                 flush = 1;
3021                 flush_write_bio(epd);
3022                 btrfs_tree_lock(eb);
3023         }
3024
3025         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3026                 btrfs_tree_unlock(eb);
3027                 if (!epd->sync_io)
3028                         return 0;
3029                 if (!flush) {
3030                         flush_write_bio(epd);
3031                         flush = 1;
3032                 }
3033                 while (1) {
3034                         wait_on_extent_buffer_writeback(eb);
3035                         btrfs_tree_lock(eb);
3036                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3037                                 break;
3038                         btrfs_tree_unlock(eb);
3039                 }
3040         }
3041
3042         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3043                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3044                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3045                 spin_lock(&fs_info->delalloc_lock);
3046                 if (fs_info->dirty_metadata_bytes >= eb->len)
3047                         fs_info->dirty_metadata_bytes -= eb->len;
3048                 else
3049                         WARN_ON(1);
3050                 spin_unlock(&fs_info->delalloc_lock);
3051                 ret = 1;
3052         }
3053
3054         btrfs_tree_unlock(eb);
3055
3056         if (!ret)
3057                 return ret;
3058
3059         num_pages = num_extent_pages(eb->start, eb->len);
3060         for (i = 0; i < num_pages; i++) {
3061                 struct page *p = extent_buffer_page(eb, i);
3062
3063                 if (!trylock_page(p)) {
3064                         if (!flush) {
3065                                 flush_write_bio(epd);
3066                                 flush = 1;
3067                         }
3068                         lock_page(p);
3069                 }
3070         }
3071
3072         return ret;
3073 }
3074
3075 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3076 {
3077         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3078         smp_mb__after_clear_bit();
3079         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3080 }
3081
3082 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3083 {
3084         int uptodate = err == 0;
3085         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
3086         struct extent_buffer *eb;
3087         int done;
3088
3089         do {
3090                 struct page *page = bvec->bv_page;
3091
3092                 bvec--;
3093                 eb = (struct extent_buffer *)page->private;
3094                 BUG_ON(!eb);
3095                 done = atomic_dec_and_test(&eb->io_pages);
3096
3097                 if (!uptodate || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3098                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3099                         ClearPageUptodate(page);
3100                         SetPageError(page);
3101                 }
3102
3103                 end_page_writeback(page);
3104
3105                 if (!done)
3106                         continue;
3107
3108                 end_extent_buffer_writeback(eb);
3109         } while (bvec >= bio->bi_io_vec);
3110
3111         bio_put(bio);
3112
3113 }
3114
3115 static int write_one_eb(struct extent_buffer *eb,
3116                         struct btrfs_fs_info *fs_info,
3117                         struct writeback_control *wbc,
3118                         struct extent_page_data *epd)
3119 {
3120         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3121         u64 offset = eb->start;
3122         unsigned long i, num_pages;
3123         int rw = (epd->sync_io ? WRITE_SYNC : WRITE);
3124         int ret;
3125
3126         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3127         num_pages = num_extent_pages(eb->start, eb->len);
3128         atomic_set(&eb->io_pages, num_pages);
3129         for (i = 0; i < num_pages; i++) {
3130                 struct page *p = extent_buffer_page(eb, i);
3131
3132                 clear_page_dirty_for_io(p);
3133                 set_page_writeback(p);
3134                 ret = submit_extent_page(rw, eb->tree, p, offset >> 9,
3135                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3136                                          -1, end_bio_extent_buffer_writepage,
3137                                          0, 0, 0);
3138                 if (ret) {
3139                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3140                         SetPageError(p);
3141                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3142                                 end_extent_buffer_writeback(eb);
3143                         ret = -EIO;
3144                         break;
3145                 }
3146                 offset += PAGE_CACHE_SIZE;
3147                 update_nr_written(p, wbc, 1);
3148                 unlock_page(p);
3149         }
3150
3151         if (unlikely(ret)) {
3152                 for (; i < num_pages; i++) {
3153                         struct page *p = extent_buffer_page(eb, i);
3154                         unlock_page(p);
3155                 }
3156         }
3157
3158         return ret;
3159 }
3160
3161 int btree_write_cache_pages(struct address_space *mapping,
3162                                    struct writeback_control *wbc)
3163 {
3164         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3165         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3166         struct extent_buffer *eb, *prev_eb = NULL;
3167         struct extent_page_data epd = {
3168                 .bio = NULL,
3169                 .tree = tree,
3170                 .extent_locked = 0,
3171                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3172         };
3173         int ret = 0;
3174         int done = 0;
3175         int nr_to_write_done = 0;
3176         struct pagevec pvec;
3177         int nr_pages;
3178         pgoff_t index;
3179         pgoff_t end;            /* Inclusive */
3180         int scanned = 0;
3181         int tag;
3182
3183         pagevec_init(&pvec, 0);
3184         if (wbc->range_cyclic) {
3185                 index = mapping->writeback_index; /* Start from prev offset */
3186                 end = -1;
3187         } else {
3188                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3189                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3190                 scanned = 1;
3191         }
3192         if (wbc->sync_mode == WB_SYNC_ALL)
3193                 tag = PAGECACHE_TAG_TOWRITE;
3194         else
3195                 tag = PAGECACHE_TAG_DIRTY;
3196 retry:
3197         if (wbc->sync_mode == WB_SYNC_ALL)
3198                 tag_pages_for_writeback(mapping, index, end);
3199         while (!done && !nr_to_write_done && (index <= end) &&
3200                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3201                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3202                 unsigned i;
3203
3204                 scanned = 1;
3205                 for (i = 0; i < nr_pages; i++) {
3206                         struct page *page = pvec.pages[i];
3207
3208                         if (!PagePrivate(page))
3209                                 continue;
3210
3211                         if (!wbc->range_cyclic && page->index > end) {
3212                                 done = 1;
3213                                 break;
3214                         }
3215
3216                         eb = (struct extent_buffer *)page->private;
3217                         if (!eb) {
3218                                 WARN_ON(1);
3219                                 continue;
3220                         }
3221
3222                         if (eb == prev_eb)
3223                                 continue;
3224
3225                         if (!atomic_inc_not_zero(&eb->refs)) {
3226                                 WARN_ON(1);
3227                                 continue;
3228                         }
3229
3230                         prev_eb = eb;
3231                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3232                         if (!ret) {
3233                                 free_extent_buffer(eb);
3234                                 continue;
3235                         }
3236
3237                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3238                         if (ret) {
3239                                 done = 1;
3240                                 free_extent_buffer(eb);
3241                                 break;
3242                         }
3243                         free_extent_buffer(eb);
3244
3245                         /*
3246                          * the filesystem may choose to bump up nr_to_write.
3247                          * We have to make sure to honor the new nr_to_write
3248                          * at any time
3249                          */
3250                         nr_to_write_done = wbc->nr_to_write <= 0;
3251                 }
3252                 pagevec_release(&pvec);
3253                 cond_resched();
3254         }
3255         if (!scanned && !done) {
3256                 /*
3257                  * We hit the last page and there is more work to be done: wrap
3258                  * back to the start of the file
3259                  */
3260                 scanned = 1;
3261                 index = 0;
3262                 goto retry;
3263         }
3264         flush_write_bio(&epd);
3265         return ret;
3266 }
3267
3268 /**
3269  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3270  * @mapping: address space structure to write
3271  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3272  * @writepage: function called for each page
3273  * @data: data passed to writepage function
3274  *
3275  * If a page is already under I/O, write_cache_pages() skips it, even
3276  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3277  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3278  * and msync() need to guarantee that all the data which was dirty at the time
3279  * the call was made get new I/O started against them.  If wbc->sync_mode is
3280  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3281  * existing IO to complete.
3282  */
3283 static int extent_write_cache_pages(struct extent_io_tree *tree,
3284                              struct address_space *mapping,
3285                              struct writeback_control *wbc,
3286                              writepage_t writepage, void *data,
3287                              void (*flush_fn)(void *))
3288 {
3289         int ret = 0;
3290         int done = 0;
3291         int nr_to_write_done = 0;
3292         struct pagevec pvec;
3293         int nr_pages;
3294         pgoff_t index;
3295         pgoff_t end;            /* Inclusive */
3296         int scanned = 0;
3297         int tag;
3298
3299         pagevec_init(&pvec, 0);
3300         if (wbc->range_cyclic) {
3301                 index = mapping->writeback_index; /* Start from prev offset */
3302                 end = -1;
3303         } else {
3304                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3305                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3306                 scanned = 1;
3307         }
3308         if (wbc->sync_mode == WB_SYNC_ALL)
3309                 tag = PAGECACHE_TAG_TOWRITE;
3310         else
3311                 tag = PAGECACHE_TAG_DIRTY;
3312 retry:
3313         if (wbc->sync_mode == WB_SYNC_ALL)
3314                 tag_pages_for_writeback(mapping, index, end);
3315         while (!done && !nr_to_write_done && (index <= end) &&
3316                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3317                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3318                 unsigned i;
3319
3320                 scanned = 1;
3321                 for (i = 0; i < nr_pages; i++) {
3322                         struct page *page = pvec.pages[i];
3323
3324                         /*
3325                          * At this point we hold neither mapping->tree_lock nor
3326                          * lock on the page itself: the page may be truncated or
3327                          * invalidated (changing page->mapping to NULL), or even
3328                          * swizzled back from swapper_space to tmpfs file
3329                          * mapping
3330                          */
3331                         if (tree->ops &&
3332                             tree->ops->write_cache_pages_lock_hook) {
3333                                 tree->ops->write_cache_pages_lock_hook(page,
3334                                                                data, flush_fn);
3335                         } else {
3336                                 if (!trylock_page(page)) {
3337                                         flush_fn(data);
3338                                         lock_page(page);
3339                                 }
3340                         }
3341
3342                         if (unlikely(page->mapping != mapping)) {
3343                                 unlock_page(page);
3344                                 continue;
3345                         }
3346
3347                         if (!wbc->range_cyclic && page->index > end) {
3348                                 done = 1;
3349                                 unlock_page(page);
3350                                 continue;
3351                         }
3352
3353                         if (wbc->sync_mode != WB_SYNC_NONE) {
3354                                 if (PageWriteback(page))
3355                                         flush_fn(data);
3356                                 wait_on_page_writeback(page);
3357                         }
3358
3359                         if (PageWriteback(page) ||
3360                             !clear_page_dirty_for_io(page)) {
3361                                 unlock_page(page);
3362                                 continue;
3363                         }
3364
3365                         ret = (*writepage)(page, wbc, data);
3366
3367                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3368                                 unlock_page(page);
3369                                 ret = 0;
3370                         }
3371                         if (ret)
3372                                 done = 1;
3373
3374                         /*
3375                          * the filesystem may choose to bump up nr_to_write.
3376                          * We have to make sure to honor the new nr_to_write
3377                          * at any time
3378                          */
3379                         nr_to_write_done = wbc->nr_to_write <= 0;
3380                 }
3381                 pagevec_release(&pvec);
3382                 cond_resched();
3383         }
3384         if (!scanned && !done) {
3385                 /*
3386                  * We hit the last page and there is more work to be done: wrap
3387                  * back to the start of the file
3388                  */
3389                 scanned = 1;
3390                 index = 0;
3391                 goto retry;
3392         }
3393         return ret;
3394 }
3395
3396 static void flush_epd_write_bio(struct extent_page_data *epd)
3397 {
3398         if (epd->bio) {
3399                 if (epd->sync_io)
3400                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
3401                 else
3402                         submit_one_bio(WRITE, epd->bio, 0, 0);
3403                 epd->bio = NULL;
3404         }
3405 }
3406
3407 static noinline void flush_write_bio(void *data)
3408 {
3409         struct extent_page_data *epd = data;
3410         flush_epd_write_bio(epd);
3411 }
3412
3413 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3414                           get_extent_t *get_extent,
3415                           struct writeback_control *wbc)
3416 {
3417         int ret;
3418         struct extent_page_data epd = {
3419                 .bio = NULL,
3420                 .tree = tree,
3421                 .get_extent = get_extent,
3422                 .extent_locked = 0,
3423                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3424         };
3425
3426         ret = __extent_writepage(page, wbc, &epd);
3427
3428         flush_epd_write_bio(&epd);
3429         return ret;
3430 }
3431
3432 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3433                               u64 start, u64 end, get_extent_t *get_extent,
3434                               int mode)
3435 {
3436         int ret = 0;
3437         struct address_space *mapping = inode->i_mapping;
3438         struct page *page;
3439         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3440                 PAGE_CACHE_SHIFT;
3441
3442         struct extent_page_data epd = {
3443                 .bio = NULL,
3444                 .tree = tree,
3445                 .get_extent = get_extent,
3446                 .extent_locked = 1,
3447                 .sync_io = mode == WB_SYNC_ALL,
3448         };
3449         struct writeback_control wbc_writepages = {
3450                 .sync_mode      = mode,
3451                 .nr_to_write    = nr_pages * 2,
3452                 .range_start    = start,
3453                 .range_end      = end + 1,
3454         };
3455
3456         while (start <= end) {
3457                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3458                 if (clear_page_dirty_for_io(page))
3459                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3460                 else {
3461                         if (tree->ops && tree->ops->writepage_end_io_hook)
3462                                 tree->ops->writepage_end_io_hook(page, start,
3463                                                  start + PAGE_CACHE_SIZE - 1,
3464                                                  NULL, 1);
3465                         unlock_page(page);
3466                 }
3467                 page_cache_release(page);
3468                 start += PAGE_CACHE_SIZE;
3469         }
3470
3471         flush_epd_write_bio(&epd);
3472         return ret;
3473 }
3474
3475 int extent_writepages(struct extent_io_tree *tree,
3476                       struct address_space *mapping,
3477                       get_extent_t *get_extent,
3478                       struct writeback_control *wbc)
3479 {
3480         int ret = 0;
3481         struct extent_page_data epd = {
3482                 .bio = NULL,
3483                 .tree = tree,
3484                 .get_extent = get_extent,
3485                 .extent_locked = 0,
3486                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3487         };
3488
3489         ret = extent_write_cache_pages(tree, mapping, wbc,
3490                                        __extent_writepage, &epd,
3491                                        flush_write_bio);
3492         flush_epd_write_bio(&epd);
3493         return ret;
3494 }
3495
3496 int extent_readpages(struct extent_io_tree *tree,
3497                      struct address_space *mapping,
3498                      struct list_head *pages, unsigned nr_pages,
3499                      get_extent_t get_extent)
3500 {
3501         struct bio *bio = NULL;
3502         unsigned page_idx;
3503         unsigned long bio_flags = 0;
3504
3505         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3506                 struct page *page = list_entry(pages->prev, struct page, lru);
3507
3508                 prefetchw(&page->flags);
3509                 list_del(&page->lru);
3510                 if (!add_to_page_cache_lru(page, mapping,
3511                                         page->index, GFP_NOFS)) {
3512                         __extent_read_full_page(tree, page, get_extent,
3513                                                 &bio, 0, &bio_flags);
3514                 }
3515                 page_cache_release(page);
3516         }
3517         BUG_ON(!list_empty(pages));
3518         if (bio)
3519                 submit_one_bio(READ, bio, 0, bio_flags);
3520         return 0;
3521 }
3522
3523 /*
3524  * basic invalidatepage code, this waits on any locked or writeback
3525  * ranges corresponding to the page, and then deletes any extent state
3526  * records from the tree
3527  */
3528 int extent_invalidatepage(struct extent_io_tree *tree,
3529                           struct page *page, unsigned long offset)
3530 {
3531         struct extent_state *cached_state = NULL;
3532         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
3533         u64 end = start + PAGE_CACHE_SIZE - 1;
3534         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3535
3536         start += (offset + blocksize - 1) & ~(blocksize - 1);
3537         if (start > end)
3538                 return 0;
3539
3540         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
3541         wait_on_page_writeback(page);
3542         clear_extent_bit(tree, start, end,
3543                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3544                          EXTENT_DO_ACCOUNTING,
3545                          1, 1, &cached_state, GFP_NOFS);
3546         return 0;
3547 }
3548
3549 /*
3550  * a helper for releasepage, this tests for areas of the page that
3551  * are locked or under IO and drops the related state bits if it is safe
3552  * to drop the page.
3553  */
3554 int try_release_extent_state(struct extent_map_tree *map,
3555                              struct extent_io_tree *tree, struct page *page,
3556                              gfp_t mask)
3557 {
3558         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3559         u64 end = start + PAGE_CACHE_SIZE - 1;
3560         int ret = 1;
3561
3562         if (test_range_bit(tree, start, end,
3563                            EXTENT_IOBITS, 0, NULL))
3564                 ret = 0;
3565         else {
3566                 if ((mask & GFP_NOFS) == GFP_NOFS)
3567                         mask = GFP_NOFS;
3568                 /*
3569                  * at this point we can safely clear everything except the
3570                  * locked bit and the nodatasum bit
3571                  */
3572                 ret = clear_extent_bit(tree, start, end,
3573                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3574                                  0, 0, NULL, mask);
3575
3576                 /* if clear_extent_bit failed for enomem reasons,
3577                  * we can't allow the release to continue.
3578                  */
3579                 if (ret < 0)
3580                         ret = 0;
3581                 else
3582                         ret = 1;
3583         }
3584         return ret;
3585 }
3586
3587 /*
3588  * a helper for releasepage.  As long as there are no locked extents
3589  * in the range corresponding to the page, both state records and extent
3590  * map records are removed
3591  */
3592 int try_release_extent_mapping(struct extent_map_tree *map,
3593                                struct extent_io_tree *tree, struct page *page,
3594                                gfp_t mask)
3595 {
3596         struct extent_map *em;
3597         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3598         u64 end = start + PAGE_CACHE_SIZE - 1;
3599
3600         if ((mask & __GFP_WAIT) &&
3601             page->mapping->host->i_size > 16 * 1024 * 1024) {
3602                 u64 len;
3603                 while (start <= end) {
3604                         len = end - start + 1;
3605                         write_lock(&map->lock);
3606                         em = lookup_extent_mapping(map, start, len);
3607                         if (!em) {
3608                                 write_unlock(&map->lock);
3609                                 break;
3610                         }
3611                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3612                             em->start != start) {
3613                                 write_unlock(&map->lock);
3614                                 free_extent_map(em);
3615                                 break;
3616                         }
3617                         if (!test_range_bit(tree, em->start,
3618                                             extent_map_end(em) - 1,
3619                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
3620                                             0, NULL)) {
3621                                 remove_extent_mapping(map, em);
3622                                 /* once for the rb tree */
3623                                 free_extent_map(em);
3624                         }
3625                         start = extent_map_end(em);
3626                         write_unlock(&map->lock);
3627
3628                         /* once for us */
3629                         free_extent_map(em);
3630                 }
3631         }
3632         return try_release_extent_state(map, tree, page, mask);
3633 }
3634
3635 /*
3636  * helper function for fiemap, which doesn't want to see any holes.
3637  * This maps until we find something past 'last'
3638  */
3639 static struct extent_map *get_extent_skip_holes(struct inode *inode,
3640                                                 u64 offset,
3641                                                 u64 last,
3642                                                 get_extent_t *get_extent)
3643 {
3644         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
3645         struct extent_map *em;
3646         u64 len;
3647
3648         if (offset >= last)
3649                 return NULL;
3650
3651         while(1) {
3652                 len = last - offset;
3653                 if (len == 0)
3654                         break;
3655                 len = (len + sectorsize - 1) & ~(sectorsize - 1);
3656                 em = get_extent(inode, NULL, 0, offset, len, 0);
3657                 if (IS_ERR_OR_NULL(em))
3658                         return em;
3659
3660                 /* if this isn't a hole return it */
3661                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
3662                     em->block_start != EXTENT_MAP_HOLE) {
3663                         return em;
3664                 }
3665
3666                 /* this is a hole, advance to the next extent */
3667                 offset = extent_map_end(em);
3668                 free_extent_map(em);
3669                 if (offset >= last)
3670                         break;
3671         }
3672         return NULL;
3673 }
3674
3675 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
3676                 __u64 start, __u64 len, get_extent_t *get_extent)
3677 {
3678         int ret = 0;
3679         u64 off = start;
3680         u64 max = start + len;
3681         u32 flags = 0;
3682         u32 found_type;
3683         u64 last;
3684         u64 last_for_get_extent = 0;
3685         u64 disko = 0;
3686         u64 isize = i_size_read(inode);
3687         struct btrfs_key found_key;
3688         struct extent_map *em = NULL;
3689         struct extent_state *cached_state = NULL;
3690         struct btrfs_path *path;
3691         struct btrfs_file_extent_item *item;
3692         int end = 0;
3693         u64 em_start = 0;
3694         u64 em_len = 0;
3695         u64 em_end = 0;
3696         unsigned long emflags;
3697
3698         if (len == 0)
3699                 return -EINVAL;
3700
3701         path = btrfs_alloc_path();
3702         if (!path)
3703                 return -ENOMEM;
3704         path->leave_spinning = 1;
3705
3706         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
3707         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
3708
3709         /*
3710          * lookup the last file extent.  We're not using i_size here
3711          * because there might be preallocation past i_size
3712          */
3713         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3714                                        path, btrfs_ino(inode), -1, 0);
3715         if (ret < 0) {
3716                 btrfs_free_path(path);
3717                 return ret;
3718         }
3719         WARN_ON(!ret);
3720         path->slots[0]--;
3721         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3722                               struct btrfs_file_extent_item);
3723         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3724         found_type = btrfs_key_type(&found_key);
3725
3726         /* No extents, but there might be delalloc bits */
3727         if (found_key.objectid != btrfs_ino(inode) ||
3728             found_type != BTRFS_EXTENT_DATA_KEY) {
3729                 /* have to trust i_size as the end */
3730                 last = (u64)-1;
3731                 last_for_get_extent = isize;
3732         } else {
3733                 /*
3734                  * remember the start of the last extent.  There are a
3735                  * bunch of different factors that go into the length of the
3736                  * extent, so its much less complex to remember where it started
3737                  */
3738                 last = found_key.offset;
3739                 last_for_get_extent = last + 1;
3740         }
3741         btrfs_free_path(path);
3742
3743         /*
3744          * we might have some extents allocated but more delalloc past those
3745          * extents.  so, we trust isize unless the start of the last extent is
3746          * beyond isize
3747          */
3748         if (last < isize) {
3749                 last = (u64)-1;
3750                 last_for_get_extent = isize;
3751         }
3752
3753         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3754                          &cached_state, GFP_NOFS);
3755
3756         em = get_extent_skip_holes(inode, start, last_for_get_extent,
3757                                    get_extent);
3758         if (!em)
3759                 goto out;
3760         if (IS_ERR(em)) {
3761                 ret = PTR_ERR(em);
3762                 goto out;
3763         }
3764
3765         while (!end) {
3766                 u64 offset_in_extent;
3767
3768                 /* break if the extent we found is outside the range */
3769                 if (em->start >= max || extent_map_end(em) < off)
3770                         break;
3771
3772                 /*
3773                  * get_extent may return an extent that starts before our
3774                  * requested range.  We have to make sure the ranges
3775                  * we return to fiemap always move forward and don't
3776                  * overlap, so adjust the offsets here
3777                  */
3778                 em_start = max(em->start, off);
3779
3780                 /*
3781                  * record the offset from the start of the extent
3782                  * for adjusting the disk offset below
3783                  */
3784                 offset_in_extent = em_start - em->start;
3785                 em_end = extent_map_end(em);
3786                 em_len = em_end - em_start;
3787                 emflags = em->flags;
3788                 disko = 0;
3789                 flags = 0;
3790
3791                 /*
3792                  * bump off for our next call to get_extent
3793                  */
3794                 off = extent_map_end(em);
3795                 if (off >= max)
3796                         end = 1;
3797
3798                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3799                         end = 1;
3800                         flags |= FIEMAP_EXTENT_LAST;
3801                 } else if (em->block_start == EXTENT_MAP_INLINE) {
3802                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
3803                                   FIEMAP_EXTENT_NOT_ALIGNED);
3804                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3805                         flags |= (FIEMAP_EXTENT_DELALLOC |
3806                                   FIEMAP_EXTENT_UNKNOWN);
3807                 } else {
3808                         disko = em->block_start + offset_in_extent;
3809                 }
3810                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3811                         flags |= FIEMAP_EXTENT_ENCODED;
3812
3813                 free_extent_map(em);
3814                 em = NULL;
3815                 if ((em_start >= last) || em_len == (u64)-1 ||
3816                    (last == (u64)-1 && isize <= em_end)) {
3817                         flags |= FIEMAP_EXTENT_LAST;
3818                         end = 1;
3819                 }
3820
3821                 /* now scan forward to see if this is really the last extent. */
3822                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
3823                                            get_extent);
3824                 if (IS_ERR(em)) {
3825                         ret = PTR_ERR(em);
3826                         goto out;
3827                 }
3828                 if (!em) {
3829                         flags |= FIEMAP_EXTENT_LAST;
3830                         end = 1;
3831                 }
3832                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3833                                               em_len, flags);
3834                 if (ret)
3835                         goto out_free;
3836         }
3837 out_free:
3838         free_extent_map(em);
3839 out:
3840         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3841                              &cached_state, GFP_NOFS);
3842         return ret;
3843 }
3844
3845 inline struct page *extent_buffer_page(struct extent_buffer *eb,
3846                                               unsigned long i)
3847 {
3848         return eb->pages[i];
3849 }
3850
3851 inline unsigned long num_extent_pages(u64 start, u64 len)
3852 {
3853         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3854                 (start >> PAGE_CACHE_SHIFT);
3855 }
3856
3857 static void __free_extent_buffer(struct extent_buffer *eb)
3858 {
3859 #if LEAK_DEBUG
3860         unsigned long flags;
3861         spin_lock_irqsave(&leak_lock, flags);
3862         list_del(&eb->leak_list);
3863         spin_unlock_irqrestore(&leak_lock, flags);
3864 #endif
3865         if (eb->pages && eb->pages != eb->inline_pages)
3866                 kfree(eb->pages);
3867         kmem_cache_free(extent_buffer_cache, eb);
3868 }
3869
3870 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3871                                                    u64 start,
3872                                                    unsigned long len,
3873                                                    gfp_t mask)
3874 {
3875         struct extent_buffer *eb = NULL;
3876 #if LEAK_DEBUG
3877         unsigned long flags;
3878 #endif
3879
3880         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3881         if (eb == NULL)
3882                 return NULL;
3883         eb->start = start;
3884         eb->len = len;
3885         eb->tree = tree;
3886         rwlock_init(&eb->lock);
3887         atomic_set(&eb->write_locks, 0);
3888         atomic_set(&eb->read_locks, 0);
3889         atomic_set(&eb->blocking_readers, 0);
3890         atomic_set(&eb->blocking_writers, 0);
3891         atomic_set(&eb->spinning_readers, 0);
3892         atomic_set(&eb->spinning_writers, 0);
3893         eb->lock_nested = 0;
3894         init_waitqueue_head(&eb->write_lock_wq);
3895         init_waitqueue_head(&eb->read_lock_wq);
3896
3897 #if LEAK_DEBUG
3898         spin_lock_irqsave(&leak_lock, flags);
3899         list_add(&eb->leak_list, &buffers);
3900         spin_unlock_irqrestore(&leak_lock, flags);
3901 #endif
3902         spin_lock_init(&eb->refs_lock);
3903         atomic_set(&eb->refs, 1);
3904         atomic_set(&eb->io_pages, 0);
3905
3906         if (len > MAX_INLINE_EXTENT_BUFFER_SIZE) {
3907                 struct page **pages;
3908                 int num_pages = (len + PAGE_CACHE_SIZE - 1) >>
3909                         PAGE_CACHE_SHIFT;
3910                 pages = kzalloc(num_pages, mask);
3911                 if (!pages) {
3912                         __free_extent_buffer(eb);
3913                         return NULL;
3914                 }
3915                 eb->pages = pages;
3916         } else {
3917                 eb->pages = eb->inline_pages;
3918         }
3919
3920         return eb;
3921 }
3922
3923 static int extent_buffer_under_io(struct extent_buffer *eb)
3924 {
3925         return (atomic_read(&eb->io_pages) ||
3926                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
3927                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3928 }
3929
3930 /*
3931  * Helper for releasing extent buffer page.
3932  */
3933 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3934                                                 unsigned long start_idx)
3935 {
3936         unsigned long index;
3937         struct page *page;
3938
3939         BUG_ON(extent_buffer_under_io(eb));
3940
3941         index = num_extent_pages(eb->start, eb->len);
3942         if (start_idx >= index)
3943                 return;
3944
3945         do {
3946                 index--;
3947                 page = extent_buffer_page(eb, index);
3948                 if (page) {
3949                         spin_lock(&page->mapping->private_lock);
3950                         /*
3951                          * We do this since we'll remove the pages after we've
3952                          * removed the eb from the radix tree, so we could race
3953                          * and have this page now attached to the new eb.  So
3954                          * only clear page_private if it's still connected to
3955                          * this eb.
3956                          */
3957                         if (PagePrivate(page) &&
3958                             page->private == (unsigned long)eb) {
3959                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3960                                 BUG_ON(PageDirty(page));
3961                                 BUG_ON(PageWriteback(page));
3962                                 /*
3963                                  * We need to make sure we haven't be attached
3964                                  * to a new eb.
3965                                  */
3966                                 ClearPagePrivate(page);
3967                                 set_page_private(page, 0);
3968                                 /* One for the page private */
3969                                 page_cache_release(page);
3970                         }
3971                         spin_unlock(&page->mapping->private_lock);
3972
3973                         /* One for when we alloced the page */
3974                         page_cache_release(page);
3975                 }
3976         } while (index != start_idx);
3977 }
3978
3979 /*
3980  * Helper for releasing the extent buffer.
3981  */
3982 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3983 {
3984         btrfs_release_extent_buffer_page(eb, 0);
3985         __free_extent_buffer(eb);
3986 }
3987
3988 static void check_buffer_tree_ref(struct extent_buffer *eb)
3989 {
3990         /* the ref bit is tricky.  We have to make sure it is set
3991          * if we have the buffer dirty.   Otherwise the
3992          * code to free a buffer can end up dropping a dirty
3993          * page
3994          *
3995          * Once the ref bit is set, it won't go away while the
3996          * buffer is dirty or in writeback, and it also won't
3997          * go away while we have the reference count on the
3998          * eb bumped.
3999          *
4000          * We can't just set the ref bit without bumping the
4001          * ref on the eb because free_extent_buffer might
4002          * see the ref bit and try to clear it.  If this happens
4003          * free_extent_buffer might end up dropping our original
4004          * ref by mistake and freeing the page before we are able
4005          * to add one more ref.
4006          *
4007          * So bump the ref count first, then set the bit.  If someone
4008          * beat us to it, drop the ref we added.
4009          */
4010         if (!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4011                 atomic_inc(&eb->refs);
4012                 if (test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4013                         atomic_dec(&eb->refs);
4014         }
4015 }
4016
4017 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4018 {
4019         unsigned long num_pages, i;
4020
4021         check_buffer_tree_ref(eb);
4022
4023         num_pages = num_extent_pages(eb->start, eb->len);
4024         for (i = 0; i < num_pages; i++) {
4025                 struct page *p = extent_buffer_page(eb, i);
4026                 mark_page_accessed(p);
4027         }
4028 }
4029
4030 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
4031                                           u64 start, unsigned long len)
4032 {
4033         unsigned long num_pages = num_extent_pages(start, len);
4034         unsigned long i;
4035         unsigned long index = start >> PAGE_CACHE_SHIFT;
4036         struct extent_buffer *eb;
4037         struct extent_buffer *exists = NULL;
4038         struct page *p;
4039         struct address_space *mapping = tree->mapping;
4040         int uptodate = 1;
4041         int ret;
4042
4043         rcu_read_lock();
4044         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4045         if (eb && atomic_inc_not_zero(&eb->refs)) {
4046                 rcu_read_unlock();
4047                 mark_extent_buffer_accessed(eb);
4048                 return eb;
4049         }
4050         rcu_read_unlock();
4051
4052         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
4053         if (!eb)
4054                 return NULL;
4055
4056         for (i = 0; i < num_pages; i++, index++) {
4057                 p = find_or_create_page(mapping, index, GFP_NOFS);
4058                 if (!p) {
4059                         WARN_ON(1);
4060                         goto free_eb;
4061                 }
4062
4063                 spin_lock(&mapping->private_lock);
4064                 if (PagePrivate(p)) {
4065                         /*
4066                          * We could have already allocated an eb for this page
4067                          * and attached one so lets see if we can get a ref on
4068                          * the existing eb, and if we can we know it's good and
4069                          * we can just return that one, else we know we can just
4070                          * overwrite page->private.
4071                          */
4072                         exists = (struct extent_buffer *)p->private;
4073                         if (atomic_inc_not_zero(&exists->refs)) {
4074                                 spin_unlock(&mapping->private_lock);
4075                                 unlock_page(p);
4076                                 mark_extent_buffer_accessed(exists);
4077                                 goto free_eb;
4078                         }
4079
4080                         /*
4081                          * Do this so attach doesn't complain and we need to
4082                          * drop the ref the old guy had.
4083                          */
4084                         ClearPagePrivate(p);
4085                         WARN_ON(PageDirty(p));
4086                         page_cache_release(p);
4087                 }
4088                 attach_extent_buffer_page(eb, p);
4089                 spin_unlock(&mapping->private_lock);
4090                 WARN_ON(PageDirty(p));
4091                 mark_page_accessed(p);
4092                 eb->pages[i] = p;
4093                 if (!PageUptodate(p))
4094                         uptodate = 0;
4095
4096                 /*
4097                  * see below about how we avoid a nasty race with release page
4098                  * and why we unlock later
4099                  */
4100         }
4101         if (uptodate)
4102                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4103 again:
4104         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4105         if (ret)
4106                 goto free_eb;
4107
4108         spin_lock(&tree->buffer_lock);
4109         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
4110         if (ret == -EEXIST) {
4111                 exists = radix_tree_lookup(&tree->buffer,
4112                                                 start >> PAGE_CACHE_SHIFT);
4113                 if (!atomic_inc_not_zero(&exists->refs)) {
4114                         spin_unlock(&tree->buffer_lock);
4115                         radix_tree_preload_end();
4116                         exists = NULL;
4117                         goto again;
4118                 }
4119                 spin_unlock(&tree->buffer_lock);
4120                 radix_tree_preload_end();
4121                 mark_extent_buffer_accessed(exists);
4122                 goto free_eb;
4123         }
4124         /* add one reference for the tree */
4125         spin_lock(&eb->refs_lock);
4126         check_buffer_tree_ref(eb);
4127         spin_unlock(&eb->refs_lock);
4128         spin_unlock(&tree->buffer_lock);
4129         radix_tree_preload_end();
4130
4131         /*
4132          * there is a race where release page may have
4133          * tried to find this extent buffer in the radix
4134          * but failed.  It will tell the VM it is safe to
4135          * reclaim the, and it will clear the page private bit.
4136          * We must make sure to set the page private bit properly
4137          * after the extent buffer is in the radix tree so
4138          * it doesn't get lost
4139          */
4140         SetPageChecked(eb->pages[0]);
4141         for (i = 1; i < num_pages; i++) {
4142                 p = extent_buffer_page(eb, i);
4143                 ClearPageChecked(p);
4144                 unlock_page(p);
4145         }
4146         unlock_page(eb->pages[0]);
4147         return eb;
4148
4149 free_eb:
4150         for (i = 0; i < num_pages; i++) {
4151                 if (eb->pages[i])
4152                         unlock_page(eb->pages[i]);
4153         }
4154
4155         if (!atomic_dec_and_test(&eb->refs))
4156                 return exists;
4157         btrfs_release_extent_buffer(eb);
4158         return exists;
4159 }
4160
4161 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
4162                                          u64 start, unsigned long len)
4163 {
4164         struct extent_buffer *eb;
4165
4166         rcu_read_lock();
4167         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4168         if (eb && atomic_inc_not_zero(&eb->refs)) {
4169                 rcu_read_unlock();
4170                 mark_extent_buffer_accessed(eb);
4171                 return eb;
4172         }
4173         rcu_read_unlock();
4174
4175         return NULL;
4176 }
4177
4178 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4179 {
4180         struct extent_buffer *eb =
4181                         container_of(head, struct extent_buffer, rcu_head);
4182
4183         __free_extent_buffer(eb);
4184 }
4185
4186 /* Expects to have eb->eb_lock already held */
4187 static void release_extent_buffer(struct extent_buffer *eb, gfp_t mask)
4188 {
4189         WARN_ON(atomic_read(&eb->refs) == 0);
4190         if (atomic_dec_and_test(&eb->refs)) {
4191                 struct extent_io_tree *tree = eb->tree;
4192
4193                 spin_unlock(&eb->refs_lock);
4194
4195                 spin_lock(&tree->buffer_lock);
4196                 radix_tree_delete(&tree->buffer,
4197                                   eb->start >> PAGE_CACHE_SHIFT);
4198                 spin_unlock(&tree->buffer_lock);
4199
4200                 /* Should be safe to release our pages at this point */
4201                 btrfs_release_extent_buffer_page(eb, 0);
4202
4203                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4204                 return;
4205         }
4206         spin_unlock(&eb->refs_lock);
4207 }
4208
4209 void free_extent_buffer(struct extent_buffer *eb)
4210 {
4211         if (!eb)
4212                 return;
4213
4214         spin_lock(&eb->refs_lock);
4215         if (atomic_read(&eb->refs) == 2 &&
4216             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4217             !extent_buffer_under_io(eb) &&
4218             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4219                 atomic_dec(&eb->refs);
4220
4221         /*
4222          * I know this is terrible, but it's temporary until we stop tracking
4223          * the uptodate bits and such for the extent buffers.
4224          */
4225         release_extent_buffer(eb, GFP_ATOMIC);
4226 }
4227
4228 void free_extent_buffer_stale(struct extent_buffer *eb)
4229 {
4230         if (!eb)
4231                 return;
4232
4233         spin_lock(&eb->refs_lock);
4234         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4235
4236         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4237             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4238                 atomic_dec(&eb->refs);
4239         release_extent_buffer(eb, GFP_NOFS);
4240 }
4241
4242 int clear_extent_buffer_dirty(struct extent_buffer *eb)
4243 {
4244         unsigned long i;
4245         unsigned long num_pages;
4246         struct page *page;
4247
4248         num_pages = num_extent_pages(eb->start, eb->len);
4249         WARN_ON(atomic_read(&eb->refs) == 0);
4250
4251         for (i = 0; i < num_pages; i++) {
4252                 page = extent_buffer_page(eb, i);
4253                 if (!PageDirty(page))
4254                         continue;
4255
4256                 lock_page(page);
4257                 WARN_ON(!PagePrivate(page));
4258
4259                 clear_page_dirty_for_io(page);
4260                 spin_lock_irq(&page->mapping->tree_lock);
4261                 if (!PageDirty(page)) {
4262                         radix_tree_tag_clear(&page->mapping->page_tree,
4263                                                 page_index(page),
4264                                                 PAGECACHE_TAG_DIRTY);
4265                 }
4266                 spin_unlock_irq(&page->mapping->tree_lock);
4267                 ClearPageError(page);
4268                 unlock_page(page);
4269         }
4270         WARN_ON(atomic_read(&eb->refs) == 0);
4271         return 0;
4272 }
4273
4274 int set_extent_buffer_dirty(struct extent_buffer *eb)
4275 {
4276         unsigned long i;
4277         unsigned long num_pages;
4278         int was_dirty = 0;
4279
4280         check_buffer_tree_ref(eb);
4281
4282         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4283
4284         num_pages = num_extent_pages(eb->start, eb->len);
4285         WARN_ON(atomic_read(&eb->refs) == 0);
4286         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4287
4288         for (i = 0; i < num_pages; i++)
4289                 set_page_dirty(extent_buffer_page(eb, i));
4290         return was_dirty;
4291 }
4292
4293 static int range_straddles_pages(u64 start, u64 len)
4294 {
4295         if (len < PAGE_CACHE_SIZE)
4296                 return 1;
4297         if (start & (PAGE_CACHE_SIZE - 1))
4298                 return 1;
4299         if ((start + len) & (PAGE_CACHE_SIZE - 1))
4300                 return 1;
4301         return 0;
4302 }
4303
4304 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4305 {
4306         unsigned long i;
4307         struct page *page;
4308         unsigned long num_pages;
4309
4310         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4311         num_pages = num_extent_pages(eb->start, eb->len);
4312         for (i = 0; i < num_pages; i++) {
4313                 page = extent_buffer_page(eb, i);
4314                 if (page)
4315                         ClearPageUptodate(page);
4316         }
4317         return 0;
4318 }
4319
4320 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4321 {
4322         unsigned long i;
4323         struct page *page;
4324         unsigned long num_pages;
4325
4326         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4327         num_pages = num_extent_pages(eb->start, eb->len);
4328         for (i = 0; i < num_pages; i++) {
4329                 page = extent_buffer_page(eb, i);
4330                 SetPageUptodate(page);
4331         }
4332         return 0;
4333 }
4334
4335 int extent_range_uptodate(struct extent_io_tree *tree,
4336                           u64 start, u64 end)
4337 {
4338         struct page *page;
4339         int ret;
4340         int pg_uptodate = 1;
4341         int uptodate;
4342         unsigned long index;
4343
4344         if (range_straddles_pages(start, end - start + 1)) {
4345                 ret = test_range_bit(tree, start, end,
4346                                      EXTENT_UPTODATE, 1, NULL);
4347                 if (ret)
4348                         return 1;
4349         }
4350         while (start <= end) {
4351                 index = start >> PAGE_CACHE_SHIFT;
4352                 page = find_get_page(tree->mapping, index);
4353                 if (!page)
4354                         return 1;
4355                 uptodate = PageUptodate(page);
4356                 page_cache_release(page);
4357                 if (!uptodate) {
4358                         pg_uptodate = 0;
4359                         break;
4360                 }
4361                 start += PAGE_CACHE_SIZE;
4362         }
4363         return pg_uptodate;
4364 }
4365
4366 int extent_buffer_uptodate(struct extent_buffer *eb)
4367 {
4368         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4369 }
4370
4371 int read_extent_buffer_pages(struct extent_io_tree *tree,
4372                              struct extent_buffer *eb, u64 start, int wait,
4373                              get_extent_t *get_extent, int mirror_num)
4374 {
4375         unsigned long i;
4376         unsigned long start_i;
4377         struct page *page;
4378         int err;
4379         int ret = 0;
4380         int locked_pages = 0;
4381         int all_uptodate = 1;
4382         unsigned long num_pages;
4383         unsigned long num_reads = 0;
4384         struct bio *bio = NULL;
4385         unsigned long bio_flags = 0;
4386
4387         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4388                 return 0;
4389
4390         if (start) {
4391                 WARN_ON(start < eb->start);
4392                 start_i = (start >> PAGE_CACHE_SHIFT) -
4393                         (eb->start >> PAGE_CACHE_SHIFT);
4394         } else {
4395                 start_i = 0;
4396         }
4397
4398         num_pages = num_extent_pages(eb->start, eb->len);
4399         for (i = start_i; i < num_pages; i++) {
4400                 page = extent_buffer_page(eb, i);
4401                 if (wait == WAIT_NONE) {
4402                         if (!trylock_page(page))
4403                                 goto unlock_exit;
4404                 } else {
4405                         lock_page(page);
4406                 }
4407                 locked_pages++;
4408                 if (!PageUptodate(page)) {
4409                         num_reads++;
4410                         all_uptodate = 0;
4411                 }
4412         }
4413         if (all_uptodate) {
4414                 if (start_i == 0)
4415                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4416                 goto unlock_exit;
4417         }
4418
4419         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4420         eb->failed_mirror = 0;
4421         atomic_set(&eb->io_pages, num_reads);
4422         for (i = start_i; i < num_pages; i++) {
4423                 page = extent_buffer_page(eb, i);
4424                 if (!PageUptodate(page)) {
4425                         ClearPageError(page);
4426                         err = __extent_read_full_page(tree, page,
4427                                                       get_extent, &bio,
4428                                                       mirror_num, &bio_flags);
4429                         if (err)
4430                                 ret = err;
4431                 } else {
4432                         unlock_page(page);
4433                 }
4434         }
4435
4436         if (bio)
4437                 submit_one_bio(READ, bio, mirror_num, bio_flags);
4438
4439         if (ret || wait != WAIT_COMPLETE)
4440                 return ret;
4441
4442         for (i = start_i; i < num_pages; i++) {
4443                 page = extent_buffer_page(eb, i);
4444                 wait_on_page_locked(page);
4445                 if (!PageUptodate(page))
4446                         ret = -EIO;
4447         }
4448
4449         return ret;
4450
4451 unlock_exit:
4452         i = start_i;
4453         while (locked_pages > 0) {
4454                 page = extent_buffer_page(eb, i);
4455                 i++;
4456                 unlock_page(page);
4457                 locked_pages--;
4458         }
4459         return ret;
4460 }
4461
4462 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4463                         unsigned long start,
4464                         unsigned long len)
4465 {
4466         size_t cur;
4467         size_t offset;
4468         struct page *page;
4469         char *kaddr;
4470         char *dst = (char *)dstv;
4471         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4472         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4473
4474         WARN_ON(start > eb->len);
4475         WARN_ON(start + len > eb->start + eb->len);
4476
4477         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4478
4479         while (len > 0) {
4480                 page = extent_buffer_page(eb, i);
4481
4482                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4483                 kaddr = page_address(page);
4484                 memcpy(dst, kaddr + offset, cur);
4485
4486                 dst += cur;
4487                 len -= cur;
4488                 offset = 0;
4489                 i++;
4490         }
4491 }
4492
4493 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4494                                unsigned long min_len, char **map,
4495                                unsigned long *map_start,
4496                                unsigned long *map_len)
4497 {
4498         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4499         char *kaddr;
4500         struct page *p;
4501         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4502         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4503         unsigned long end_i = (start_offset + start + min_len - 1) >>
4504                 PAGE_CACHE_SHIFT;
4505
4506         if (i != end_i)
4507                 return -EINVAL;
4508
4509         if (i == 0) {
4510                 offset = start_offset;
4511                 *map_start = 0;
4512         } else {
4513                 offset = 0;
4514                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4515         }
4516
4517         if (start + min_len > eb->len) {
4518                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4519                        "wanted %lu %lu\n", (unsigned long long)eb->start,
4520                        eb->len, start, min_len);
4521                 WARN_ON(1);
4522                 return -EINVAL;
4523         }
4524
4525         p = extent_buffer_page(eb, i);
4526         kaddr = page_address(p);
4527         *map = kaddr + offset;
4528         *map_len = PAGE_CACHE_SIZE - offset;
4529         return 0;
4530 }
4531
4532 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4533                           unsigned long start,
4534                           unsigned long len)
4535 {
4536         size_t cur;
4537         size_t offset;
4538         struct page *page;
4539         char *kaddr;
4540         char *ptr = (char *)ptrv;
4541         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4542         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4543         int ret = 0;
4544
4545         WARN_ON(start > eb->len);
4546         WARN_ON(start + len > eb->start + eb->len);
4547
4548         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4549
4550         while (len > 0) {
4551                 page = extent_buffer_page(eb, i);
4552
4553                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4554
4555                 kaddr = page_address(page);
4556                 ret = memcmp(ptr, kaddr + offset, cur);
4557                 if (ret)
4558                         break;
4559
4560                 ptr += cur;
4561                 len -= cur;
4562                 offset = 0;
4563                 i++;
4564         }
4565         return ret;
4566 }
4567
4568 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4569                          unsigned long start, unsigned long len)
4570 {
4571         size_t cur;
4572         size_t offset;
4573         struct page *page;
4574         char *kaddr;
4575         char *src = (char *)srcv;
4576         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4577         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4578
4579         WARN_ON(start > eb->len);
4580         WARN_ON(start + len > eb->start + eb->len);
4581
4582         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4583
4584         while (len > 0) {
4585                 page = extent_buffer_page(eb, i);
4586                 WARN_ON(!PageUptodate(page));
4587
4588                 cur = min(len, PAGE_CACHE_SIZE - offset);
4589                 kaddr = page_address(page);
4590                 memcpy(kaddr + offset, src, cur);
4591
4592                 src += cur;
4593                 len -= cur;
4594                 offset = 0;
4595                 i++;
4596         }
4597 }
4598
4599 void memset_extent_buffer(struct extent_buffer *eb, char c,
4600                           unsigned long start, unsigned long len)
4601 {
4602         size_t cur;
4603         size_t offset;
4604         struct page *page;
4605         char *kaddr;
4606         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4607         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4608
4609         WARN_ON(start > eb->len);
4610         WARN_ON(start + len > eb->start + eb->len);
4611
4612         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4613
4614         while (len > 0) {
4615                 page = extent_buffer_page(eb, i);
4616                 WARN_ON(!PageUptodate(page));
4617
4618                 cur = min(len, PAGE_CACHE_SIZE - offset);
4619                 kaddr = page_address(page);
4620                 memset(kaddr + offset, c, cur);
4621
4622                 len -= cur;
4623                 offset = 0;
4624                 i++;
4625         }
4626 }
4627
4628 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
4629                         unsigned long dst_offset, unsigned long src_offset,
4630                         unsigned long len)
4631 {
4632         u64 dst_len = dst->len;
4633         size_t cur;
4634         size_t offset;
4635         struct page *page;
4636         char *kaddr;
4637         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4638         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4639
4640         WARN_ON(src->len != dst_len);
4641
4642         offset = (start_offset + dst_offset) &
4643                 ((unsigned long)PAGE_CACHE_SIZE - 1);
4644
4645         while (len > 0) {
4646                 page = extent_buffer_page(dst, i);
4647                 WARN_ON(!PageUptodate(page));
4648
4649                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
4650
4651                 kaddr = page_address(page);
4652                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4653
4654                 src_offset += cur;
4655                 len -= cur;
4656                 offset = 0;
4657                 i++;
4658         }
4659 }
4660
4661 static void move_pages(struct page *dst_page, struct page *src_page,
4662                        unsigned long dst_off, unsigned long src_off,
4663                        unsigned long len)
4664 {
4665         char *dst_kaddr = page_address(dst_page);
4666         if (dst_page == src_page) {
4667                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
4668         } else {
4669                 char *src_kaddr = page_address(src_page);
4670                 char *p = dst_kaddr + dst_off + len;
4671                 char *s = src_kaddr + src_off + len;
4672
4673                 while (len--)
4674                         *--p = *--s;
4675         }
4676 }
4677
4678 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4679 {
4680         unsigned long distance = (src > dst) ? src - dst : dst - src;
4681         return distance < len;
4682 }
4683
4684 static void copy_pages(struct page *dst_page, struct page *src_page,
4685                        unsigned long dst_off, unsigned long src_off,
4686                        unsigned long len)
4687 {
4688         char *dst_kaddr = page_address(dst_page);
4689         char *src_kaddr;
4690         int must_memmove = 0;
4691
4692         if (dst_page != src_page) {
4693                 src_kaddr = page_address(src_page);
4694         } else {
4695                 src_kaddr = dst_kaddr;
4696                 if (areas_overlap(src_off, dst_off, len))
4697                         must_memmove = 1;
4698         }
4699
4700         if (must_memmove)
4701                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
4702         else
4703                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
4704 }
4705
4706 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4707                            unsigned long src_offset, unsigned long len)
4708 {
4709         size_t cur;
4710         size_t dst_off_in_page;
4711         size_t src_off_in_page;
4712         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4713         unsigned long dst_i;
4714         unsigned long src_i;
4715
4716         if (src_offset + len > dst->len) {
4717                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4718                        "len %lu dst len %lu\n", src_offset, len, dst->len);
4719                 BUG_ON(1);
4720         }
4721         if (dst_offset + len > dst->len) {
4722                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4723                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
4724                 BUG_ON(1);
4725         }
4726
4727         while (len > 0) {
4728                 dst_off_in_page = (start_offset + dst_offset) &
4729                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4730                 src_off_in_page = (start_offset + src_offset) &
4731                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4732
4733                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4734                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
4735
4736                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
4737                                                src_off_in_page));
4738                 cur = min_t(unsigned long, cur,
4739                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
4740
4741                 copy_pages(extent_buffer_page(dst, dst_i),
4742                            extent_buffer_page(dst, src_i),
4743                            dst_off_in_page, src_off_in_page, cur);
4744
4745                 src_offset += cur;
4746                 dst_offset += cur;
4747                 len -= cur;
4748         }
4749 }
4750
4751 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4752                            unsigned long src_offset, unsigned long len)
4753 {
4754         size_t cur;
4755         size_t dst_off_in_page;
4756         size_t src_off_in_page;
4757         unsigned long dst_end = dst_offset + len - 1;
4758         unsigned long src_end = src_offset + len - 1;
4759         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4760         unsigned long dst_i;
4761         unsigned long src_i;
4762
4763         if (src_offset + len > dst->len) {
4764                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4765                        "len %lu len %lu\n", src_offset, len, dst->len);
4766                 BUG_ON(1);
4767         }
4768         if (dst_offset + len > dst->len) {
4769                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4770                        "len %lu len %lu\n", dst_offset, len, dst->len);
4771                 BUG_ON(1);
4772         }
4773         if (dst_offset < src_offset) {
4774                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4775                 return;
4776         }
4777         while (len > 0) {
4778                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4779                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4780
4781                 dst_off_in_page = (start_offset + dst_end) &
4782                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4783                 src_off_in_page = (start_offset + src_end) &
4784                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4785
4786                 cur = min_t(unsigned long, len, src_off_in_page + 1);
4787                 cur = min(cur, dst_off_in_page + 1);
4788                 move_pages(extent_buffer_page(dst, dst_i),
4789                            extent_buffer_page(dst, src_i),
4790                            dst_off_in_page - cur + 1,
4791                            src_off_in_page - cur + 1, cur);
4792
4793                 dst_end -= cur;
4794                 src_end -= cur;
4795                 len -= cur;
4796         }
4797 }
4798
4799 int try_release_extent_buffer(struct page *page, gfp_t mask)
4800 {
4801         struct extent_buffer *eb;
4802
4803         /*
4804          * We need to make sure noboody is attaching this page to an eb right
4805          * now.
4806          */
4807         spin_lock(&page->mapping->private_lock);
4808         if (!PagePrivate(page)) {
4809                 spin_unlock(&page->mapping->private_lock);
4810                 return 1;
4811         }
4812
4813         eb = (struct extent_buffer *)page->private;
4814         BUG_ON(!eb);
4815
4816         /*
4817          * This is a little awful but should be ok, we need to make sure that
4818          * the eb doesn't disappear out from under us while we're looking at
4819          * this page.
4820          */
4821         spin_lock(&eb->refs_lock);
4822         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4823                 spin_unlock(&eb->refs_lock);
4824                 spin_unlock(&page->mapping->private_lock);
4825                 return 0;
4826         }
4827         spin_unlock(&page->mapping->private_lock);
4828
4829         if ((mask & GFP_NOFS) == GFP_NOFS)
4830                 mask = GFP_NOFS;
4831
4832         /*
4833          * If tree ref isn't set then we know the ref on this eb is a real ref,
4834          * so just return, this page will likely be freed soon anyway.
4835          */
4836         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4837                 spin_unlock(&eb->refs_lock);
4838                 return 0;
4839         }
4840         release_extent_buffer(eb, mask);
4841
4842         return 1;
4843 }