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 static inline struct btrfs_fs_info *
  59 tree_fs_info(struct extent_io_tree *tree)
  60 {
  61         return btrfs_sb(tree->mapping->host->i_sb);
  62 }
  63
  64 int __init extent_io_init(void)
  65 {
  66         extent_state_cache = kmem_cache_create("extent_state",
  67                         sizeof(struct extent_state), 0,
  68                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  69         if (!extent_state_cache)
  70                 return -ENOMEM;
  71
  72         extent_buffer_cache = kmem_cache_create("extent_buffers",
  73                         sizeof(struct extent_buffer), 0,
  74                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
  75         if (!extent_buffer_cache)
  76                 goto free_state_cache;
  77         return 0;
  78
  79 free_state_cache:
  80         kmem_cache_destroy(extent_state_cache);
  81         return -ENOMEM;
  82 }
  83
  84 void extent_io_exit(void)
  85 {
  86         struct extent_state *state;
  87         struct extent_buffer *eb;
  88
  89         while (!list_empty(&states)) {
  90                 state = list_entry(states.next, struct extent_state, leak_list);
  91                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
  92                        "state %lu in tree %p refs %d\n",
  93                        (unsigned long long)state->start,
  94                        (unsigned long long)state->end,
  95                        state->state, state->tree, atomic_read(&state->refs));
  96                 list_del(&state->leak_list);
  97                 kmem_cache_free(extent_state_cache, state);
  98
  99         }
 100
 101         while (!list_empty(&buffers)) {
 102                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
 103                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
 104                        "refs %d\n", (unsigned long long)eb->start,
 105                        eb->len, atomic_read(&eb->refs));
 106                 list_del(&eb->leak_list);
 107                 kmem_cache_free(extent_buffer_cache, eb);
 108         }
 109         if (extent_state_cache)
 110                 kmem_cache_destroy(extent_state_cache);
 111         if (extent_buffer_cache)
 112                 kmem_cache_destroy(extent_buffer_cache);
 113 }
 114
 115 void extent_io_tree_init(struct extent_io_tree *tree,
 116                          struct address_space *mapping)
 117 {
 118         tree->state = RB_ROOT;
 119         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
 120         tree->ops = NULL;
 121         tree->dirty_bytes = 0;
 122         spin_lock_init(&tree->lock);
 123         spin_lock_init(&tree->buffer_lock);
 124         tree->mapping = mapping;
 125 }
 126
 127 static struct extent_state *alloc_extent_state(gfp_t mask)
 128 {
 129         struct extent_state *state;
 130 #if LEAK_DEBUG
 131         unsigned long flags;
 132 #endif
 133
 134         state = kmem_cache_alloc(extent_state_cache, mask);
 135         if (!state)
 136                 return state;
 137         state->state = 0;
 138         state->private = 0;
 139         state->tree = NULL;
 140 #if LEAK_DEBUG
 141         spin_lock_irqsave(&leak_lock, flags);
 142         list_add(&state->leak_list, &states);
 143         spin_unlock_irqrestore(&leak_lock, flags);
 144 #endif
 145         atomic_set(&state->refs, 1);
 146         init_waitqueue_head(&state->wq);
 147         trace_alloc_extent_state(state, mask, _RET_IP_);
 148         return state;
 149 }
 150
 151 void free_extent_state(struct extent_state *state)
 152 {
 153         if (!state)
 154                 return;
 155         if (atomic_dec_and_test(&state->refs)) {
 156 #if LEAK_DEBUG
 157                 unsigned long flags;
 158 #endif
 159                 WARN_ON(state->tree);
 160 #if LEAK_DEBUG
 161                 spin_lock_irqsave(&leak_lock, flags);
 162                 list_del(&state->leak_list);
 163                 spin_unlock_irqrestore(&leak_lock, flags);
 164 #endif
 165                 trace_free_extent_state(state, _RET_IP_);
 166                 kmem_cache_free(extent_state_cache, state);
 167         }
 168 }
 169
 170 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 171                                    struct rb_node *node)
 172 {
 173         struct rb_node **p = &root->rb_node;
 174         struct rb_node *parent = NULL;
 175         struct tree_entry *entry;
 176
 177         while (*p) {
 178                 parent = *p;
 179                 entry = rb_entry(parent, struct tree_entry, rb_node);
 180
 181                 if (offset < entry->start)
 182                         p = &(*p)->rb_left;
 183                 else if (offset > entry->end)
 184                         p = &(*p)->rb_right;
 185                 else
 186                         return parent;
 187         }
 188
 189         entry = rb_entry(node, struct tree_entry, rb_node);
 190         rb_link_node(node, parent, p);
 191         rb_insert_color(node, root);
 192         return NULL;
 193 }
 194
 195 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 196                                      struct rb_node **prev_ret,
 197                                      struct rb_node **next_ret)
 198 {
 199         struct rb_root *root = &tree->state;
 200         struct rb_node *n = root->rb_node;
 201         struct rb_node *prev = NULL;
 202         struct rb_node *orig_prev = NULL;
 203         struct tree_entry *entry;
 204         struct tree_entry *prev_entry = NULL;
 205
 206         while (n) {
 207                 entry = rb_entry(n, struct tree_entry, rb_node);
 208                 prev = n;
 209                 prev_entry = entry;
 210
 211                 if (offset < entry->start)
 212                         n = n->rb_left;
 213                 else if (offset > entry->end)
 214                         n = n->rb_right;
 215                 else
 216                         return n;
 217         }
 218
 219         if (prev_ret) {
 220                 orig_prev = prev;
 221                 while (prev && offset > prev_entry->end) {
 222                         prev = rb_next(prev);
 223                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 224                 }
 225                 *prev_ret = prev;
 226                 prev = orig_prev;
 227         }
 228
 229         if (next_ret) {
 230                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 231                 while (prev && offset < prev_entry->start) {
 232                         prev = rb_prev(prev);
 233                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 234                 }
 235                 *next_ret = prev;
 236         }
 237         return NULL;
 238 }
 239
 240 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 241                                           u64 offset)
 242 {
 243         struct rb_node *prev = NULL;
 244         struct rb_node *ret;
 245
 246         ret = __etree_search(tree, offset, &prev, NULL);
 247         if (!ret)
 248                 return prev;
 249         return ret;
 250 }
 251
 252 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 253                      struct extent_state *other)
 254 {
 255         if (tree->ops && tree->ops->merge_extent_hook)
 256                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 257                                              other);
 258 }
 259
 260 /*
 261  * utility function to look for merge candidates inside a given range.
 262  * Any extents with matching state are merged together into a single
 263  * extent in the tree.  Extents with EXTENT_IO in their state field
 264  * are not merged because the end_io handlers need to be able to do
 265  * operations on them without sleeping (or doing allocations/splits).
 266  *
 267  * This should be called with the tree lock held.
 268  */
 269 static void merge_state(struct extent_io_tree *tree,
 270                         struct extent_state *state)
 271 {
 272         struct extent_state *other;
 273         struct rb_node *other_node;
 274
 275         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 276                 return;
 277
 278         other_node = rb_prev(&state->rb_node);
 279         if (other_node) {
 280                 other = rb_entry(other_node, struct extent_state, rb_node);
 281                 if (other->end == state->start - 1 &&
 282                     other->state == state->state) {
 283                         merge_cb(tree, state, other);
 284                         state->start = other->start;
 285                         other->tree = NULL;
 286                         rb_erase(&other->rb_node, &tree->state);
 287                         free_extent_state(other);
 288                 }
 289         }
 290         other_node = rb_next(&state->rb_node);
 291         if (other_node) {
 292                 other = rb_entry(other_node, struct extent_state, rb_node);
 293                 if (other->start == state->end + 1 &&
 294                     other->state == state->state) {
 295                         merge_cb(tree, state, other);
 296                         state->end = other->end;
 297                         other->tree = NULL;
 298                         rb_erase(&other->rb_node, &tree->state);
 299                         free_extent_state(other);
 300                 }
 301         }
 302 }
 303
 304 static void set_state_cb(struct extent_io_tree *tree,
 305                          struct extent_state *state, int *bits)
 306 {
 307         if (tree->ops && tree->ops->set_bit_hook)
 308                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 309 }
 310
 311 static void clear_state_cb(struct extent_io_tree *tree,
 312                            struct extent_state *state, int *bits)
 313 {
 314         if (tree->ops && tree->ops->clear_bit_hook)
 315                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 316 }
 317
 318 static void set_state_bits(struct extent_io_tree *tree,
 319                            struct extent_state *state, int *bits);
 320
 321 /*
 322  * insert an extent_state struct into the tree.  'bits' are set on the
 323  * struct before it is inserted.
 324  *
 325  * This may return -EEXIST if the extent is already there, in which case the
 326  * state struct is freed.
 327  *
 328  * The tree lock is not taken internally.  This is a utility function and
 329  * probably isn't what you want to call (see set/clear_extent_bit).
 330  */
 331 static int insert_state(struct extent_io_tree *tree,
 332                         struct extent_state *state, u64 start, u64 end,
 333                         int *bits)
 334 {
 335         struct rb_node *node;
 336
 337         if (end < start) {
 338                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
 339                        (unsigned long long)end,
 340                        (unsigned long long)start);
 341                 WARN_ON(1);
 342         }
 343         state->start = start;
 344         state->end = end;
 345
 346         set_state_bits(tree, state, bits);
 347
 348         node = tree_insert(&tree->state, end, &state->rb_node);
 349         if (node) {
 350                 struct extent_state *found;
 351                 found = rb_entry(node, struct extent_state, rb_node);
 352                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
 353                        "%llu %llu\n", (unsigned long long)found->start,
 354                        (unsigned long long)found->end,
 355                        (unsigned long long)start, (unsigned long long)end);
 356                 return -EEXIST;
 357         }
 358         state->tree = tree;
 359         merge_state(tree, state);
 360         return 0;
 361 }
 362
 363 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 364                      u64 split)
 365 {
 366         if (tree->ops && tree->ops->split_extent_hook)
 367                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 368 }
 369
 370 /*
 371  * split a given extent state struct in two, inserting the preallocated
 372  * struct 'prealloc' as the newly created second half.  'split' indicates an
 373  * offset inside 'orig' where it should be split.
 374  *
 375  * Before calling,
 376  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 377  * are two extent state structs in the tree:
 378  * prealloc: [orig->start, split - 1]
 379  * orig: [ split, orig->end ]
 380  *
 381  * The tree locks are not taken by this function. They need to be held
 382  * by the caller.
 383  */
 384 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 385                        struct extent_state *prealloc, u64 split)
 386 {
 387         struct rb_node *node;
 388
 389         split_cb(tree, orig, split);
 390
 391         prealloc->start = orig->start;
 392         prealloc->end = split - 1;
 393         prealloc->state = orig->state;
 394         orig->start = split;
 395
 396         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
 397         if (node) {
 398                 free_extent_state(prealloc);
 399                 return -EEXIST;
 400         }
 401         prealloc->tree = tree;
 402         return 0;
 403 }
 404
 405 static struct extent_state *next_state(struct extent_state *state)
 406 {
 407         struct rb_node *next = rb_next(&state->rb_node);
 408         if (next)
 409                 return rb_entry(next, struct extent_state, rb_node);
 410         else
 411                 return NULL;
 412 }
 413
 414 /*
 415  * utility function to clear some bits in an extent state struct.
 416  * it will optionally wake up any one waiting on this state (wake == 1)
 417  *
 418  * If no bits are set on the state struct after clearing things, the
 419  * struct is freed and removed from the tree
 420  */
 421 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 422                                             struct extent_state *state,
 423                                             int *bits, int wake)
 424 {
 425         struct extent_state *next;
 426         int bits_to_clear = *bits & ~EXTENT_CTLBITS;
 427
 428         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 429                 u64 range = state->end - state->start + 1;
 430                 WARN_ON(range > tree->dirty_bytes);
 431                 tree->dirty_bytes -= range;
 432         }
 433         clear_state_cb(tree, state, bits);
 434         state->state &= ~bits_to_clear;
 435         if (wake)
 436                 wake_up(&state->wq);
 437         if (state->state == 0) {
 438                 next = next_state(state);
 439                 if (state->tree) {
 440                         rb_erase(&state->rb_node, &tree->state);
 441                         state->tree = NULL;
 442                         free_extent_state(state);
 443                 } else {
 444                         WARN_ON(1);
 445                 }
 446         } else {
 447                 merge_state(tree, state);
 448                 next = next_state(state);
 449         }
 450         return next;
 451 }
 452
 453 static struct extent_state *
 454 alloc_extent_state_atomic(struct extent_state *prealloc)
 455 {
 456         if (!prealloc)
 457                 prealloc = alloc_extent_state(GFP_ATOMIC);
 458
 459         return prealloc;
 460 }
 461
 462 void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 463 {
 464         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 465                     "Extent tree was modified by another "
 466                     "thread while locked.");
 467 }
 468
 469 /*
 470  * clear some bits on a range in the tree.  This may require splitting
 471  * or inserting elements in the tree, so the gfp mask is used to
 472  * indicate which allocations or sleeping are allowed.
 473  *
 474  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 475  * the given range from the tree regardless of state (ie for truncate).
 476  *
 477  * the range [start, end] is inclusive.
 478  *
 479  * This takes the tree lock, and returns 0 on success and < 0 on error.
 480  */
 481 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 482                      int bits, int wake, int delete,
 483                      struct extent_state **cached_state,
 484                      gfp_t mask)
 485 {
 486         struct extent_state *state;
 487         struct extent_state *cached;
 488         struct extent_state *prealloc = NULL;
 489         struct rb_node *node;
 490         u64 last_end;
 491         int err;
 492         int clear = 0;
 493
 494         if (delete)
 495                 bits |= ~EXTENT_CTLBITS;
 496         bits |= EXTENT_FIRST_DELALLOC;
 497
 498         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 499                 clear = 1;
 500 again:
 501         if (!prealloc && (mask & __GFP_WAIT)) {
 502                 prealloc = alloc_extent_state(mask);
 503                 if (!prealloc)
 504                         return -ENOMEM;
 505         }
 506
 507         spin_lock(&tree->lock);
 508         if (cached_state) {
 509                 cached = *cached_state;
 510
 511                 if (clear) {
 512                         *cached_state = NULL;
 513                         cached_state = NULL;
 514                 }
 515
 516                 if (cached && cached->tree && cached->start <= start &&
 517                     cached->end > start) {
 518                         if (clear)
 519                                 atomic_dec(&cached->refs);
 520                         state = cached;
 521                         goto hit_next;
 522                 }
 523                 if (clear)
 524                         free_extent_state(cached);
 525         }
 526         /*
 527          * this search will find the extents that end after
 528          * our range starts
 529          */
 530         node = tree_search(tree, start);
 531         if (!node)
 532                 goto out;
 533         state = rb_entry(node, struct extent_state, rb_node);
 534 hit_next:
 535         if (state->start > end)
 536                 goto out;
 537         WARN_ON(state->end < start);
 538         last_end = state->end;
 539
 540         /* the state doesn't have the wanted bits, go ahead */
 541         if (!(state->state & bits)) {
 542                 state = next_state(state);
 543                 goto next;
 544         }
 545
 546         /*
 547          *     | ---- desired range ---- |
 548          *  | state | or
 549          *  | ------------- state -------------- |
 550          *
 551          * We need to split the extent we found, and may flip
 552          * bits on second half.
 553          *
 554          * If the extent we found extends past our range, we
 555          * just split and search again.  It'll get split again
 556          * the next time though.
 557          *
 558          * If the extent we found is inside our range, we clear
 559          * the desired bit on it.
 560          */
 561
 562         if (state->start < start) {
 563                 prealloc = alloc_extent_state_atomic(prealloc);
 564                 BUG_ON(!prealloc);
 565                 err = split_state(tree, state, prealloc, start);
 566                 if (err)
 567                         extent_io_tree_panic(tree, err);
 568
 569                 prealloc = NULL;
 570                 if (err)
 571                         goto out;
 572                 if (state->end <= end) {
 573                         clear_state_bit(tree, state, &bits, wake);
 574                         if (last_end == (u64)-1)
 575                                 goto out;
 576                         start = last_end + 1;
 577                 }
 578                 goto search_again;
 579         }
 580         /*
 581          * | ---- desired range ---- |
 582          *                        | state |
 583          * We need to split the extent, and clear the bit
 584          * on the first half
 585          */
 586         if (state->start <= end && state->end > end) {
 587                 prealloc = alloc_extent_state_atomic(prealloc);
 588                 BUG_ON(!prealloc);
 589                 err = split_state(tree, state, prealloc, end + 1);
 590                 if (err)
 591                         extent_io_tree_panic(tree, err);
 592
 593                 if (wake)
 594                         wake_up(&state->wq);
 595
 596                 clear_state_bit(tree, prealloc, &bits, wake);
 597
 598                 prealloc = NULL;
 599                 goto out;
 600         }
 601
 602         state = clear_state_bit(tree, state, &bits, wake);
 603 next:
 604         if (last_end == (u64)-1)
 605                 goto out;
 606         start = last_end + 1;
 607         if (start <= end && state && !need_resched())
 608                 goto hit_next;
 609         goto search_again;
 610
 611 out:
 612         spin_unlock(&tree->lock);
 613         if (prealloc)
 614                 free_extent_state(prealloc);
 615
 616         return 0;
 617
 618 search_again:
 619         if (start > end)
 620                 goto out;
 621         spin_unlock(&tree->lock);
 622         if (mask & __GFP_WAIT)
 623                 cond_resched();
 624         goto again;
 625 }
 626
 627 static void wait_on_state(struct extent_io_tree *tree,
 628                           struct extent_state *state)
 629                 __releases(tree->lock)
 630                 __acquires(tree->lock)
 631 {
 632         DEFINE_WAIT(wait);
 633         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 634         spin_unlock(&tree->lock);
 635         schedule();
 636         spin_lock(&tree->lock);
 637         finish_wait(&state->wq, &wait);
 638 }
 639
 640 /*
 641  * waits for one or more bits to clear on a range in the state tree.
 642  * The range [start, end] is inclusive.
 643  * The tree lock is taken by this function
 644  */
 645 void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
 646 {
 647         struct extent_state *state;
 648         struct rb_node *node;
 649
 650         spin_lock(&tree->lock);
 651 again:
 652         while (1) {
 653                 /*
 654                  * this search will find all the extents that end after
 655                  * our range starts
 656                  */
 657                 node = tree_search(tree, start);
 658                 if (!node)
 659                         break;
 660
 661                 state = rb_entry(node, struct extent_state, rb_node);
 662
 663                 if (state->start > end)
 664                         goto out;
 665
 666                 if (state->state & bits) {
 667                         start = state->start;
 668                         atomic_inc(&state->refs);
 669                         wait_on_state(tree, state);
 670                         free_extent_state(state);
 671                         goto again;
 672                 }
 673                 start = state->end + 1;
 674
 675                 if (start > end)
 676                         break;
 677
 678                 cond_resched_lock(&tree->lock);
 679         }
 680 out:
 681         spin_unlock(&tree->lock);
 682 }
 683
 684 static void set_state_bits(struct extent_io_tree *tree,
 685                            struct extent_state *state,
 686                            int *bits)
 687 {
 688         int bits_to_set = *bits & ~EXTENT_CTLBITS;
 689
 690         set_state_cb(tree, state, bits);
 691         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 692                 u64 range = state->end - state->start + 1;
 693                 tree->dirty_bytes += range;
 694         }
 695         state->state |= bits_to_set;
 696 }
 697
 698 static void cache_state(struct extent_state *state,
 699                         struct extent_state **cached_ptr)
 700 {
 701         if (cached_ptr && !(*cached_ptr)) {
 702                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 703                         *cached_ptr = state;
 704                         atomic_inc(&state->refs);
 705                 }
 706         }
 707 }
 708
 709 static void uncache_state(struct extent_state **cached_ptr)
 710 {
 711         if (cached_ptr && (*cached_ptr)) {
 712                 struct extent_state *state = *cached_ptr;
 713                 *cached_ptr = NULL;
 714                 free_extent_state(state);
 715         }
 716 }
 717
 718 /*
 719  * set some bits on a range in the tree.  This may require allocations or
 720  * sleeping, so the gfp mask is used to indicate what is allowed.
 721  *
 722  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 723  * part of the range already has the desired bits set.  The start of the
 724  * existing range is returned in failed_start in this case.
 725  *
 726  * [start, end] is inclusive This takes the tree lock.
 727  */
 728
 729 static int __must_check
 730 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 731                  int bits, int exclusive_bits, u64 *failed_start,
 732                  struct extent_state **cached_state, gfp_t mask)
 733 {
 734         struct extent_state *state;
 735         struct extent_state *prealloc = NULL;
 736         struct rb_node *node;
 737         int err = 0;
 738         u64 last_start;
 739         u64 last_end;
 740
 741         bits |= EXTENT_FIRST_DELALLOC;
 742 again:
 743         if (!prealloc && (mask & __GFP_WAIT)) {
 744                 prealloc = alloc_extent_state(mask);
 745                 BUG_ON(!prealloc);
 746         }
 747
 748         spin_lock(&tree->lock);
 749         if (cached_state && *cached_state) {
 750                 state = *cached_state;
 751                 if (state->start <= start && state->end > start &&
 752                     state->tree) {
 753                         node = &state->rb_node;
 754                         goto hit_next;
 755                 }
 756         }
 757         /*
 758          * this search will find all the extents that end after
 759          * our range starts.
 760          */
 761         node = tree_search(tree, start);
 762         if (!node) {
 763                 prealloc = alloc_extent_state_atomic(prealloc);
 764                 BUG_ON(!prealloc);
 765                 err = insert_state(tree, prealloc, start, end, &bits);
 766                 if (err)
 767                         extent_io_tree_panic(tree, err);
 768
 769                 prealloc = NULL;
 770                 goto out;
 771         }
 772         state = rb_entry(node, struct extent_state, rb_node);
 773 hit_next:
 774         last_start = state->start;
 775         last_end = state->end;
 776
 777         /*
 778          * | ---- desired range ---- |
 779          * | state |
 780          *
 781          * Just lock what we found and keep going
 782          */
 783         if (state->start == start && state->end <= end) {
 784                 struct rb_node *next_node;
 785                 if (state->state & exclusive_bits) {
 786                         *failed_start = state->start;
 787                         err = -EEXIST;
 788                         goto out;
 789                 }
 790
 791                 set_state_bits(tree, state, &bits);
 792
 793                 cache_state(state, cached_state);
 794                 merge_state(tree, state);
 795                 if (last_end == (u64)-1)
 796                         goto out;
 797
 798                 start = last_end + 1;
 799                 next_node = rb_next(&state->rb_node);
 800                 if (next_node && start < end && prealloc && !need_resched()) {
 801                         state = rb_entry(next_node, struct extent_state,
 802                                          rb_node);
 803                         if (state->start == start)
 804                                 goto hit_next;
 805                 }
 806                 goto search_again;
 807         }
 808
 809         /*
 810          *     | ---- desired range ---- |
 811          * | state |
 812          *   or
 813          * | ------------- state -------------- |
 814          *
 815          * We need to split the extent we found, and may flip bits on
 816          * second half.
 817          *
 818          * If the extent we found extends past our
 819          * range, we just split and search again.  It'll get split
 820          * again the next time though.
 821          *
 822          * If the extent we found is inside our range, we set the
 823          * desired bit on it.
 824          */
 825         if (state->start < start) {
 826                 if (state->state & exclusive_bits) {
 827                         *failed_start = start;
 828                         err = -EEXIST;
 829                         goto out;
 830                 }
 831
 832                 prealloc = alloc_extent_state_atomic(prealloc);
 833                 BUG_ON(!prealloc);
 834                 err = split_state(tree, state, prealloc, start);
 835                 if (err)
 836                         extent_io_tree_panic(tree, err);
 837
 838                 prealloc = NULL;
 839                 if (err)
 840                         goto out;
 841                 if (state->end <= end) {
 842                         set_state_bits(tree, state, &bits);
 843                         cache_state(state, cached_state);
 844                         merge_state(tree, state);
 845                         if (last_end == (u64)-1)
 846                                 goto out;
 847                         start = last_end + 1;
 848                 }
 849                 goto search_again;
 850         }
 851         /*
 852          * | ---- desired range ---- |
 853          *     | state | or               | state |
 854          *
 855          * There's a hole, we need to insert something in it and
 856          * ignore the extent we found.
 857          */
 858         if (state->start > start) {
 859                 u64 this_end;
 860                 if (end < last_start)
 861                         this_end = end;
 862                 else
 863                         this_end = last_start - 1;
 864
 865                 prealloc = alloc_extent_state_atomic(prealloc);
 866                 BUG_ON(!prealloc);
 867
 868                 /*
 869                  * Avoid to free 'prealloc' if it can be merged with
 870                  * the later extent.
 871                  */
 872                 err = insert_state(tree, prealloc, start, this_end,
 873                                    &bits);
 874                 if (err)
 875                         extent_io_tree_panic(tree, err);
 876
 877                 cache_state(prealloc, cached_state);
 878                 prealloc = NULL;
 879                 start = this_end + 1;
 880                 goto search_again;
 881         }
 882         /*
 883          * | ---- desired range ---- |
 884          *                        | state |
 885          * We need to split the extent, and set the bit
 886          * on the first half
 887          */
 888         if (state->start <= end && state->end > end) {
 889                 if (state->state & exclusive_bits) {
 890                         *failed_start = start;
 891                         err = -EEXIST;
 892                         goto out;
 893                 }
 894
 895                 prealloc = alloc_extent_state_atomic(prealloc);
 896                 BUG_ON(!prealloc);
 897                 err = split_state(tree, state, prealloc, end + 1);
 898                 if (err)
 899                         extent_io_tree_panic(tree, err);
 900
 901                 set_state_bits(tree, prealloc, &bits);
 902                 cache_state(prealloc, cached_state);
 903                 merge_state(tree, prealloc);
 904                 prealloc = NULL;
 905                 goto out;
 906         }
 907
 908         goto search_again;
 909
 910 out:
 911         spin_unlock(&tree->lock);
 912         if (prealloc)
 913                 free_extent_state(prealloc);
 914
 915         return err;
 916
 917 search_again:
 918         if (start > end)
 919                 goto out;
 920         spin_unlock(&tree->lock);
 921         if (mask & __GFP_WAIT)
 922                 cond_resched();
 923         goto again;
 924 }
 925
 926 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits,
 927                    u64 *failed_start, struct extent_state **cached_state,
 928                    gfp_t mask)
 929 {
 930         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
 931                                 cached_state, mask);
 932 }
 933
 934
 935 /**
 936  * convert_extent - convert all bits in a given range from one bit to another
 937  * @tree:       the io tree to search
 938  * @start:      the start offset in bytes
 939  * @end:        the end offset in bytes (inclusive)
 940  * @bits:       the bits to set in this range
 941  * @clear_bits: the bits to clear in this range
 942  * @mask:       the allocation mask
 943  *
 944  * This will go through and set bits for the given range.  If any states exist
 945  * already in this range they are set with the given bit and cleared of the
 946  * clear_bits.  This is only meant to be used by things that are mergeable, ie
 947  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
 948  * boundary bits like LOCK.
 949  */
 950 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 951                        int bits, int clear_bits, gfp_t mask)
 952 {
 953         struct extent_state *state;
 954         struct extent_state *prealloc = NULL;
 955         struct rb_node *node;
 956         int err = 0;
 957         u64 last_start;
 958         u64 last_end;
 959
 960 again:
 961         if (!prealloc && (mask & __GFP_WAIT)) {
 962                 prealloc = alloc_extent_state(mask);
 963                 if (!prealloc)
 964                         return -ENOMEM;
 965         }
 966
 967         spin_lock(&tree->lock);
 968         /*
 969          * this search will find all the extents that end after
 970          * our range starts.
 971          */
 972         node = tree_search(tree, start);
 973         if (!node) {
 974                 prealloc = alloc_extent_state_atomic(prealloc);
 975                 if (!prealloc) {
 976                         err = -ENOMEM;
 977                         goto out;
 978                 }
 979                 err = insert_state(tree, prealloc, start, end, &bits);
 980                 prealloc = NULL;
 981                 if (err)
 982                         extent_io_tree_panic(tree, err);
 983                 goto out;
 984         }
 985         state = rb_entry(node, struct extent_state, rb_node);
 986 hit_next:
 987         last_start = state->start;
 988         last_end = state->end;
 989
 990         /*
 991          * | ---- desired range ---- |
 992          * | state |
 993          *
 994          * Just lock what we found and keep going
 995          */
 996         if (state->start == start && state->end <= end) {
 997                 struct rb_node *next_node;
 998
 999                 set_state_bits(tree, state, &bits);
1000                 clear_state_bit(tree, state, &clear_bits, 0);
1001                 if (last_end == (u64)-1)
1002                         goto out;
1003
1004                 start = last_end + 1;
1005                 next_node = rb_next(&state->rb_node);
1006                 if (next_node && start < end && prealloc && !need_resched()) {
1007                         state = rb_entry(next_node, struct extent_state,
1008                                          rb_node);
1009                         if (state->start == start)
1010                                 goto hit_next;
1011                 }
1012                 goto search_again;
1013         }
1014
1015         /*
1016          *     | ---- desired range ---- |
1017          * | state |
1018          *   or
1019          * | ------------- state -------------- |
1020          *
1021          * We need to split the extent we found, and may flip bits on
1022          * second half.
1023          *
1024          * If the extent we found extends past our
1025          * range, we just split and search again.  It'll get split
1026          * again the next time though.
1027          *
1028          * If the extent we found is inside our range, we set the
1029          * desired bit on it.
1030          */
1031         if (state->start < start) {
1032                 prealloc = alloc_extent_state_atomic(prealloc);
1033                 if (!prealloc) {
1034                         err = -ENOMEM;
1035                         goto out;
1036                 }
1037                 err = split_state(tree, state, prealloc, start);
1038                 if (err)
1039                         extent_io_tree_panic(tree, err);
1040                 prealloc = NULL;
1041                 if (err)
1042                         goto out;
1043                 if (state->end <= end) {
1044                         set_state_bits(tree, state, &bits);
1045                         clear_state_bit(tree, state, &clear_bits, 0);
1046                         if (last_end == (u64)-1)
1047                                 goto out;
1048                         start = last_end + 1;
1049                 }
1050                 goto search_again;
1051         }
1052         /*
1053          * | ---- desired range ---- |
1054          *     | state | or               | state |
1055          *
1056          * There's a hole, we need to insert something in it and
1057          * ignore the extent we found.
1058          */
1059         if (state->start > start) {
1060                 u64 this_end;
1061                 if (end < last_start)
1062                         this_end = end;
1063                 else
1064                         this_end = last_start - 1;
1065
1066                 prealloc = alloc_extent_state_atomic(prealloc);
1067                 if (!prealloc) {
1068                         err = -ENOMEM;
1069                         goto out;
1070                 }
1071
1072                 /*
1073                  * Avoid to free 'prealloc' if it can be merged with
1074                  * the later extent.
1075                  */
1076                 err = insert_state(tree, prealloc, start, this_end,
1077                                    &bits);
1078                 if (err)
1079                         extent_io_tree_panic(tree, err);
1080                 prealloc = NULL;
1081                 start = this_end + 1;
1082                 goto search_again;
1083         }
1084         /*
1085          * | ---- desired range ---- |
1086          *                        | state |
1087          * We need to split the extent, and set the bit
1088          * on the first half
1089          */
1090         if (state->start <= end && state->end > end) {
1091                 prealloc = alloc_extent_state_atomic(prealloc);
1092                 if (!prealloc) {
1093                         err = -ENOMEM;
1094                         goto out;
1095                 }
1096
1097                 err = split_state(tree, state, prealloc, end + 1);
1098                 if (err)
1099                         extent_io_tree_panic(tree, err);
1100
1101                 set_state_bits(tree, prealloc, &bits);
1102                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1103                 prealloc = NULL;
1104                 goto out;
1105         }
1106
1107         goto search_again;
1108
1109 out:
1110         spin_unlock(&tree->lock);
1111         if (prealloc)
1112                 free_extent_state(prealloc);
1113
1114         return err;
1115
1116 search_again:
1117         if (start > end)
1118                 goto out;
1119         spin_unlock(&tree->lock);
1120         if (mask & __GFP_WAIT)
1121                 cond_resched();
1122         goto again;
1123 }
1124
1125 /* wrappers around set/clear extent bit */
1126 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1127                      gfp_t mask)
1128 {
1129         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1130                               NULL, mask);
1131 }
1132
1133 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1134                     int bits, gfp_t mask)
1135 {
1136         return set_extent_bit(tree, start, end, bits, NULL,
1137                               NULL, mask);
1138 }
1139
1140 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1141                       int bits, gfp_t mask)
1142 {
1143         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1144 }
1145
1146 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1147                         struct extent_state **cached_state, gfp_t mask)
1148 {
1149         return set_extent_bit(tree, start, end,
1150                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1151                               NULL, cached_state, mask);
1152 }
1153
1154 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1155                        gfp_t mask)
1156 {
1157         return clear_extent_bit(tree, start, end,
1158                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1159                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1160 }
1161
1162 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1163                      gfp_t mask)
1164 {
1165         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1166                               NULL, mask);
1167 }
1168
1169 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1170                         struct extent_state **cached_state, gfp_t mask)
1171 {
1172         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
1173                               cached_state, mask);
1174 }
1175
1176 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
1177                                  u64 end, struct extent_state **cached_state,
1178                                  gfp_t mask)
1179 {
1180         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1181                                 cached_state, mask);
1182 }
1183
1184 /*
1185  * either insert or lock state struct between start and end use mask to tell
1186  * us if waiting is desired.
1187  */
1188 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1189                      int bits, struct extent_state **cached_state)
1190 {
1191         int err;
1192         u64 failed_start;
1193         while (1) {
1194                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1195                                        EXTENT_LOCKED, &failed_start,
1196                                        cached_state, GFP_NOFS);
1197                 if (err == -EEXIST) {
1198                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1199                         start = failed_start;
1200                 } else
1201                         break;
1202                 WARN_ON(start > end);
1203         }
1204         return err;
1205 }
1206
1207 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1208 {
1209         return lock_extent_bits(tree, start, end, 0, NULL);
1210 }
1211
1212 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1213 {
1214         int err;
1215         u64 failed_start;
1216
1217         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1218                                &failed_start, NULL, GFP_NOFS);
1219         if (err == -EEXIST) {
1220                 if (failed_start > start)
1221                         clear_extent_bit(tree, start, failed_start - 1,
1222                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1223                 return 0;
1224         }
1225         return 1;
1226 }
1227
1228 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1229                          struct extent_state **cached, gfp_t mask)
1230 {
1231         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1232                                 mask);
1233 }
1234
1235 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1236 {
1237         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1238                                 GFP_NOFS);
1239 }
1240
1241 /*
1242  * helper function to set both pages and extents in the tree writeback
1243  */
1244 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1245 {
1246         unsigned long index = start >> PAGE_CACHE_SHIFT;
1247         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1248         struct page *page;
1249
1250         while (index <= end_index) {
1251                 page = find_get_page(tree->mapping, index);
1252                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1253                 set_page_writeback(page);
1254                 page_cache_release(page);
1255                 index++;
1256         }
1257         return 0;
1258 }
1259
1260 /* find the first state struct with 'bits' set after 'start', and
1261  * return it.  tree->lock must be held.  NULL will returned if
1262  * nothing was found after 'start'
1263  */
1264 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1265                                                  u64 start, int bits)
1266 {
1267         struct rb_node *node;
1268         struct extent_state *state;
1269
1270         /*
1271          * this search will find all the extents that end after
1272          * our range starts.
1273          */
1274         node = tree_search(tree, start);
1275         if (!node)
1276                 goto out;
1277
1278         while (1) {
1279                 state = rb_entry(node, struct extent_state, rb_node);
1280                 if (state->end >= start && (state->state & bits))
1281                         return state;
1282
1283                 node = rb_next(node);
1284                 if (!node)
1285                         break;
1286         }
1287 out:
1288         return NULL;
1289 }
1290
1291 /*
1292  * find the first offset in the io tree with 'bits' set. zero is
1293  * returned if we find something, and *start_ret and *end_ret are
1294  * set to reflect the state struct that was found.
1295  *
1296  * If nothing was found, 1 is returned, < 0 on error
1297  */
1298 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1299                           u64 *start_ret, u64 *end_ret, int bits)
1300 {
1301         struct extent_state *state;
1302         int ret = 1;
1303
1304         spin_lock(&tree->lock);
1305         state = find_first_extent_bit_state(tree, start, bits);
1306         if (state) {
1307                 *start_ret = state->start;
1308                 *end_ret = state->end;
1309                 ret = 0;
1310         }
1311         spin_unlock(&tree->lock);
1312         return ret;
1313 }
1314
1315 /*
1316  * find a contiguous range of bytes in the file marked as delalloc, not
1317  * more than 'max_bytes'.  start and end are used to return the range,
1318  *
1319  * 1 is returned if we find something, 0 if nothing was in the tree
1320  */
1321 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1322                                         u64 *start, u64 *end, u64 max_bytes,
1323                                         struct extent_state **cached_state)
1324 {
1325         struct rb_node *node;
1326         struct extent_state *state;
1327         u64 cur_start = *start;
1328         u64 found = 0;
1329         u64 total_bytes = 0;
1330
1331         spin_lock(&tree->lock);
1332
1333         /*
1334          * this search will find all the extents that end after
1335          * our range starts.
1336          */
1337         node = tree_search(tree, cur_start);
1338         if (!node) {
1339                 if (!found)
1340                         *end = (u64)-1;
1341                 goto out;
1342         }
1343
1344         while (1) {
1345                 state = rb_entry(node, struct extent_state, rb_node);
1346                 if (found && (state->start != cur_start ||
1347                               (state->state & EXTENT_BOUNDARY))) {
1348                         goto out;
1349                 }
1350                 if (!(state->state & EXTENT_DELALLOC)) {
1351                         if (!found)
1352                                 *end = state->end;
1353                         goto out;
1354                 }
1355                 if (!found) {
1356                         *start = state->start;
1357                         *cached_state = state;
1358                         atomic_inc(&state->refs);
1359                 }
1360                 found++;
1361                 *end = state->end;
1362                 cur_start = state->end + 1;
1363                 node = rb_next(node);
1364                 if (!node)
1365                         break;
1366                 total_bytes += state->end - state->start + 1;
1367                 if (total_bytes >= max_bytes)
1368                         break;
1369         }
1370 out:
1371         spin_unlock(&tree->lock);
1372         return found;
1373 }
1374
1375 static noinline void __unlock_for_delalloc(struct inode *inode,
1376                                            struct page *locked_page,
1377                                            u64 start, u64 end)
1378 {
1379         int ret;
1380         struct page *pages[16];
1381         unsigned long index = start >> PAGE_CACHE_SHIFT;
1382         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1383         unsigned long nr_pages = end_index - index + 1;
1384         int i;
1385
1386         if (index == locked_page->index && end_index == index)
1387                 return;
1388
1389         while (nr_pages > 0) {
1390                 ret = find_get_pages_contig(inode->i_mapping, index,
1391                                      min_t(unsigned long, nr_pages,
1392                                      ARRAY_SIZE(pages)), pages);
1393                 for (i = 0; i < ret; i++) {
1394                         if (pages[i] != locked_page)
1395                                 unlock_page(pages[i]);
1396                         page_cache_release(pages[i]);
1397                 }
1398                 nr_pages -= ret;
1399                 index += ret;
1400                 cond_resched();
1401         }
1402 }
1403
1404 static noinline int lock_delalloc_pages(struct inode *inode,
1405                                         struct page *locked_page,
1406                                         u64 delalloc_start,
1407                                         u64 delalloc_end)
1408 {
1409         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1410         unsigned long start_index = index;
1411         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1412         unsigned long pages_locked = 0;
1413         struct page *pages[16];
1414         unsigned long nrpages;
1415         int ret;
1416         int i;
1417
1418         /* the caller is responsible for locking the start index */
1419         if (index == locked_page->index && index == end_index)
1420                 return 0;
1421
1422         /* skip the page at the start index */
1423         nrpages = end_index - index + 1;
1424         while (nrpages > 0) {
1425                 ret = find_get_pages_contig(inode->i_mapping, index,
1426                                      min_t(unsigned long,
1427                                      nrpages, ARRAY_SIZE(pages)), pages);
1428                 if (ret == 0) {
1429                         ret = -EAGAIN;
1430                         goto done;
1431                 }
1432                 /* now we have an array of pages, lock them all */
1433                 for (i = 0; i < ret; i++) {
1434                         /*
1435                          * the caller is taking responsibility for
1436                          * locked_page
1437                          */
1438                         if (pages[i] != locked_page) {
1439                                 lock_page(pages[i]);
1440                                 if (!PageDirty(pages[i]) ||
1441                                     pages[i]->mapping != inode->i_mapping) {
1442                                         ret = -EAGAIN;
1443                                         unlock_page(pages[i]);
1444                                         page_cache_release(pages[i]);
1445                                         goto done;
1446                                 }
1447                         }
1448                         page_cache_release(pages[i]);
1449                         pages_locked++;
1450                 }
1451                 nrpages -= ret;
1452                 index += ret;
1453                 cond_resched();
1454         }
1455         ret = 0;
1456 done:
1457         if (ret && pages_locked) {
1458                 __unlock_for_delalloc(inode, locked_page,
1459                               delalloc_start,
1460                               ((u64)(start_index + pages_locked - 1)) <<
1461                               PAGE_CACHE_SHIFT);
1462         }
1463         return ret;
1464 }
1465
1466 /*
1467  * find a contiguous range of bytes in the file marked as delalloc, not
1468  * more than 'max_bytes'.  start and end are used to return the range,
1469  *
1470  * 1 is returned if we find something, 0 if nothing was in the tree
1471  */
1472 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1473                                              struct extent_io_tree *tree,
1474                                              struct page *locked_page,
1475                                              u64 *start, u64 *end,
1476                                              u64 max_bytes)
1477 {
1478         u64 delalloc_start;
1479         u64 delalloc_end;
1480         u64 found;
1481         struct extent_state *cached_state = NULL;
1482         int ret;
1483         int loops = 0;
1484
1485 again:
1486         /* step one, find a bunch of delalloc bytes starting at start */
1487         delalloc_start = *start;
1488         delalloc_end = 0;
1489         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1490                                     max_bytes, &cached_state);
1491         if (!found || delalloc_end <= *start) {
1492                 *start = delalloc_start;
1493                 *end = delalloc_end;
1494                 free_extent_state(cached_state);
1495                 return found;
1496         }
1497
1498         /*
1499          * start comes from the offset of locked_page.  We have to lock
1500          * pages in order, so we can't process delalloc bytes before
1501          * locked_page
1502          */
1503         if (delalloc_start < *start)
1504                 delalloc_start = *start;
1505
1506         /*
1507          * make sure to limit the number of pages we try to lock down
1508          * if we're looping.
1509          */
1510         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1511                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1512
1513         /* step two, lock all the pages after the page that has start */
1514         ret = lock_delalloc_pages(inode, locked_page,
1515                                   delalloc_start, delalloc_end);
1516         if (ret == -EAGAIN) {
1517                 /* some of the pages are gone, lets avoid looping by
1518                  * shortening the size of the delalloc range we're searching
1519                  */
1520                 free_extent_state(cached_state);
1521                 if (!loops) {
1522                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1523                         max_bytes = PAGE_CACHE_SIZE - offset;
1524                         loops = 1;
1525                         goto again;
1526                 } else {
1527                         found = 0;
1528                         goto out_failed;
1529                 }
1530         }
1531         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1532
1533         /* step three, lock the state bits for the whole range */
1534         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1535
1536         /* then test to make sure it is all still delalloc */
1537         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1538                              EXTENT_DELALLOC, 1, cached_state);
1539         if (!ret) {
1540                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1541                                      &cached_state, GFP_NOFS);
1542                 __unlock_for_delalloc(inode, locked_page,
1543                               delalloc_start, delalloc_end);
1544                 cond_resched();
1545                 goto again;
1546         }
1547         free_extent_state(cached_state);
1548         *start = delalloc_start;
1549         *end = delalloc_end;
1550 out_failed:
1551         return found;
1552 }
1553
1554 int extent_clear_unlock_delalloc(struct inode *inode,
1555                                 struct extent_io_tree *tree,
1556                                 u64 start, u64 end, struct page *locked_page,
1557                                 unsigned long op)
1558 {
1559         int ret;
1560         struct page *pages[16];
1561         unsigned long index = start >> PAGE_CACHE_SHIFT;
1562         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1563         unsigned long nr_pages = end_index - index + 1;
1564         int i;
1565         int clear_bits = 0;
1566
1567         if (op & EXTENT_CLEAR_UNLOCK)
1568                 clear_bits |= EXTENT_LOCKED;
1569         if (op & EXTENT_CLEAR_DIRTY)
1570                 clear_bits |= EXTENT_DIRTY;
1571
1572         if (op & EXTENT_CLEAR_DELALLOC)
1573                 clear_bits |= EXTENT_DELALLOC;
1574
1575         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1576         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1577                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1578                     EXTENT_SET_PRIVATE2)))
1579                 return 0;
1580
1581         while (nr_pages > 0) {
1582                 ret = find_get_pages_contig(inode->i_mapping, index,
1583                                      min_t(unsigned long,
1584                                      nr_pages, ARRAY_SIZE(pages)), pages);
1585                 for (i = 0; i < ret; i++) {
1586
1587                         if (op & EXTENT_SET_PRIVATE2)
1588                                 SetPagePrivate2(pages[i]);
1589
1590                         if (pages[i] == locked_page) {
1591                                 page_cache_release(pages[i]);
1592                                 continue;
1593                         }
1594                         if (op & EXTENT_CLEAR_DIRTY)
1595                                 clear_page_dirty_for_io(pages[i]);
1596                         if (op & EXTENT_SET_WRITEBACK)
1597                                 set_page_writeback(pages[i]);
1598                         if (op & EXTENT_END_WRITEBACK)
1599                                 end_page_writeback(pages[i]);
1600                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1601                                 unlock_page(pages[i]);
1602                         page_cache_release(pages[i]);
1603                 }
1604                 nr_pages -= ret;
1605                 index += ret;
1606                 cond_resched();
1607         }
1608         return 0;
1609 }
1610
1611 /*
1612  * count the number of bytes in the tree that have a given bit(s)
1613  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1614  * cached.  The total number found is returned.
1615  */
1616 u64 count_range_bits(struct extent_io_tree *tree,
1617                      u64 *start, u64 search_end, u64 max_bytes,
1618                      unsigned long bits, int contig)
1619 {
1620         struct rb_node *node;
1621         struct extent_state *state;
1622         u64 cur_start = *start;
1623         u64 total_bytes = 0;
1624         u64 last = 0;
1625         int found = 0;
1626
1627         if (search_end <= cur_start) {
1628                 WARN_ON(1);
1629                 return 0;
1630         }
1631
1632         spin_lock(&tree->lock);
1633         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1634                 total_bytes = tree->dirty_bytes;
1635                 goto out;
1636         }
1637         /*
1638          * this search will find all the extents that end after
1639          * our range starts.
1640          */
1641         node = tree_search(tree, cur_start);
1642         if (!node)
1643                 goto out;
1644
1645         while (1) {
1646                 state = rb_entry(node, struct extent_state, rb_node);
1647                 if (state->start > search_end)
1648                         break;
1649                 if (contig && found && state->start > last + 1)
1650                         break;
1651                 if (state->end >= cur_start && (state->state & bits) == bits) {
1652                         total_bytes += min(search_end, state->end) + 1 -
1653                                        max(cur_start, state->start);
1654                         if (total_bytes >= max_bytes)
1655                                 break;
1656                         if (!found) {
1657                                 *start = max(cur_start, state->start);
1658                                 found = 1;
1659                         }
1660                         last = state->end;
1661                 } else if (contig && found) {
1662                         break;
1663                 }
1664                 node = rb_next(node);
1665                 if (!node)
1666                         break;
1667         }
1668 out:
1669         spin_unlock(&tree->lock);
1670         return total_bytes;
1671 }
1672
1673 /*
1674  * set the private field for a given byte offset in the tree.  If there isn't
1675  * an extent_state there already, this does nothing.
1676  */
1677 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1678 {
1679         struct rb_node *node;
1680         struct extent_state *state;
1681         int ret = 0;
1682
1683         spin_lock(&tree->lock);
1684         /*
1685          * this search will find all the extents that end after
1686          * our range starts.
1687          */
1688         node = tree_search(tree, start);
1689         if (!node) {
1690                 ret = -ENOENT;
1691                 goto out;
1692         }
1693         state = rb_entry(node, struct extent_state, rb_node);
1694         if (state->start != start) {
1695                 ret = -ENOENT;
1696                 goto out;
1697         }
1698         state->private = private;
1699 out:
1700         spin_unlock(&tree->lock);
1701         return ret;
1702 }
1703
1704 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1705 {
1706         struct rb_node *node;
1707         struct extent_state *state;
1708         int ret = 0;
1709
1710         spin_lock(&tree->lock);
1711         /*
1712          * this search will find all the extents that end after
1713          * our range starts.
1714          */
1715         node = tree_search(tree, start);
1716         if (!node) {
1717                 ret = -ENOENT;
1718                 goto out;
1719         }
1720         state = rb_entry(node, struct extent_state, rb_node);
1721         if (state->start != start) {
1722                 ret = -ENOENT;
1723                 goto out;
1724         }
1725         *private = state->private;
1726 out:
1727         spin_unlock(&tree->lock);
1728         return ret;
1729 }
1730
1731 /*
1732  * searches a range in the state tree for a given mask.
1733  * If 'filled' == 1, this returns 1 only if every extent in the tree
1734  * has the bits set.  Otherwise, 1 is returned if any bit in the
1735  * range is found set.
1736  */
1737 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1738                    int bits, int filled, struct extent_state *cached)
1739 {
1740         struct extent_state *state = NULL;
1741         struct rb_node *node;
1742         int bitset = 0;
1743
1744         spin_lock(&tree->lock);
1745         if (cached && cached->tree && cached->start <= start &&
1746             cached->end > start)
1747                 node = &cached->rb_node;
1748         else
1749                 node = tree_search(tree, start);
1750         while (node && start <= end) {
1751                 state = rb_entry(node, struct extent_state, rb_node);
1752
1753                 if (filled && state->start > start) {
1754                         bitset = 0;
1755                         break;
1756                 }
1757
1758                 if (state->start > end)
1759                         break;
1760
1761                 if (state->state & bits) {
1762                         bitset = 1;
1763                         if (!filled)
1764                                 break;
1765                 } else if (filled) {
1766                         bitset = 0;
1767                         break;
1768                 }
1769
1770                 if (state->end == (u64)-1)
1771                         break;
1772
1773                 start = state->end + 1;
1774                 if (start > end)
1775                         break;
1776                 node = rb_next(node);
1777                 if (!node) {
1778                         if (filled)
1779                                 bitset = 0;
1780                         break;
1781                 }
1782         }
1783         spin_unlock(&tree->lock);
1784         return bitset;
1785 }
1786
1787 /*
1788  * helper function to set a given page up to date if all the
1789  * extents in the tree for that page are up to date
1790  */
1791 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1792 {
1793         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1794         u64 end = start + PAGE_CACHE_SIZE - 1;
1795         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1796                 SetPageUptodate(page);
1797 }
1798
1799 /*
1800  * helper function to unlock a page if all the extents in the tree
1801  * for that page are unlocked
1802  */
1803 static void check_page_locked(struct extent_io_tree *tree, struct page *page)
1804 {
1805         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1806         u64 end = start + PAGE_CACHE_SIZE - 1;
1807         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1808                 unlock_page(page);
1809 }
1810
1811 /*
1812  * helper function to end page writeback if all the extents
1813  * in the tree for that page are done with writeback
1814  */
1815 static void check_page_writeback(struct extent_io_tree *tree,
1816                                  struct page *page)
1817 {
1818         end_page_writeback(page);
1819 }
1820
1821 /*
1822  * When IO fails, either with EIO or csum verification fails, we
1823  * try other mirrors that might have a good copy of the data.  This
1824  * io_failure_record is used to record state as we go through all the
1825  * mirrors.  If another mirror has good data, the page is set up to date
1826  * and things continue.  If a good mirror can't be found, the original
1827  * bio end_io callback is called to indicate things have failed.
1828  */
1829 struct io_failure_record {
1830         struct page *page;
1831         u64 start;
1832         u64 len;
1833         u64 logical;
1834         unsigned long bio_flags;
1835         int this_mirror;
1836         int failed_mirror;
1837         int in_validation;
1838 };
1839
1840 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1841                                 int did_repair)
1842 {
1843         int ret;
1844         int err = 0;
1845         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1846
1847         set_state_private(failure_tree, rec->start, 0);
1848         ret = clear_extent_bits(failure_tree, rec->start,
1849                                 rec->start + rec->len - 1,
1850                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1851         if (ret)
1852                 err = ret;
1853
1854         if (did_repair) {
1855                 ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1856                                         rec->start + rec->len - 1,
1857                                         EXTENT_DAMAGED, GFP_NOFS);
1858                 if (ret && !err)
1859                         err = ret;
1860         }
1861
1862         kfree(rec);
1863         return err;
1864 }
1865
1866 static void repair_io_failure_callback(struct bio *bio, int err)
1867 {
1868         complete(bio->bi_private);
1869 }
1870
1871 /*
1872  * this bypasses the standard btrfs submit functions deliberately, as
1873  * the standard behavior is to write all copies in a raid setup. here we only
1874  * want to write the one bad copy. so we do the mapping for ourselves and issue
1875  * submit_bio directly.
1876  * to avoid any synchonization issues, wait for the data after writing, which
1877  * actually prevents the read that triggered the error from finishing.
1878  * currently, there can be no more than two copies of every data bit. thus,
1879  * exactly one rewrite is required.
1880  */
1881 int repair_io_failure(struct btrfs_mapping_tree *map_tree, u64 start,
1882                         u64 length, u64 logical, struct page *page,
1883                         int mirror_num)
1884 {
1885         struct bio *bio;
1886         struct btrfs_device *dev;
1887         DECLARE_COMPLETION_ONSTACK(compl);
1888         u64 map_length = 0;
1889         u64 sector;
1890         struct btrfs_bio *bbio = NULL;
1891         int ret;
1892
1893         BUG_ON(!mirror_num);
1894
1895         bio = bio_alloc(GFP_NOFS, 1);
1896         if (!bio)
1897                 return -EIO;
1898         bio->bi_private = &compl;
1899         bio->bi_end_io = repair_io_failure_callback;
1900         bio->bi_size = 0;
1901         map_length = length;
1902
1903         ret = btrfs_map_block(map_tree, WRITE, logical,
1904                               &map_length, &bbio, mirror_num);
1905         if (ret) {
1906                 bio_put(bio);
1907                 return -EIO;
1908         }
1909         BUG_ON(mirror_num != bbio->mirror_num);
1910         sector = bbio->stripes[mirror_num-1].physical >> 9;
1911         bio->bi_sector = sector;
1912         dev = bbio->stripes[mirror_num-1].dev;
1913         kfree(bbio);
1914         if (!dev || !dev->bdev || !dev->writeable) {
1915                 bio_put(bio);
1916                 return -EIO;
1917         }
1918         bio->bi_bdev = dev->bdev;
1919         bio_add_page(bio, page, length, start-page_offset(page));
1920         btrfsic_submit_bio(WRITE_SYNC, bio);
1921         wait_for_completion(&compl);
1922
1923         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
1924                 /* try to remap that extent elsewhere? */
1925                 bio_put(bio);
1926                 return -EIO;
1927         }
1928
1929         printk(KERN_INFO "btrfs read error corrected: ino %lu off %llu (dev %s "
1930                         "sector %llu)\n", page->mapping->host->i_ino, start,
1931                         dev->name, sector);
1932
1933         bio_put(bio);
1934         return 0;
1935 }
1936
1937 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
1938                          int mirror_num)
1939 {
1940         struct btrfs_mapping_tree *map_tree = &root->fs_info->mapping_tree;
1941         u64 start = eb->start;
1942         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
1943         int ret = 0;
1944
1945         for (i = 0; i < num_pages; i++) {
1946                 struct page *p = extent_buffer_page(eb, i);
1947                 ret = repair_io_failure(map_tree, start, PAGE_CACHE_SIZE,
1948                                         start, p, mirror_num);
1949                 if (ret)
1950                         break;
1951                 start += PAGE_CACHE_SIZE;
1952         }
1953
1954         return ret;
1955 }
1956
1957 /*
1958  * each time an IO finishes, we do a fast check in the IO failure tree
1959  * to see if we need to process or clean up an io_failure_record
1960  */
1961 static int clean_io_failure(u64 start, struct page *page)
1962 {
1963         u64 private;
1964         u64 private_failure;
1965         struct io_failure_record *failrec;
1966         struct btrfs_mapping_tree *map_tree;
1967         struct extent_state *state;
1968         int num_copies;
1969         int did_repair = 0;
1970         int ret;
1971         struct inode *inode = page->mapping->host;
1972
1973         private = 0;
1974         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
1975                                 (u64)-1, 1, EXTENT_DIRTY, 0);
1976         if (!ret)
1977                 return 0;
1978
1979         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
1980                                 &private_failure);
1981         if (ret)
1982                 return 0;
1983
1984         failrec = (struct io_failure_record *)(unsigned long) private_failure;
1985         BUG_ON(!failrec->this_mirror);
1986
1987         if (failrec->in_validation) {
1988                 /* there was no real error, just free the record */
1989                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
1990                          failrec->start);
1991                 did_repair = 1;
1992                 goto out;
1993         }
1994
1995         spin_lock(&BTRFS_I(inode)->io_tree.lock);
1996         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
1997                                             failrec->start,
1998                                             EXTENT_LOCKED);
1999         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2000
2001         if (state && state->start == failrec->start) {
2002                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
2003                 num_copies = btrfs_num_copies(map_tree, failrec->logical,
2004                                                 failrec->len);
2005                 if (num_copies > 1)  {
2006                         ret = repair_io_failure(map_tree, start, failrec->len,
2007                                                 failrec->logical, page,
2008                                                 failrec->failed_mirror);
2009                         did_repair = !ret;
2010                 }
2011         }
2012
2013 out:
2014         if (!ret)
2015                 ret = free_io_failure(inode, failrec, did_repair);
2016
2017         return ret;
2018 }
2019
2020 /*
2021  * this is a generic handler for readpage errors (default
2022  * readpage_io_failed_hook). if other copies exist, read those and write back
2023  * good data to the failed position. does not investigate in remapping the
2024  * failed extent elsewhere, hoping the device will be smart enough to do this as
2025  * needed
2026  */
2027
2028 static int bio_readpage_error(struct bio *failed_bio, struct page *page,
2029                                 u64 start, u64 end, int failed_mirror,
2030                                 struct extent_state *state)
2031 {
2032         struct io_failure_record *failrec = NULL;
2033         u64 private;
2034         struct extent_map *em;
2035         struct inode *inode = page->mapping->host;
2036         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2037         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2038         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2039         struct bio *bio;
2040         int num_copies;
2041         int ret;
2042         int read_mode;
2043         u64 logical;
2044
2045         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2046
2047         ret = get_state_private(failure_tree, start, &private);
2048         if (ret) {
2049                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2050                 if (!failrec)
2051                         return -ENOMEM;
2052                 failrec->start = start;
2053                 failrec->len = end - start + 1;
2054                 failrec->this_mirror = 0;
2055                 failrec->bio_flags = 0;
2056                 failrec->in_validation = 0;
2057
2058                 read_lock(&em_tree->lock);
2059                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2060                 if (!em) {
2061                         read_unlock(&em_tree->lock);
2062                         kfree(failrec);
2063                         return -EIO;
2064                 }
2065
2066                 if (em->start > start || em->start + em->len < start) {
2067                         free_extent_map(em);
2068                         em = NULL;
2069                 }
2070                 read_unlock(&em_tree->lock);
2071
2072                 if (!em || IS_ERR(em)) {
2073                         kfree(failrec);
2074                         return -EIO;
2075                 }
2076                 logical = start - em->start;
2077                 logical = em->block_start + logical;
2078                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2079                         logical = em->block_start;
2080                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2081                         extent_set_compress_type(&failrec->bio_flags,
2082                                                  em->compress_type);
2083                 }
2084                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2085                          "len=%llu\n", logical, start, failrec->len);
2086                 failrec->logical = logical;
2087                 free_extent_map(em);
2088
2089                 /* set the bits in the private failure tree */
2090                 ret = set_extent_bits(failure_tree, start, end,
2091                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2092                 if (ret >= 0)
2093                         ret = set_state_private(failure_tree, start,
2094                                                 (u64)(unsigned long)failrec);
2095                 /* set the bits in the inode's tree */
2096                 if (ret >= 0)
2097                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2098                                                 GFP_NOFS);
2099                 if (ret < 0) {
2100                         kfree(failrec);
2101                         return ret;
2102                 }
2103         } else {
2104                 failrec = (struct io_failure_record *)(unsigned long)private;
2105                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2106                          "start=%llu, len=%llu, validation=%d\n",
2107                          failrec->logical, failrec->start, failrec->len,
2108                          failrec->in_validation);
2109                 /*
2110                  * when data can be on disk more than twice, add to failrec here
2111                  * (e.g. with a list for failed_mirror) to make
2112                  * clean_io_failure() clean all those errors at once.
2113                  */
2114         }
2115         num_copies = btrfs_num_copies(
2116                               &BTRFS_I(inode)->root->fs_info->mapping_tree,
2117                               failrec->logical, failrec->len);
2118         if (num_copies == 1) {
2119                 /*
2120                  * we only have a single copy of the data, so don't bother with
2121                  * all the retry and error correction code that follows. no
2122                  * matter what the error is, it is very likely to persist.
2123                  */
2124                 pr_debug("bio_readpage_error: cannot repair, num_copies == 1. "
2125                          "state=%p, num_copies=%d, next_mirror %d, "
2126                          "failed_mirror %d\n", state, num_copies,
2127                          failrec->this_mirror, failed_mirror);
2128                 free_io_failure(inode, failrec, 0);
2129                 return -EIO;
2130         }
2131
2132         if (!state) {
2133                 spin_lock(&tree->lock);
2134                 state = find_first_extent_bit_state(tree, failrec->start,
2135                                                     EXTENT_LOCKED);
2136                 if (state && state->start != failrec->start)
2137                         state = NULL;
2138                 spin_unlock(&tree->lock);
2139         }
2140
2141         /*
2142          * there are two premises:
2143          *      a) deliver good data to the caller
2144          *      b) correct the bad sectors on disk
2145          */
2146         if (failed_bio->bi_vcnt > 1) {
2147                 /*
2148                  * to fulfill b), we need to know the exact failing sectors, as
2149                  * we don't want to rewrite any more than the failed ones. thus,
2150                  * we need separate read requests for the failed bio
2151                  *
2152                  * if the following BUG_ON triggers, our validation request got
2153                  * merged. we need separate requests for our algorithm to work.
2154                  */
2155                 BUG_ON(failrec->in_validation);
2156                 failrec->in_validation = 1;
2157                 failrec->this_mirror = failed_mirror;
2158                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2159         } else {
2160                 /*
2161                  * we're ready to fulfill a) and b) alongside. get a good copy
2162                  * of the failed sector and if we succeed, we have setup
2163                  * everything for repair_io_failure to do the rest for us.
2164                  */
2165                 if (failrec->in_validation) {
2166                         BUG_ON(failrec->this_mirror != failed_mirror);
2167                         failrec->in_validation = 0;
2168                         failrec->this_mirror = 0;
2169                 }
2170                 failrec->failed_mirror = failed_mirror;
2171                 failrec->this_mirror++;
2172                 if (failrec->this_mirror == failed_mirror)
2173                         failrec->this_mirror++;
2174                 read_mode = READ_SYNC;
2175         }
2176
2177         if (!state || failrec->this_mirror > num_copies) {
2178                 pr_debug("bio_readpage_error: (fail) state=%p, num_copies=%d, "
2179                          "next_mirror %d, failed_mirror %d\n", state,
2180                          num_copies, failrec->this_mirror, failed_mirror);
2181                 free_io_failure(inode, failrec, 0);
2182                 return -EIO;
2183         }
2184
2185         bio = bio_alloc(GFP_NOFS, 1);
2186         if (!bio) {
2187                 free_io_failure(inode, failrec, 0);
2188                 return -EIO;
2189         }
2190         bio->bi_private = state;
2191         bio->bi_end_io = failed_bio->bi_end_io;
2192         bio->bi_sector = failrec->logical >> 9;
2193         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2194         bio->bi_size = 0;
2195
2196         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2197
2198         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2199                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2200                  failrec->this_mirror, num_copies, failrec->in_validation);
2201
2202         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2203                                          failrec->this_mirror,
2204                                          failrec->bio_flags, 0);
2205         return ret;
2206 }
2207
2208 /* lots and lots of room for performance fixes in the end_bio funcs */
2209
2210 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2211 {
2212         int uptodate = (err == 0);
2213         struct extent_io_tree *tree;
2214         int ret;
2215
2216         tree = &BTRFS_I(page->mapping->host)->io_tree;
2217
2218         if (tree->ops && tree->ops->writepage_end_io_hook) {
2219                 ret = tree->ops->writepage_end_io_hook(page, start,
2220                                                end, NULL, uptodate);
2221                 if (ret)
2222                         uptodate = 0;
2223         }
2224
2225         if (!uptodate && tree->ops &&
2226             tree->ops->writepage_io_failed_hook) {
2227                 ret = tree->ops->writepage_io_failed_hook(NULL, page,
2228                                                  start, end, NULL);
2229                 /* Writeback already completed */
2230                 if (ret == 0)
2231                         return 1;
2232         }
2233
2234         if (!uptodate) {
2235                 clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
2236                 ClearPageUptodate(page);
2237                 SetPageError(page);
2238         }
2239         return 0;
2240 }
2241
2242 /*
2243  * after a writepage IO is done, we need to:
2244  * clear the uptodate bits on error
2245  * clear the writeback bits in the extent tree for this IO
2246  * end_page_writeback if the page has no more pending IO
2247  *
2248  * Scheduling is not allowed, so the extent state tree is expected
2249  * to have one and only one object corresponding to this IO.
2250  */
2251 static void end_bio_extent_writepage(struct bio *bio, int err)
2252 {
2253         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2254         struct extent_io_tree *tree;
2255         u64 start;
2256         u64 end;
2257         int whole_page;
2258
2259         do {
2260                 struct page *page = bvec->bv_page;
2261                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2262
2263                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2264                          bvec->bv_offset;
2265                 end = start + bvec->bv_len - 1;
2266
2267                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2268                         whole_page = 1;
2269                 else
2270                         whole_page = 0;
2271
2272                 if (--bvec >= bio->bi_io_vec)
2273                         prefetchw(&bvec->bv_page->flags);
2274
2275                 if (end_extent_writepage(page, err, start, end))
2276                         continue;
2277
2278                 if (whole_page)
2279                         end_page_writeback(page);
2280                 else
2281                         check_page_writeback(tree, page);
2282         } while (bvec >= bio->bi_io_vec);
2283
2284         bio_put(bio);
2285 }
2286
2287 /*
2288  * after a readpage IO is done, we need to:
2289  * clear the uptodate bits on error
2290  * set the uptodate bits if things worked
2291  * set the page up to date if all extents in the tree are uptodate
2292  * clear the lock bit in the extent tree
2293  * unlock the page if there are no other extents locked for it
2294  *
2295  * Scheduling is not allowed, so the extent state tree is expected
2296  * to have one and only one object corresponding to this IO.
2297  */
2298 static void end_bio_extent_readpage(struct bio *bio, int err)
2299 {
2300         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2301         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2302         struct bio_vec *bvec = bio->bi_io_vec;
2303         struct extent_io_tree *tree;
2304         u64 start;
2305         u64 end;
2306         int whole_page;
2307         int mirror;
2308         int ret;
2309
2310         if (err)
2311                 uptodate = 0;
2312
2313         do {
2314                 struct page *page = bvec->bv_page;
2315                 struct extent_state *cached = NULL;
2316                 struct extent_state *state;
2317
2318                 pr_debug("end_bio_extent_readpage: bi_vcnt=%d, idx=%d, err=%d, "
2319                          "mirror=%ld\n", bio->bi_vcnt, bio->bi_idx, err,
2320                          (long int)bio->bi_bdev);
2321                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2322
2323                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
2324                         bvec->bv_offset;
2325                 end = start + bvec->bv_len - 1;
2326
2327                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2328                         whole_page = 1;
2329                 else
2330                         whole_page = 0;
2331
2332                 if (++bvec <= bvec_end)
2333                         prefetchw(&bvec->bv_page->flags);
2334
2335                 spin_lock(&tree->lock);
2336                 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
2337                 if (state && state->start == start) {
2338                         /*
2339                          * take a reference on the state, unlock will drop
2340                          * the ref
2341                          */
2342                         cache_state(state, &cached);
2343                 }
2344                 spin_unlock(&tree->lock);
2345
2346                 mirror = (int)(unsigned long)bio->bi_bdev;
2347                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
2348                         ret = tree->ops->readpage_end_io_hook(page, start, end,
2349                                                               state, mirror);
2350                         if (ret)
2351                                 uptodate = 0;
2352                         else
2353                                 clean_io_failure(start, page);
2354                 }
2355
2356                 if (!uptodate && tree->ops && tree->ops->readpage_io_failed_hook) {
2357                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2358                         if (!ret && !err &&
2359                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2360                                 uptodate = 1;
2361                 } else if (!uptodate) {
2362                         /*
2363                          * The generic bio_readpage_error handles errors the
2364                          * following way: If possible, new read requests are
2365                          * created and submitted and will end up in
2366                          * end_bio_extent_readpage as well (if we're lucky, not
2367                          * in the !uptodate case). In that case it returns 0 and
2368                          * we just go on with the next page in our bio. If it
2369                          * can't handle the error it will return -EIO and we
2370                          * remain responsible for that page.
2371                          */
2372                         ret = bio_readpage_error(bio, page, start, end, mirror, NULL);
2373                         if (ret == 0) {
2374                                 uptodate =
2375                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2376                                 if (err)
2377                                         uptodate = 0;
2378                                 uncache_state(&cached);
2379                                 continue;
2380                         }
2381                 }
2382
2383                 if (uptodate && tree->track_uptodate) {
2384                         set_extent_uptodate(tree, start, end, &cached,
2385                                             GFP_ATOMIC);
2386                 }
2387                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2388
2389                 if (whole_page) {
2390                         if (uptodate) {
2391                                 SetPageUptodate(page);
2392                         } else {
2393                                 ClearPageUptodate(page);
2394                                 SetPageError(page);
2395                         }
2396                         unlock_page(page);
2397                 } else {
2398                         if (uptodate) {
2399                                 check_page_uptodate(tree, page);
2400                         } else {
2401                                 ClearPageUptodate(page);
2402                                 SetPageError(page);
2403                         }
2404                         check_page_locked(tree, page);
2405                 }
2406         } while (bvec <= bvec_end);
2407
2408         bio_put(bio);
2409 }
2410
2411 struct bio *
2412 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2413                 gfp_t gfp_flags)
2414 {
2415         struct bio *bio;
2416
2417         bio = bio_alloc(gfp_flags, nr_vecs);
2418
2419         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2420                 while (!bio && (nr_vecs /= 2))
2421                         bio = bio_alloc(gfp_flags, nr_vecs);
2422         }
2423
2424         if (bio) {
2425                 bio->bi_size = 0;
2426                 bio->bi_bdev = bdev;
2427                 bio->bi_sector = first_sector;
2428         }
2429         return bio;
2430 }
2431
2432 /*
2433  * Since writes are async, they will only return -ENOMEM.
2434  * Reads can return the full range of I/O error conditions.
2435  */
2436 static int __must_check submit_one_bio(int rw, struct bio *bio,
2437                                        int mirror_num, unsigned long bio_flags)
2438 {
2439         int ret = 0;
2440         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2441         struct page *page = bvec->bv_page;
2442         struct extent_io_tree *tree = bio->bi_private;
2443         u64 start;
2444
2445         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
2446
2447         bio->bi_private = NULL;
2448
2449         bio_get(bio);
2450
2451         if (tree->ops && tree->ops->submit_bio_hook)
2452                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2453                                            mirror_num, bio_flags, start);
2454         else
2455                 btrfsic_submit_bio(rw, bio);
2456
2457         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2458                 ret = -EOPNOTSUPP;
2459         bio_put(bio);
2460         return ret;
2461 }
2462
2463 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2464                      unsigned long offset, size_t size, struct bio *bio,
2465                      unsigned long bio_flags)
2466 {
2467         int ret = 0;
2468         if (tree->ops && tree->ops->merge_bio_hook)
2469                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2470                                                 bio_flags);
2471         BUG_ON(ret < 0);
2472         return ret;
2473
2474 }
2475
2476 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2477                               struct page *page, sector_t sector,
2478                               size_t size, unsigned long offset,
2479                               struct block_device *bdev,
2480                               struct bio **bio_ret,
2481                               unsigned long max_pages,
2482                               bio_end_io_t end_io_func,
2483                               int mirror_num,
2484                               unsigned long prev_bio_flags,
2485                               unsigned long bio_flags)
2486 {
2487         int ret = 0;
2488         struct bio *bio;
2489         int nr;
2490         int contig = 0;
2491         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2492         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2493         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2494
2495         if (bio_ret && *bio_ret) {
2496                 bio = *bio_ret;
2497                 if (old_compressed)
2498                         contig = bio->bi_sector == sector;
2499                 else
2500                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
2501                                 sector;
2502
2503                 if (prev_bio_flags != bio_flags || !contig ||
2504                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2505                     bio_add_page(bio, page, page_size, offset) < page_size) {
2506                         ret = submit_one_bio(rw, bio, mirror_num,
2507                                              prev_bio_flags);
2508                         if (ret < 0)
2509                                 return ret;
2510                         bio = NULL;
2511                 } else {
2512                         return 0;
2513                 }
2514         }
2515         if (this_compressed)
2516                 nr = BIO_MAX_PAGES;
2517         else
2518                 nr = bio_get_nr_vecs(bdev);
2519
2520         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2521         if (!bio)
2522                 return -ENOMEM;
2523
2524         bio_add_page(bio, page, page_size, offset);
2525         bio->bi_end_io = end_io_func;
2526         bio->bi_private = tree;
2527
2528         if (bio_ret)
2529                 *bio_ret = bio;
2530         else
2531                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2532
2533         return ret;
2534 }
2535
2536 void attach_extent_buffer_page(struct extent_buffer *eb, struct page *page)
2537 {
2538         if (!PagePrivate(page)) {
2539                 SetPagePrivate(page);
2540                 page_cache_get(page);
2541                 set_page_private(page, (unsigned long)eb);
2542         } else {
2543                 WARN_ON(page->private != (unsigned long)eb);
2544         }
2545 }
2546
2547 void set_page_extent_mapped(struct page *page)
2548 {
2549         if (!PagePrivate(page)) {
2550                 SetPagePrivate(page);
2551                 page_cache_get(page);
2552                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2553         }
2554 }
2555
2556 /*
2557  * basic readpage implementation.  Locked extent state structs are inserted
2558  * into the tree that are removed when the IO is done (by the end_io
2559  * handlers)
2560  * XXX JDM: This needs looking at to ensure proper page locking
2561  */
2562 static int __extent_read_full_page(struct extent_io_tree *tree,
2563                                    struct page *page,
2564                                    get_extent_t *get_extent,
2565                                    struct bio **bio, int mirror_num,
2566                                    unsigned long *bio_flags)
2567 {
2568         struct inode *inode = page->mapping->host;
2569         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2570         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2571         u64 end;
2572         u64 cur = start;
2573         u64 extent_offset;
2574         u64 last_byte = i_size_read(inode);
2575         u64 block_start;
2576         u64 cur_end;
2577         sector_t sector;
2578         struct extent_map *em;
2579         struct block_device *bdev;
2580         struct btrfs_ordered_extent *ordered;
2581         int ret;
2582         int nr = 0;
2583         size_t pg_offset = 0;
2584         size_t iosize;
2585         size_t disk_io_size;
2586         size_t blocksize = inode->i_sb->s_blocksize;
2587         unsigned long this_bio_flag = 0;
2588
2589         set_page_extent_mapped(page);
2590
2591         if (!PageUptodate(page)) {
2592                 if (cleancache_get_page(page) == 0) {
2593                         BUG_ON(blocksize != PAGE_SIZE);
2594                         goto out;
2595                 }
2596         }
2597
2598         end = page_end;
2599         while (1) {
2600                 lock_extent(tree, start, end);
2601                 ordered = btrfs_lookup_ordered_extent(inode, start);
2602                 if (!ordered)
2603                         break;
2604                 unlock_extent(tree, start, end);
2605                 btrfs_start_ordered_extent(inode, ordered, 1);
2606                 btrfs_put_ordered_extent(ordered);
2607         }
2608
2609         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2610                 char *userpage;
2611                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2612
2613                 if (zero_offset) {
2614                         iosize = PAGE_CACHE_SIZE - zero_offset;
2615                         userpage = kmap_atomic(page);
2616                         memset(userpage + zero_offset, 0, iosize);
2617                         flush_dcache_page(page);
2618                         kunmap_atomic(userpage);
2619                 }
2620         }
2621         while (cur <= end) {
2622                 if (cur >= last_byte) {
2623                         char *userpage;
2624                         struct extent_state *cached = NULL;
2625
2626                         iosize = PAGE_CACHE_SIZE - pg_offset;
2627                         userpage = kmap_atomic(page);
2628                         memset(userpage + pg_offset, 0, iosize);
2629                         flush_dcache_page(page);
2630                         kunmap_atomic(userpage);
2631                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2632                                             &cached, GFP_NOFS);
2633                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2634                                              &cached, GFP_NOFS);
2635                         break;
2636                 }
2637                 em = get_extent(inode, page, pg_offset, cur,
2638                                 end - cur + 1, 0);
2639                 if (IS_ERR_OR_NULL(em)) {
2640                         SetPageError(page);
2641                         unlock_extent(tree, cur, end);
2642                         break;
2643                 }
2644                 extent_offset = cur - em->start;
2645                 BUG_ON(extent_map_end(em) <= cur);
2646                 BUG_ON(end < cur);
2647
2648                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2649                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2650                         extent_set_compress_type(&this_bio_flag,
2651                                                  em->compress_type);
2652                 }
2653
2654                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2655                 cur_end = min(extent_map_end(em) - 1, end);
2656                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2657                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2658                         disk_io_size = em->block_len;
2659                         sector = em->block_start >> 9;
2660                 } else {
2661                         sector = (em->block_start + extent_offset) >> 9;
2662                         disk_io_size = iosize;
2663                 }
2664                 bdev = em->bdev;
2665                 block_start = em->block_start;
2666                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2667                         block_start = EXTENT_MAP_HOLE;
2668                 free_extent_map(em);
2669                 em = NULL;
2670
2671                 /* we've found a hole, just zero and go on */
2672                 if (block_start == EXTENT_MAP_HOLE) {
2673                         char *userpage;
2674                         struct extent_state *cached = NULL;
2675
2676                         userpage = kmap_atomic(page);
2677                         memset(userpage + pg_offset, 0, iosize);
2678                         flush_dcache_page(page);
2679                         kunmap_atomic(userpage);
2680
2681                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2682                                             &cached, GFP_NOFS);
2683                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2684                                              &cached, GFP_NOFS);
2685                         cur = cur + iosize;
2686                         pg_offset += iosize;
2687                         continue;
2688                 }
2689                 /* the get_extent function already copied into the page */
2690                 if (test_range_bit(tree, cur, cur_end,
2691                                    EXTENT_UPTODATE, 1, NULL)) {
2692                         check_page_uptodate(tree, page);
2693                         unlock_extent(tree, cur, cur + iosize - 1);
2694                         cur = cur + iosize;
2695                         pg_offset += iosize;
2696                         continue;
2697                 }
2698                 /* we have an inline extent but it didn't get marked up
2699                  * to date.  Error out
2700                  */
2701                 if (block_start == EXTENT_MAP_INLINE) {
2702                         SetPageError(page);
2703                         unlock_extent(tree, cur, cur + iosize - 1);
2704                         cur = cur + iosize;
2705                         pg_offset += iosize;
2706                         continue;
2707                 }
2708
2709                 ret = 0;
2710                 if (tree->ops && tree->ops->readpage_io_hook) {
2711                         ret = tree->ops->readpage_io_hook(page, cur,
2712                                                           cur + iosize - 1);
2713                 }
2714                 if (!ret) {
2715                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2716                         pnr -= page->index;
2717                         ret = submit_extent_page(READ, tree, page,
2718                                          sector, disk_io_size, pg_offset,
2719                                          bdev, bio, pnr,
2720                                          end_bio_extent_readpage, mirror_num,
2721                                          *bio_flags,
2722                                          this_bio_flag);
2723                         BUG_ON(ret == -ENOMEM);
2724                         nr++;
2725                         *bio_flags = this_bio_flag;
2726                 }
2727                 if (ret)
2728                         SetPageError(page);
2729                 cur = cur + iosize;
2730                 pg_offset += iosize;
2731         }
2732 out:
2733         if (!nr) {
2734                 if (!PageError(page))
2735                         SetPageUptodate(page);
2736                 unlock_page(page);
2737         }
2738         return 0;
2739 }
2740
2741 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2742                             get_extent_t *get_extent, int mirror_num)
2743 {
2744         struct bio *bio = NULL;
2745         unsigned long bio_flags = 0;
2746         int ret;
2747
2748         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
2749                                       &bio_flags);
2750         if (bio)
2751                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
2752         return ret;
2753 }
2754
2755 static noinline void update_nr_written(struct page *page,
2756                                       struct writeback_control *wbc,
2757                                       unsigned long nr_written)
2758 {
2759         wbc->nr_to_write -= nr_written;
2760         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2761             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2762                 page->mapping->writeback_index = page->index + nr_written;
2763 }
2764
2765 /*
2766  * the writepage semantics are similar to regular writepage.  extent
2767  * records are inserted to lock ranges in the tree, and as dirty areas
2768  * are found, they are marked writeback.  Then the lock bits are removed
2769  * and the end_io handler clears the writeback ranges
2770  */
2771 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2772                               void *data)
2773 {
2774         struct inode *inode = page->mapping->host;
2775         struct extent_page_data *epd = data;
2776         struct extent_io_tree *tree = epd->tree;
2777         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2778         u64 delalloc_start;
2779         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2780         u64 end;
2781         u64 cur = start;
2782         u64 extent_offset;
2783         u64 last_byte = i_size_read(inode);
2784         u64 block_start;
2785         u64 iosize;
2786         sector_t sector;
2787         struct extent_state *cached_state = NULL;
2788         struct extent_map *em;
2789         struct block_device *bdev;
2790         int ret;
2791         int nr = 0;
2792         size_t pg_offset = 0;
2793         size_t blocksize;
2794         loff_t i_size = i_size_read(inode);
2795         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2796         u64 nr_delalloc;
2797         u64 delalloc_end;
2798         int page_started;
2799         int compressed;
2800         int write_flags;
2801         unsigned long nr_written = 0;
2802         bool fill_delalloc = true;
2803
2804         if (wbc->sync_mode == WB_SYNC_ALL)
2805                 write_flags = WRITE_SYNC;
2806         else
2807                 write_flags = WRITE;
2808
2809         trace___extent_writepage(page, inode, wbc);
2810
2811         WARN_ON(!PageLocked(page));
2812
2813         ClearPageError(page);
2814
2815         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2816         if (page->index > end_index ||
2817            (page->index == end_index && !pg_offset)) {
2818                 page->mapping->a_ops->invalidatepage(page, 0);
2819                 unlock_page(page);
2820                 return 0;
2821         }
2822
2823         if (page->index == end_index) {
2824                 char *userpage;
2825
2826                 userpage = kmap_atomic(page);
2827                 memset(userpage + pg_offset, 0,
2828                        PAGE_CACHE_SIZE - pg_offset);
2829                 kunmap_atomic(userpage);
2830                 flush_dcache_page(page);
2831         }
2832         pg_offset = 0;
2833
2834         set_page_extent_mapped(page);
2835
2836         if (!tree->ops || !tree->ops->fill_delalloc)
2837                 fill_delalloc = false;
2838
2839         delalloc_start = start;
2840         delalloc_end = 0;
2841         page_started = 0;
2842         if (!epd->extent_locked && fill_delalloc) {
2843                 u64 delalloc_to_write = 0;
2844                 /*
2845                  * make sure the wbc mapping index is at least updated
2846                  * to this page.
2847                  */
2848                 update_nr_written(page, wbc, 0);
2849
2850                 while (delalloc_end < page_end) {
2851                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2852                                                        page,
2853                                                        &delalloc_start,
2854                                                        &delalloc_end,
2855                                                        128 * 1024 * 1024);
2856                         if (nr_delalloc == 0) {
2857                                 delalloc_start = delalloc_end + 1;
2858                                 continue;
2859                         }
2860                         ret = tree->ops->fill_delalloc(inode, page,
2861                                                        delalloc_start,
2862                                                        delalloc_end,
2863                                                        &page_started,
2864                                                        &nr_written);
2865                         /* File system has been set read-only */
2866                         if (ret) {
2867                                 SetPageError(page);
2868                                 goto done;
2869                         }
2870                         /*
2871                          * delalloc_end is already one less than the total
2872                          * length, so we don't subtract one from
2873                          * PAGE_CACHE_SIZE
2874                          */
2875                         delalloc_to_write += (delalloc_end - delalloc_start +
2876                                               PAGE_CACHE_SIZE) >>
2877                                               PAGE_CACHE_SHIFT;
2878                         delalloc_start = delalloc_end + 1;
2879                 }
2880                 if (wbc->nr_to_write < delalloc_to_write) {
2881                         int thresh = 8192;
2882
2883                         if (delalloc_to_write < thresh * 2)
2884                                 thresh = delalloc_to_write;
2885                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2886                                                  thresh);
2887                 }
2888
2889                 /* did the fill delalloc function already unlock and start
2890                  * the IO?
2891                  */
2892                 if (page_started) {
2893                         ret = 0;
2894                         /*
2895                          * we've unlocked the page, so we can't update
2896                          * the mapping's writeback index, just update
2897                          * nr_to_write.
2898                          */
2899                         wbc->nr_to_write -= nr_written;
2900                         goto done_unlocked;
2901                 }
2902         }
2903         if (tree->ops && tree->ops->writepage_start_hook) {
2904                 ret = tree->ops->writepage_start_hook(page, start,
2905                                                       page_end);
2906                 if (ret) {
2907                         /* Fixup worker will requeue */
2908                         if (ret == -EBUSY)
2909                                 wbc->pages_skipped++;
2910                         else
2911                                 redirty_page_for_writepage(wbc, page);
2912                         update_nr_written(page, wbc, nr_written);
2913                         unlock_page(page);
2914                         ret = 0;
2915                         goto done_unlocked;
2916                 }
2917         }
2918
2919         /*
2920          * we don't want to touch the inode after unlocking the page,
2921          * so we update the mapping writeback index now
2922          */
2923         update_nr_written(page, wbc, nr_written + 1);
2924
2925         end = page_end;
2926         if (last_byte <= start) {
2927                 if (tree->ops && tree->ops->writepage_end_io_hook)
2928                         tree->ops->writepage_end_io_hook(page, start,
2929                                                          page_end, NULL, 1);
2930                 goto done;
2931         }
2932
2933         blocksize = inode->i_sb->s_blocksize;
2934
2935         while (cur <= end) {
2936                 if (cur >= last_byte) {
2937                         if (tree->ops && tree->ops->writepage_end_io_hook)
2938                                 tree->ops->writepage_end_io_hook(page, cur,
2939                                                          page_end, NULL, 1);
2940                         break;
2941                 }
2942                 em = epd->get_extent(inode, page, pg_offset, cur,
2943                                      end - cur + 1, 1);
2944                 if (IS_ERR_OR_NULL(em)) {
2945                         SetPageError(page);
2946                         break;
2947                 }
2948
2949                 extent_offset = cur - em->start;
2950                 BUG_ON(extent_map_end(em) <= cur);
2951                 BUG_ON(end < cur);
2952                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2953                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2954                 sector = (em->block_start + extent_offset) >> 9;
2955                 bdev = em->bdev;
2956                 block_start = em->block_start;
2957                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2958                 free_extent_map(em);
2959                 em = NULL;
2960
2961                 /*
2962                  * compressed and inline extents are written through other
2963                  * paths in the FS
2964                  */
2965                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2966                     block_start == EXTENT_MAP_INLINE) {
2967                         /*
2968                          * end_io notification does not happen here for
2969                          * compressed extents
2970                          */
2971                         if (!compressed && tree->ops &&
2972                             tree->ops->writepage_end_io_hook)
2973                                 tree->ops->writepage_end_io_hook(page, cur,
2974                                                          cur + iosize - 1,
2975                                                          NULL, 1);
2976                         else if (compressed) {
2977                                 /* we don't want to end_page_writeback on
2978                                  * a compressed extent.  this happens
2979                                  * elsewhere
2980                                  */
2981                                 nr++;
2982                         }
2983
2984                         cur += iosize;
2985                         pg_offset += iosize;
2986                         continue;
2987                 }
2988                 /* leave this out until we have a page_mkwrite call */
2989                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2990                                    EXTENT_DIRTY, 0, NULL)) {
2991                         cur = cur + iosize;
2992                         pg_offset += iosize;
2993                         continue;
2994                 }
2995
2996                 if (tree->ops && tree->ops->writepage_io_hook) {
2997                         ret = tree->ops->writepage_io_hook(page, cur,
2998                                                 cur + iosize - 1);
2999                 } else {
3000                         ret = 0;
3001                 }
3002                 if (ret) {
3003                         SetPageError(page);
3004                 } else {
3005                         unsigned long max_nr = end_index + 1;
3006
3007                         set_range_writeback(tree, cur, cur + iosize - 1);
3008                         if (!PageWriteback(page)) {
3009                                 printk(KERN_ERR "btrfs warning page %lu not "
3010                                        "writeback, cur %llu end %llu\n",
3011                                        page->index, (unsigned long long)cur,
3012                                        (unsigned long long)end);
3013                         }
3014
3015                         ret = submit_extent_page(write_flags, tree, page,
3016                                                  sector, iosize, pg_offset,
3017                                                  bdev, &epd->bio, max_nr,
3018                                                  end_bio_extent_writepage,
3019                                                  0, 0, 0);
3020                         if (ret)
3021                                 SetPageError(page);
3022                 }
3023                 cur = cur + iosize;
3024                 pg_offset += iosize;
3025                 nr++;
3026         }
3027 done:
3028         if (nr == 0) {
3029                 /* make sure the mapping tag for page dirty gets cleared */
3030                 set_page_writeback(page);
3031                 end_page_writeback(page);
3032         }
3033         unlock_page(page);
3034
3035 done_unlocked:
3036
3037         /* drop our reference on any cached states */
3038         free_extent_state(cached_state);
3039         return 0;
3040 }
3041
3042 static int eb_wait(void *word)
3043 {
3044         io_schedule();
3045         return 0;
3046 }
3047
3048 static void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3049 {
3050         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3051                     TASK_UNINTERRUPTIBLE);
3052 }
3053
3054 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3055                                      struct btrfs_fs_info *fs_info,
3056                                      struct extent_page_data *epd)
3057 {
3058         unsigned long i, num_pages;
3059         int flush = 0;
3060         int ret = 0;
3061
3062         if (!btrfs_try_tree_write_lock(eb)) {
3063                 flush = 1;
3064                 flush_write_bio(epd);
3065                 btrfs_tree_lock(eb);
3066         }
3067
3068         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3069                 btrfs_tree_unlock(eb);
3070                 if (!epd->sync_io)
3071                         return 0;
3072                 if (!flush) {
3073                         flush_write_bio(epd);
3074                         flush = 1;
3075                 }
3076                 while (1) {
3077                         wait_on_extent_buffer_writeback(eb);
3078                         btrfs_tree_lock(eb);
3079                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3080                                 break;
3081                         btrfs_tree_unlock(eb);
3082                 }
3083         }
3084
3085         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3086                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3087                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3088                 spin_lock(&fs_info->delalloc_lock);
3089                 if (fs_info->dirty_metadata_bytes >= eb->len)
3090                         fs_info->dirty_metadata_bytes -= eb->len;
3091                 else
3092                         WARN_ON(1);
3093                 spin_unlock(&fs_info->delalloc_lock);
3094                 ret = 1;
3095         }
3096
3097         btrfs_tree_unlock(eb);
3098
3099         if (!ret)
3100                 return ret;
3101
3102         num_pages = num_extent_pages(eb->start, eb->len);
3103         for (i = 0; i < num_pages; i++) {
3104                 struct page *p = extent_buffer_page(eb, i);
3105
3106                 if (!trylock_page(p)) {
3107                         if (!flush) {
3108                                 flush_write_bio(epd);
3109                                 flush = 1;
3110                         }
3111                         lock_page(p);
3112                 }
3113         }
3114
3115         return ret;
3116 }
3117
3118 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3119 {
3120         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3121         smp_mb__after_clear_bit();
3122         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3123 }
3124
3125 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3126 {
3127         int uptodate = err == 0;
3128         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
3129         struct extent_buffer *eb;
3130         int done;
3131
3132         do {
3133                 struct page *page = bvec->bv_page;
3134
3135                 bvec--;
3136                 eb = (struct extent_buffer *)page->private;
3137                 BUG_ON(!eb);
3138                 done = atomic_dec_and_test(&eb->io_pages);
3139
3140                 if (!uptodate || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3141                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3142                         ClearPageUptodate(page);
3143                         SetPageError(page);
3144                 }
3145
3146                 end_page_writeback(page);
3147
3148                 if (!done)
3149                         continue;
3150
3151                 end_extent_buffer_writeback(eb);
3152         } while (bvec >= bio->bi_io_vec);
3153
3154         bio_put(bio);
3155
3156 }
3157
3158 static int write_one_eb(struct extent_buffer *eb,
3159                         struct btrfs_fs_info *fs_info,
3160                         struct writeback_control *wbc,
3161                         struct extent_page_data *epd)
3162 {
3163         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3164         u64 offset = eb->start;
3165         unsigned long i, num_pages;
3166         int rw = (epd->sync_io ? WRITE_SYNC : WRITE);
3167         int ret;
3168
3169         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3170         num_pages = num_extent_pages(eb->start, eb->len);
3171         atomic_set(&eb->io_pages, num_pages);
3172         for (i = 0; i < num_pages; i++) {
3173                 struct page *p = extent_buffer_page(eb, i);
3174
3175                 clear_page_dirty_for_io(p);
3176                 set_page_writeback(p);
3177                 ret = submit_extent_page(rw, eb->tree, p, offset >> 9,
3178                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3179                                          -1, end_bio_extent_buffer_writepage,
3180                                          0, 0, 0);
3181                 if (ret) {
3182                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3183                         SetPageError(p);
3184                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3185                                 end_extent_buffer_writeback(eb);
3186                         ret = -EIO;
3187                         break;
3188                 }
3189                 offset += PAGE_CACHE_SIZE;
3190                 update_nr_written(p, wbc, 1);
3191                 unlock_page(p);
3192         }
3193
3194         if (unlikely(ret)) {
3195                 for (; i < num_pages; i++) {
3196                         struct page *p = extent_buffer_page(eb, i);
3197                         unlock_page(p);
3198                 }
3199         }
3200
3201         return ret;
3202 }
3203
3204 int btree_write_cache_pages(struct address_space *mapping,
3205                                    struct writeback_control *wbc)
3206 {
3207         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3208         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3209         struct extent_buffer *eb, *prev_eb = NULL;
3210         struct extent_page_data epd = {
3211                 .bio = NULL,
3212                 .tree = tree,
3213                 .extent_locked = 0,
3214                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3215         };
3216         int ret = 0;
3217         int done = 0;
3218         int nr_to_write_done = 0;
3219         struct pagevec pvec;
3220         int nr_pages;
3221         pgoff_t index;
3222         pgoff_t end;            /* Inclusive */
3223         int scanned = 0;
3224         int tag;
3225
3226         pagevec_init(&pvec, 0);
3227         if (wbc->range_cyclic) {
3228                 index = mapping->writeback_index; /* Start from prev offset */
3229                 end = -1;
3230         } else {
3231                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3232                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3233                 scanned = 1;
3234         }
3235         if (wbc->sync_mode == WB_SYNC_ALL)
3236                 tag = PAGECACHE_TAG_TOWRITE;
3237         else
3238                 tag = PAGECACHE_TAG_DIRTY;
3239 retry:
3240         if (wbc->sync_mode == WB_SYNC_ALL)
3241                 tag_pages_for_writeback(mapping, index, end);
3242         while (!done && !nr_to_write_done && (index <= end) &&
3243                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3244                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3245                 unsigned i;
3246
3247                 scanned = 1;
3248                 for (i = 0; i < nr_pages; i++) {
3249                         struct page *page = pvec.pages[i];
3250
3251                         if (!PagePrivate(page))
3252                                 continue;
3253
3254                         if (!wbc->range_cyclic && page->index > end) {
3255                                 done = 1;
3256                                 break;
3257                         }
3258
3259                         eb = (struct extent_buffer *)page->private;
3260                         if (!eb) {
3261                                 WARN_ON(1);
3262                                 continue;
3263                         }
3264
3265                         if (eb == prev_eb)
3266                                 continue;
3267
3268                         if (!atomic_inc_not_zero(&eb->refs)) {
3269                                 WARN_ON(1);
3270                                 continue;
3271                         }
3272
3273                         prev_eb = eb;
3274                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3275                         if (!ret) {
3276                                 free_extent_buffer(eb);
3277                                 continue;
3278                         }
3279
3280                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3281                         if (ret) {
3282                                 done = 1;
3283                                 free_extent_buffer(eb);
3284                                 break;
3285                         }
3286                         free_extent_buffer(eb);
3287
3288                         /*
3289                          * the filesystem may choose to bump up nr_to_write.
3290                          * We have to make sure to honor the new nr_to_write
3291                          * at any time
3292                          */
3293                         nr_to_write_done = wbc->nr_to_write <= 0;
3294                 }
3295                 pagevec_release(&pvec);
3296                 cond_resched();
3297         }
3298         if (!scanned && !done) {
3299                 /*
3300                  * We hit the last page and there is more work to be done: wrap
3301                  * back to the start of the file
3302                  */
3303                 scanned = 1;
3304                 index = 0;
3305                 goto retry;
3306         }
3307         flush_write_bio(&epd);
3308         return ret;
3309 }
3310
3311 /**
3312  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3313  * @mapping: address space structure to write
3314  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3315  * @writepage: function called for each page
3316  * @data: data passed to writepage function
3317  *
3318  * If a page is already under I/O, write_cache_pages() skips it, even
3319  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3320  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3321  * and msync() need to guarantee that all the data which was dirty at the time
3322  * the call was made get new I/O started against them.  If wbc->sync_mode is
3323  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3324  * existing IO to complete.
3325  */
3326 static int extent_write_cache_pages(struct extent_io_tree *tree,
3327                              struct address_space *mapping,
3328                              struct writeback_control *wbc,
3329                              writepage_t writepage, void *data,
3330                              void (*flush_fn)(void *))
3331 {
3332         int ret = 0;
3333         int done = 0;
3334         int nr_to_write_done = 0;
3335         struct pagevec pvec;
3336         int nr_pages;
3337         pgoff_t index;
3338         pgoff_t end;            /* Inclusive */
3339         int scanned = 0;
3340         int tag;
3341
3342         pagevec_init(&pvec, 0);
3343         if (wbc->range_cyclic) {
3344                 index = mapping->writeback_index; /* Start from prev offset */
3345                 end = -1;
3346         } else {
3347                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3348                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3349                 scanned = 1;
3350         }
3351         if (wbc->sync_mode == WB_SYNC_ALL)
3352                 tag = PAGECACHE_TAG_TOWRITE;
3353         else
3354                 tag = PAGECACHE_TAG_DIRTY;
3355 retry:
3356         if (wbc->sync_mode == WB_SYNC_ALL)
3357                 tag_pages_for_writeback(mapping, index, end);
3358         while (!done && !nr_to_write_done && (index <= end) &&
3359                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3360                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3361                 unsigned i;
3362
3363                 scanned = 1;
3364                 for (i = 0; i < nr_pages; i++) {
3365                         struct page *page = pvec.pages[i];
3366
3367                         /*
3368                          * At this point we hold neither mapping->tree_lock nor
3369                          * lock on the page itself: the page may be truncated or
3370                          * invalidated (changing page->mapping to NULL), or even
3371                          * swizzled back from swapper_space to tmpfs file
3372                          * mapping
3373                          */
3374                         if (tree->ops &&
3375                             tree->ops->write_cache_pages_lock_hook) {
3376                                 tree->ops->write_cache_pages_lock_hook(page,
3377                                                                data, flush_fn);
3378                         } else {
3379                                 if (!trylock_page(page)) {
3380                                         flush_fn(data);
3381                                         lock_page(page);
3382                                 }
3383                         }
3384
3385                         if (unlikely(page->mapping != mapping)) {
3386                                 unlock_page(page);
3387                                 continue;
3388                         }
3389
3390                         if (!wbc->range_cyclic && page->index > end) {
3391                                 done = 1;
3392                                 unlock_page(page);
3393                                 continue;
3394                         }
3395
3396                         if (wbc->sync_mode != WB_SYNC_NONE) {
3397                                 if (PageWriteback(page))
3398                                         flush_fn(data);
3399                                 wait_on_page_writeback(page);
3400                         }
3401
3402                         if (PageWriteback(page) ||
3403                             !clear_page_dirty_for_io(page)) {
3404                                 unlock_page(page);
3405                                 continue;
3406                         }
3407
3408                         ret = (*writepage)(page, wbc, data);
3409
3410                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3411                                 unlock_page(page);
3412                                 ret = 0;
3413                         }
3414                         if (ret)
3415                                 done = 1;
3416
3417                         /*
3418                          * the filesystem may choose to bump up nr_to_write.
3419                          * We have to make sure to honor the new nr_to_write
3420                          * at any time
3421                          */
3422                         nr_to_write_done = wbc->nr_to_write <= 0;
3423                 }
3424                 pagevec_release(&pvec);
3425                 cond_resched();
3426         }
3427         if (!scanned && !done) {
3428                 /*
3429                  * We hit the last page and there is more work to be done: wrap
3430                  * back to the start of the file
3431                  */
3432                 scanned = 1;
3433                 index = 0;
3434                 goto retry;
3435         }
3436         return ret;
3437 }
3438
3439 static void flush_epd_write_bio(struct extent_page_data *epd)
3440 {
3441         if (epd->bio) {
3442                 int rw = WRITE;
3443                 int ret;
3444
3445                 if (epd->sync_io)
3446                         rw = WRITE_SYNC;
3447
3448                 ret = submit_one_bio(rw, epd->bio, 0, 0);
3449                 BUG_ON(ret < 0); /* -ENOMEM */
3450                 epd->bio = NULL;
3451         }
3452 }
3453
3454 static noinline void flush_write_bio(void *data)
3455 {
3456         struct extent_page_data *epd = data;
3457         flush_epd_write_bio(epd);
3458 }
3459
3460 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3461                           get_extent_t *get_extent,
3462                           struct writeback_control *wbc)
3463 {
3464         int ret;
3465         struct extent_page_data epd = {
3466                 .bio = NULL,
3467                 .tree = tree,
3468                 .get_extent = get_extent,
3469                 .extent_locked = 0,
3470                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3471         };
3472
3473         ret = __extent_writepage(page, wbc, &epd);
3474
3475         flush_epd_write_bio(&epd);
3476         return ret;
3477 }
3478
3479 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3480                               u64 start, u64 end, get_extent_t *get_extent,
3481                               int mode)
3482 {
3483         int ret = 0;
3484         struct address_space *mapping = inode->i_mapping;
3485         struct page *page;
3486         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3487                 PAGE_CACHE_SHIFT;
3488
3489         struct extent_page_data epd = {
3490                 .bio = NULL,
3491                 .tree = tree,
3492                 .get_extent = get_extent,
3493                 .extent_locked = 1,
3494                 .sync_io = mode == WB_SYNC_ALL,
3495         };
3496         struct writeback_control wbc_writepages = {
3497                 .sync_mode      = mode,
3498                 .nr_to_write    = nr_pages * 2,
3499                 .range_start    = start,
3500                 .range_end      = end + 1,
3501         };
3502
3503         while (start <= end) {
3504                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3505                 if (clear_page_dirty_for_io(page))
3506                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3507                 else {
3508                         if (tree->ops && tree->ops->writepage_end_io_hook)
3509                                 tree->ops->writepage_end_io_hook(page, start,
3510                                                  start + PAGE_CACHE_SIZE - 1,
3511                                                  NULL, 1);
3512                         unlock_page(page);
3513                 }
3514                 page_cache_release(page);
3515                 start += PAGE_CACHE_SIZE;
3516         }
3517
3518         flush_epd_write_bio(&epd);
3519         return ret;
3520 }
3521
3522 int extent_writepages(struct extent_io_tree *tree,
3523                       struct address_space *mapping,
3524                       get_extent_t *get_extent,
3525                       struct writeback_control *wbc)
3526 {
3527         int ret = 0;
3528         struct extent_page_data epd = {
3529                 .bio = NULL,
3530                 .tree = tree,
3531                 .get_extent = get_extent,
3532                 .extent_locked = 0,
3533                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3534         };
3535
3536         ret = extent_write_cache_pages(tree, mapping, wbc,
3537                                        __extent_writepage, &epd,
3538                                        flush_write_bio);
3539         flush_epd_write_bio(&epd);
3540         return ret;
3541 }
3542
3543 int extent_readpages(struct extent_io_tree *tree,
3544                      struct address_space *mapping,
3545                      struct list_head *pages, unsigned nr_pages,
3546                      get_extent_t get_extent)
3547 {
3548         struct bio *bio = NULL;
3549         unsigned page_idx;
3550         unsigned long bio_flags = 0;
3551
3552         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3553                 struct page *page = list_entry(pages->prev, struct page, lru);
3554
3555                 prefetchw(&page->flags);
3556                 list_del(&page->lru);
3557                 if (!add_to_page_cache_lru(page, mapping,
3558                                         page->index, GFP_NOFS)) {
3559                         __extent_read_full_page(tree, page, get_extent,
3560                                                 &bio, 0, &bio_flags);
3561                 }
3562                 page_cache_release(page);
3563         }
3564         BUG_ON(!list_empty(pages));
3565         if (bio)
3566                 return submit_one_bio(READ, bio, 0, bio_flags);
3567         return 0;
3568 }
3569
3570 /*
3571  * basic invalidatepage code, this waits on any locked or writeback
3572  * ranges corresponding to the page, and then deletes any extent state
3573  * records from the tree
3574  */
3575 int extent_invalidatepage(struct extent_io_tree *tree,
3576                           struct page *page, unsigned long offset)
3577 {
3578         struct extent_state *cached_state = NULL;
3579         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
3580         u64 end = start + PAGE_CACHE_SIZE - 1;
3581         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3582
3583         start += (offset + blocksize - 1) & ~(blocksize - 1);
3584         if (start > end)
3585                 return 0;
3586
3587         lock_extent_bits(tree, start, end, 0, &cached_state);
3588         wait_on_page_writeback(page);
3589         clear_extent_bit(tree, start, end,
3590                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3591                          EXTENT_DO_ACCOUNTING,
3592                          1, 1, &cached_state, GFP_NOFS);
3593         return 0;
3594 }
3595
3596 /*
3597  * a helper for releasepage, this tests for areas of the page that
3598  * are locked or under IO and drops the related state bits if it is safe
3599  * to drop the page.
3600  */
3601 int try_release_extent_state(struct extent_map_tree *map,
3602                              struct extent_io_tree *tree, struct page *page,
3603                              gfp_t mask)
3604 {
3605         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3606         u64 end = start + PAGE_CACHE_SIZE - 1;
3607         int ret = 1;
3608
3609         if (test_range_bit(tree, start, end,
3610                            EXTENT_IOBITS, 0, NULL))
3611                 ret = 0;
3612         else {
3613                 if ((mask & GFP_NOFS) == GFP_NOFS)
3614                         mask = GFP_NOFS;
3615                 /*
3616                  * at this point we can safely clear everything except the
3617                  * locked bit and the nodatasum bit
3618                  */
3619                 ret = clear_extent_bit(tree, start, end,
3620                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3621                                  0, 0, NULL, mask);
3622
3623                 /* if clear_extent_bit failed for enomem reasons,
3624                  * we can't allow the release to continue.
3625                  */
3626                 if (ret < 0)
3627                         ret = 0;
3628                 else
3629                         ret = 1;
3630         }
3631         return ret;
3632 }
3633
3634 /*
3635  * a helper for releasepage.  As long as there are no locked extents
3636  * in the range corresponding to the page, both state records and extent
3637  * map records are removed
3638  */
3639 int try_release_extent_mapping(struct extent_map_tree *map,
3640                                struct extent_io_tree *tree, struct page *page,
3641                                gfp_t mask)
3642 {
3643         struct extent_map *em;
3644         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
3645         u64 end = start + PAGE_CACHE_SIZE - 1;
3646
3647         if ((mask & __GFP_WAIT) &&
3648             page->mapping->host->i_size > 16 * 1024 * 1024) {
3649                 u64 len;
3650                 while (start <= end) {
3651                         len = end - start + 1;
3652                         write_lock(&map->lock);
3653                         em = lookup_extent_mapping(map, start, len);
3654                         if (!em) {
3655                                 write_unlock(&map->lock);
3656                                 break;
3657                         }
3658                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3659                             em->start != start) {
3660                                 write_unlock(&map->lock);
3661                                 free_extent_map(em);
3662                                 break;
3663                         }
3664                         if (!test_range_bit(tree, em->start,
3665                                             extent_map_end(em) - 1,
3666                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
3667                                             0, NULL)) {
3668                                 remove_extent_mapping(map, em);
3669                                 /* once for the rb tree */
3670                                 free_extent_map(em);
3671                         }
3672                         start = extent_map_end(em);
3673                         write_unlock(&map->lock);
3674
3675                         /* once for us */
3676                         free_extent_map(em);
3677                 }
3678         }
3679         return try_release_extent_state(map, tree, page, mask);
3680 }
3681
3682 /*
3683  * helper function for fiemap, which doesn't want to see any holes.
3684  * This maps until we find something past 'last'
3685  */
3686 static struct extent_map *get_extent_skip_holes(struct inode *inode,
3687                                                 u64 offset,
3688                                                 u64 last,
3689                                                 get_extent_t *get_extent)
3690 {
3691         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
3692         struct extent_map *em;
3693         u64 len;
3694
3695         if (offset >= last)
3696                 return NULL;
3697
3698         while(1) {
3699                 len = last - offset;
3700                 if (len == 0)
3701                         break;
3702                 len = (len + sectorsize - 1) & ~(sectorsize - 1);
3703                 em = get_extent(inode, NULL, 0, offset, len, 0);
3704                 if (IS_ERR_OR_NULL(em))
3705                         return em;
3706
3707                 /* if this isn't a hole return it */
3708                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
3709                     em->block_start != EXTENT_MAP_HOLE) {
3710                         return em;
3711                 }
3712
3713                 /* this is a hole, advance to the next extent */
3714                 offset = extent_map_end(em);
3715                 free_extent_map(em);
3716                 if (offset >= last)
3717                         break;
3718         }
3719         return NULL;
3720 }
3721
3722 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
3723                 __u64 start, __u64 len, get_extent_t *get_extent)
3724 {
3725         int ret = 0;
3726         u64 off = start;
3727         u64 max = start + len;
3728         u32 flags = 0;
3729         u32 found_type;
3730         u64 last;
3731         u64 last_for_get_extent = 0;
3732         u64 disko = 0;
3733         u64 isize = i_size_read(inode);
3734         struct btrfs_key found_key;
3735         struct extent_map *em = NULL;
3736         struct extent_state *cached_state = NULL;
3737         struct btrfs_path *path;
3738         struct btrfs_file_extent_item *item;
3739         int end = 0;
3740         u64 em_start = 0;
3741         u64 em_len = 0;
3742         u64 em_end = 0;
3743         unsigned long emflags;
3744
3745         if (len == 0)
3746                 return -EINVAL;
3747
3748         path = btrfs_alloc_path();
3749         if (!path)
3750                 return -ENOMEM;
3751         path->leave_spinning = 1;
3752
3753         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
3754         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
3755
3756         /*
3757          * lookup the last file extent.  We're not using i_size here
3758          * because there might be preallocation past i_size
3759          */
3760         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3761                                        path, btrfs_ino(inode), -1, 0);
3762         if (ret < 0) {
3763                 btrfs_free_path(path);
3764                 return ret;
3765         }
3766         WARN_ON(!ret);
3767         path->slots[0]--;
3768         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3769                               struct btrfs_file_extent_item);
3770         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3771         found_type = btrfs_key_type(&found_key);
3772
3773         /* No extents, but there might be delalloc bits */
3774         if (found_key.objectid != btrfs_ino(inode) ||
3775             found_type != BTRFS_EXTENT_DATA_KEY) {
3776                 /* have to trust i_size as the end */
3777                 last = (u64)-1;
3778                 last_for_get_extent = isize;
3779         } else {
3780                 /*
3781                  * remember the start of the last extent.  There are a
3782                  * bunch of different factors that go into the length of the
3783                  * extent, so its much less complex to remember where it started
3784                  */
3785                 last = found_key.offset;
3786                 last_for_get_extent = last + 1;
3787         }
3788         btrfs_free_path(path);
3789
3790         /*
3791          * we might have some extents allocated but more delalloc past those
3792          * extents.  so, we trust isize unless the start of the last extent is
3793          * beyond isize
3794          */
3795         if (last < isize) {
3796                 last = (u64)-1;
3797                 last_for_get_extent = isize;
3798         }
3799
3800         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3801                          &cached_state);
3802
3803         em = get_extent_skip_holes(inode, start, last_for_get_extent,
3804                                    get_extent);
3805         if (!em)
3806                 goto out;
3807         if (IS_ERR(em)) {
3808                 ret = PTR_ERR(em);
3809                 goto out;
3810         }
3811
3812         while (!end) {
3813                 u64 offset_in_extent;
3814
3815                 /* break if the extent we found is outside the range */
3816                 if (em->start >= max || extent_map_end(em) < off)
3817                         break;
3818
3819                 /*
3820                  * get_extent may return an extent that starts before our
3821                  * requested range.  We have to make sure the ranges
3822                  * we return to fiemap always move forward and don't
3823                  * overlap, so adjust the offsets here
3824                  */
3825                 em_start = max(em->start, off);
3826
3827                 /*
3828                  * record the offset from the start of the extent
3829                  * for adjusting the disk offset below
3830                  */
3831                 offset_in_extent = em_start - em->start;
3832                 em_end = extent_map_end(em);
3833                 em_len = em_end - em_start;
3834                 emflags = em->flags;
3835                 disko = 0;
3836                 flags = 0;
3837
3838                 /*
3839                  * bump off for our next call to get_extent
3840                  */
3841                 off = extent_map_end(em);
3842                 if (off >= max)
3843                         end = 1;
3844
3845                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3846                         end = 1;
3847                         flags |= FIEMAP_EXTENT_LAST;
3848                 } else if (em->block_start == EXTENT_MAP_INLINE) {
3849                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
3850                                   FIEMAP_EXTENT_NOT_ALIGNED);
3851                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3852                         flags |= (FIEMAP_EXTENT_DELALLOC |
3853                                   FIEMAP_EXTENT_UNKNOWN);
3854                 } else {
3855                         disko = em->block_start + offset_in_extent;
3856                 }
3857                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3858                         flags |= FIEMAP_EXTENT_ENCODED;
3859
3860                 free_extent_map(em);
3861                 em = NULL;
3862                 if ((em_start >= last) || em_len == (u64)-1 ||
3863                    (last == (u64)-1 && isize <= em_end)) {
3864                         flags |= FIEMAP_EXTENT_LAST;
3865                         end = 1;
3866                 }
3867
3868                 /* now scan forward to see if this is really the last extent. */
3869                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
3870                                            get_extent);
3871                 if (IS_ERR(em)) {
3872                         ret = PTR_ERR(em);
3873                         goto out;
3874                 }
3875                 if (!em) {
3876                         flags |= FIEMAP_EXTENT_LAST;
3877                         end = 1;
3878                 }
3879                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3880                                               em_len, flags);
3881                 if (ret)
3882                         goto out_free;
3883         }
3884 out_free:
3885         free_extent_map(em);
3886 out:
3887         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3888                              &cached_state, GFP_NOFS);
3889         return ret;
3890 }
3891
3892 inline struct page *extent_buffer_page(struct extent_buffer *eb,
3893                                               unsigned long i)
3894 {
3895         return eb->pages[i];
3896 }
3897
3898 inline unsigned long num_extent_pages(u64 start, u64 len)
3899 {
3900         return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
3901                 (start >> PAGE_CACHE_SHIFT);
3902 }
3903
3904 static void __free_extent_buffer(struct extent_buffer *eb)
3905 {
3906 #if LEAK_DEBUG
3907         unsigned long flags;
3908         spin_lock_irqsave(&leak_lock, flags);
3909         list_del(&eb->leak_list);
3910         spin_unlock_irqrestore(&leak_lock, flags);
3911 #endif
3912         if (eb->pages && eb->pages != eb->inline_pages)
3913                 kfree(eb->pages);
3914         kmem_cache_free(extent_buffer_cache, eb);
3915 }
3916
3917 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3918                                                    u64 start,
3919                                                    unsigned long len,
3920                                                    gfp_t mask)
3921 {
3922         struct extent_buffer *eb = NULL;
3923 #if LEAK_DEBUG
3924         unsigned long flags;
3925 #endif
3926
3927         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
3928         if (eb == NULL)
3929                 return NULL;
3930         eb->start = start;
3931         eb->len = len;
3932         eb->tree = tree;
3933         rwlock_init(&eb->lock);
3934         atomic_set(&eb->write_locks, 0);
3935         atomic_set(&eb->read_locks, 0);
3936         atomic_set(&eb->blocking_readers, 0);
3937         atomic_set(&eb->blocking_writers, 0);
3938         atomic_set(&eb->spinning_readers, 0);
3939         atomic_set(&eb->spinning_writers, 0);
3940         eb->lock_nested = 0;
3941         init_waitqueue_head(&eb->write_lock_wq);
3942         init_waitqueue_head(&eb->read_lock_wq);
3943
3944 #if LEAK_DEBUG
3945         spin_lock_irqsave(&leak_lock, flags);
3946         list_add(&eb->leak_list, &buffers);
3947         spin_unlock_irqrestore(&leak_lock, flags);
3948 #endif
3949         spin_lock_init(&eb->refs_lock);
3950         atomic_set(&eb->refs, 1);
3951         atomic_set(&eb->io_pages, 0);
3952
3953         if (len > MAX_INLINE_EXTENT_BUFFER_SIZE) {
3954                 struct page **pages;
3955                 int num_pages = (len + PAGE_CACHE_SIZE - 1) >>
3956                         PAGE_CACHE_SHIFT;
3957                 pages = kzalloc(num_pages, mask);
3958                 if (!pages) {
3959                         __free_extent_buffer(eb);
3960                         return NULL;
3961                 }
3962                 eb->pages = pages;
3963         } else {
3964                 eb->pages = eb->inline_pages;
3965         }
3966
3967         return eb;
3968 }
3969
3970 static int extent_buffer_under_io(struct extent_buffer *eb)
3971 {
3972         return (atomic_read(&eb->io_pages) ||
3973                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
3974                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3975 }
3976
3977 /*
3978  * Helper for releasing extent buffer page.
3979  */
3980 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3981                                                 unsigned long start_idx)
3982 {
3983         unsigned long index;
3984         struct page *page;
3985
3986         BUG_ON(extent_buffer_under_io(eb));
3987
3988         index = num_extent_pages(eb->start, eb->len);
3989         if (start_idx >= index)
3990                 return;
3991
3992         do {
3993                 index--;
3994                 page = extent_buffer_page(eb, index);
3995                 if (page) {
3996                         spin_lock(&page->mapping->private_lock);
3997                         /*
3998                          * We do this since we'll remove the pages after we've
3999                          * removed the eb from the radix tree, so we could race
4000                          * and have this page now attached to the new eb.  So
4001                          * only clear page_private if it's still connected to
4002                          * this eb.
4003                          */
4004                         if (PagePrivate(page) &&
4005                             page->private == (unsigned long)eb) {
4006                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4007                                 BUG_ON(PageDirty(page));
4008                                 BUG_ON(PageWriteback(page));
4009                                 /*
4010                                  * We need to make sure we haven't be attached
4011                                  * to a new eb.
4012                                  */
4013                                 ClearPagePrivate(page);
4014                                 set_page_private(page, 0);
4015                                 /* One for the page private */
4016                                 page_cache_release(page);
4017                         }
4018                         spin_unlock(&page->mapping->private_lock);
4019
4020                         /* One for when we alloced the page */
4021                         page_cache_release(page);
4022                 }
4023         } while (index != start_idx);
4024 }
4025
4026 /*
4027  * Helper for releasing the extent buffer.
4028  */
4029 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4030 {
4031         btrfs_release_extent_buffer_page(eb, 0);
4032         __free_extent_buffer(eb);
4033 }
4034
4035 static void check_buffer_tree_ref(struct extent_buffer *eb)
4036 {
4037         /* the ref bit is tricky.  We have to make sure it is set
4038          * if we have the buffer dirty.   Otherwise the
4039          * code to free a buffer can end up dropping a dirty
4040          * page
4041          *
4042          * Once the ref bit is set, it won't go away while the
4043          * buffer is dirty or in writeback, and it also won't
4044          * go away while we have the reference count on the
4045          * eb bumped.
4046          *
4047          * We can't just set the ref bit without bumping the
4048          * ref on the eb because free_extent_buffer might
4049          * see the ref bit and try to clear it.  If this happens
4050          * free_extent_buffer might end up dropping our original
4051          * ref by mistake and freeing the page before we are able
4052          * to add one more ref.
4053          *
4054          * So bump the ref count first, then set the bit.  If someone
4055          * beat us to it, drop the ref we added.
4056          */
4057         if (!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4058                 atomic_inc(&eb->refs);
4059                 if (test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4060                         atomic_dec(&eb->refs);
4061         }
4062 }
4063
4064 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4065 {
4066         unsigned long num_pages, i;
4067
4068         check_buffer_tree_ref(eb);
4069
4070         num_pages = num_extent_pages(eb->start, eb->len);
4071         for (i = 0; i < num_pages; i++) {
4072                 struct page *p = extent_buffer_page(eb, i);
4073                 mark_page_accessed(p);
4074         }
4075 }
4076
4077 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
4078                                           u64 start, unsigned long len)
4079 {
4080         unsigned long num_pages = num_extent_pages(start, len);
4081         unsigned long i;
4082         unsigned long index = start >> PAGE_CACHE_SHIFT;
4083         struct extent_buffer *eb;
4084         struct extent_buffer *exists = NULL;
4085         struct page *p;
4086         struct address_space *mapping = tree->mapping;
4087         int uptodate = 1;
4088         int ret;
4089
4090         rcu_read_lock();
4091         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4092         if (eb && atomic_inc_not_zero(&eb->refs)) {
4093                 rcu_read_unlock();
4094                 mark_extent_buffer_accessed(eb);
4095                 return eb;
4096         }
4097         rcu_read_unlock();
4098
4099         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
4100         if (!eb)
4101                 return NULL;
4102
4103         for (i = 0; i < num_pages; i++, index++) {
4104                 p = find_or_create_page(mapping, index, GFP_NOFS);
4105                 if (!p) {
4106                         WARN_ON(1);
4107                         goto free_eb;
4108                 }
4109
4110                 spin_lock(&mapping->private_lock);
4111                 if (PagePrivate(p)) {
4112                         /*
4113                          * We could have already allocated an eb for this page
4114                          * and attached one so lets see if we can get a ref on
4115                          * the existing eb, and if we can we know it's good and
4116                          * we can just return that one, else we know we can just
4117                          * overwrite page->private.
4118                          */
4119                         exists = (struct extent_buffer *)p->private;
4120                         if (atomic_inc_not_zero(&exists->refs)) {
4121                                 spin_unlock(&mapping->private_lock);
4122                                 unlock_page(p);
4123                                 page_cache_release(p);
4124                                 mark_extent_buffer_accessed(exists);
4125                                 goto free_eb;
4126                         }
4127
4128                         /*
4129                          * Do this so attach doesn't complain and we need to
4130                          * drop the ref the old guy had.
4131                          */
4132                         ClearPagePrivate(p);
4133                         WARN_ON(PageDirty(p));
4134                         page_cache_release(p);
4135                 }
4136                 attach_extent_buffer_page(eb, p);
4137                 spin_unlock(&mapping->private_lock);
4138                 WARN_ON(PageDirty(p));
4139                 mark_page_accessed(p);
4140                 eb->pages[i] = p;
4141                 if (!PageUptodate(p))
4142                         uptodate = 0;
4143
4144                 /*
4145                  * see below about how we avoid a nasty race with release page
4146                  * and why we unlock later
4147                  */
4148         }
4149         if (uptodate)
4150                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4151 again:
4152         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4153         if (ret)
4154                 goto free_eb;
4155
4156         spin_lock(&tree->buffer_lock);
4157         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
4158         if (ret == -EEXIST) {
4159                 exists = radix_tree_lookup(&tree->buffer,
4160                                                 start >> PAGE_CACHE_SHIFT);
4161                 if (!atomic_inc_not_zero(&exists->refs)) {
4162                         spin_unlock(&tree->buffer_lock);
4163                         radix_tree_preload_end();
4164                         exists = NULL;
4165                         goto again;
4166                 }
4167                 spin_unlock(&tree->buffer_lock);
4168                 radix_tree_preload_end();
4169                 mark_extent_buffer_accessed(exists);
4170                 goto free_eb;
4171         }
4172         /* add one reference for the tree */
4173         spin_lock(&eb->refs_lock);
4174         check_buffer_tree_ref(eb);
4175         spin_unlock(&eb->refs_lock);
4176         spin_unlock(&tree->buffer_lock);
4177         radix_tree_preload_end();
4178
4179         /*
4180          * there is a race where release page may have
4181          * tried to find this extent buffer in the radix
4182          * but failed.  It will tell the VM it is safe to
4183          * reclaim the, and it will clear the page private bit.
4184          * We must make sure to set the page private bit properly
4185          * after the extent buffer is in the radix tree so
4186          * it doesn't get lost
4187          */
4188         SetPageChecked(eb->pages[0]);
4189         for (i = 1; i < num_pages; i++) {
4190                 p = extent_buffer_page(eb, i);
4191                 ClearPageChecked(p);
4192                 unlock_page(p);
4193         }
4194         unlock_page(eb->pages[0]);
4195         return eb;
4196
4197 free_eb:
4198         for (i = 0; i < num_pages; i++) {
4199                 if (eb->pages[i])
4200                         unlock_page(eb->pages[i]);
4201         }
4202
4203         WARN_ON(!atomic_dec_and_test(&eb->refs));
4204         btrfs_release_extent_buffer(eb);
4205         return exists;
4206 }
4207
4208 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
4209                                          u64 start, unsigned long len)
4210 {
4211         struct extent_buffer *eb;
4212
4213         rcu_read_lock();
4214         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4215         if (eb && atomic_inc_not_zero(&eb->refs)) {
4216                 rcu_read_unlock();
4217                 mark_extent_buffer_accessed(eb);
4218                 return eb;
4219         }
4220         rcu_read_unlock();
4221
4222         return NULL;
4223 }
4224
4225 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4226 {
4227         struct extent_buffer *eb =
4228                         container_of(head, struct extent_buffer, rcu_head);
4229
4230         __free_extent_buffer(eb);
4231 }
4232
4233 /* Expects to have eb->eb_lock already held */
4234 static void release_extent_buffer(struct extent_buffer *eb, gfp_t mask)
4235 {
4236         WARN_ON(atomic_read(&eb->refs) == 0);
4237         if (atomic_dec_and_test(&eb->refs)) {
4238                 struct extent_io_tree *tree = eb->tree;
4239
4240                 spin_unlock(&eb->refs_lock);
4241
4242                 spin_lock(&tree->buffer_lock);
4243                 radix_tree_delete(&tree->buffer,
4244                                   eb->start >> PAGE_CACHE_SHIFT);
4245                 spin_unlock(&tree->buffer_lock);
4246
4247                 /* Should be safe to release our pages at this point */
4248                 btrfs_release_extent_buffer_page(eb, 0);
4249
4250                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4251                 return;
4252         }
4253         spin_unlock(&eb->refs_lock);
4254 }
4255
4256 void free_extent_buffer(struct extent_buffer *eb)
4257 {
4258         if (!eb)
4259                 return;
4260
4261         spin_lock(&eb->refs_lock);
4262         if (atomic_read(&eb->refs) == 2 &&
4263             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4264             !extent_buffer_under_io(eb) &&
4265             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4266                 atomic_dec(&eb->refs);
4267
4268         /*
4269          * I know this is terrible, but it's temporary until we stop tracking
4270          * the uptodate bits and such for the extent buffers.
4271          */
4272         release_extent_buffer(eb, GFP_ATOMIC);
4273 }
4274
4275 void free_extent_buffer_stale(struct extent_buffer *eb)
4276 {
4277         if (!eb)
4278                 return;
4279
4280         spin_lock(&eb->refs_lock);
4281         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4282
4283         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4284             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4285                 atomic_dec(&eb->refs);
4286         release_extent_buffer(eb, GFP_NOFS);
4287 }
4288
4289 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4290 {
4291         unsigned long i;
4292         unsigned long num_pages;
4293         struct page *page;
4294
4295         num_pages = num_extent_pages(eb->start, eb->len);
4296
4297         for (i = 0; i < num_pages; i++) {
4298                 page = extent_buffer_page(eb, i);
4299                 if (!PageDirty(page))
4300                         continue;
4301
4302                 lock_page(page);
4303                 WARN_ON(!PagePrivate(page));
4304
4305                 clear_page_dirty_for_io(page);
4306                 spin_lock_irq(&page->mapping->tree_lock);
4307                 if (!PageDirty(page)) {
4308                         radix_tree_tag_clear(&page->mapping->page_tree,
4309                                                 page_index(page),
4310                                                 PAGECACHE_TAG_DIRTY);
4311                 }
4312                 spin_unlock_irq(&page->mapping->tree_lock);
4313                 ClearPageError(page);
4314                 unlock_page(page);
4315         }
4316         WARN_ON(atomic_read(&eb->refs) == 0);
4317 }
4318
4319 int set_extent_buffer_dirty(struct extent_buffer *eb)
4320 {
4321         unsigned long i;
4322         unsigned long num_pages;
4323         int was_dirty = 0;
4324
4325         check_buffer_tree_ref(eb);
4326
4327         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4328
4329         num_pages = num_extent_pages(eb->start, eb->len);
4330         WARN_ON(atomic_read(&eb->refs) == 0);
4331         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4332
4333         for (i = 0; i < num_pages; i++)
4334                 set_page_dirty(extent_buffer_page(eb, i));
4335         return was_dirty;
4336 }
4337
4338 static int range_straddles_pages(u64 start, u64 len)
4339 {
4340         if (len < PAGE_CACHE_SIZE)
4341                 return 1;
4342         if (start & (PAGE_CACHE_SIZE - 1))
4343                 return 1;
4344         if ((start + len) & (PAGE_CACHE_SIZE - 1))
4345                 return 1;
4346         return 0;
4347 }
4348
4349 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4350 {
4351         unsigned long i;
4352         struct page *page;
4353         unsigned long num_pages;
4354
4355         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4356         num_pages = num_extent_pages(eb->start, eb->len);
4357         for (i = 0; i < num_pages; i++) {
4358                 page = extent_buffer_page(eb, i);
4359                 if (page)
4360                         ClearPageUptodate(page);
4361         }
4362         return 0;
4363 }
4364
4365 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4366 {
4367         unsigned long i;
4368         struct page *page;
4369         unsigned long num_pages;
4370
4371         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4372         num_pages = num_extent_pages(eb->start, eb->len);
4373         for (i = 0; i < num_pages; i++) {
4374                 page = extent_buffer_page(eb, i);
4375                 SetPageUptodate(page);
4376         }
4377         return 0;
4378 }
4379
4380 int extent_range_uptodate(struct extent_io_tree *tree,
4381                           u64 start, u64 end)
4382 {
4383         struct page *page;
4384         int ret;
4385         int pg_uptodate = 1;
4386         int uptodate;
4387         unsigned long index;
4388
4389         if (range_straddles_pages(start, end - start + 1)) {
4390                 ret = test_range_bit(tree, start, end,
4391                                      EXTENT_UPTODATE, 1, NULL);
4392                 if (ret)
4393                         return 1;
4394         }
4395         while (start <= end) {
4396                 index = start >> PAGE_CACHE_SHIFT;
4397                 page = find_get_page(tree->mapping, index);
4398                 if (!page)
4399                         return 1;
4400                 uptodate = PageUptodate(page);
4401                 page_cache_release(page);
4402                 if (!uptodate) {
4403                         pg_uptodate = 0;
4404                         break;
4405                 }
4406                 start += PAGE_CACHE_SIZE;
4407         }
4408         return pg_uptodate;
4409 }
4410
4411 int extent_buffer_uptodate(struct extent_buffer *eb)
4412 {
4413         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4414 }
4415
4416 int read_extent_buffer_pages(struct extent_io_tree *tree,
4417                              struct extent_buffer *eb, u64 start, int wait,
4418                              get_extent_t *get_extent, int mirror_num)
4419 {
4420         unsigned long i;
4421         unsigned long start_i;
4422         struct page *page;
4423         int err;
4424         int ret = 0;
4425         int locked_pages = 0;
4426         int all_uptodate = 1;
4427         unsigned long num_pages;
4428         unsigned long num_reads = 0;
4429         struct bio *bio = NULL;
4430         unsigned long bio_flags = 0;
4431
4432         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4433                 return 0;
4434
4435         if (start) {
4436                 WARN_ON(start < eb->start);
4437                 start_i = (start >> PAGE_CACHE_SHIFT) -
4438                         (eb->start >> PAGE_CACHE_SHIFT);
4439         } else {
4440                 start_i = 0;
4441         }
4442
4443         num_pages = num_extent_pages(eb->start, eb->len);
4444         for (i = start_i; i < num_pages; i++) {
4445                 page = extent_buffer_page(eb, i);
4446                 if (wait == WAIT_NONE) {
4447                         if (!trylock_page(page))
4448                                 goto unlock_exit;
4449                 } else {
4450                         lock_page(page);
4451                 }
4452                 locked_pages++;
4453                 if (!PageUptodate(page)) {
4454                         num_reads++;
4455                         all_uptodate = 0;
4456                 }
4457         }
4458         if (all_uptodate) {
4459                 if (start_i == 0)
4460                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4461                 goto unlock_exit;
4462         }
4463
4464         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4465         eb->read_mirror = 0;
4466         atomic_set(&eb->io_pages, num_reads);
4467         for (i = start_i; i < num_pages; i++) {
4468                 page = extent_buffer_page(eb, i);
4469                 if (!PageUptodate(page)) {
4470                         ClearPageError(page);
4471                         err = __extent_read_full_page(tree, page,
4472                                                       get_extent, &bio,
4473                                                       mirror_num, &bio_flags);
4474                         if (err)
4475                                 ret = err;
4476                 } else {
4477                         unlock_page(page);
4478                 }
4479         }
4480
4481         if (bio) {
4482                 err = submit_one_bio(READ, bio, mirror_num, bio_flags);
4483                 if (err)
4484                         return err;
4485         }
4486
4487         if (ret || wait != WAIT_COMPLETE)
4488                 return ret;
4489
4490         for (i = start_i; i < num_pages; i++) {
4491                 page = extent_buffer_page(eb, i);
4492                 wait_on_page_locked(page);
4493                 if (!PageUptodate(page))
4494                         ret = -EIO;
4495         }
4496
4497         return ret;
4498
4499 unlock_exit:
4500         i = start_i;
4501         while (locked_pages > 0) {
4502                 page = extent_buffer_page(eb, i);
4503                 i++;
4504                 unlock_page(page);
4505                 locked_pages--;
4506         }
4507         return ret;
4508 }
4509
4510 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4511                         unsigned long start,
4512                         unsigned long len)
4513 {
4514         size_t cur;
4515         size_t offset;
4516         struct page *page;
4517         char *kaddr;
4518         char *dst = (char *)dstv;
4519         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4520         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4521
4522         WARN_ON(start > eb->len);
4523         WARN_ON(start + len > eb->start + eb->len);
4524
4525         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4526
4527         while (len > 0) {
4528                 page = extent_buffer_page(eb, i);
4529
4530                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4531                 kaddr = page_address(page);
4532                 memcpy(dst, kaddr + offset, cur);
4533
4534                 dst += cur;
4535                 len -= cur;
4536                 offset = 0;
4537                 i++;
4538         }
4539 }
4540
4541 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4542                                unsigned long min_len, char **map,
4543                                unsigned long *map_start,
4544                                unsigned long *map_len)
4545 {
4546         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4547         char *kaddr;
4548         struct page *p;
4549         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4550         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4551         unsigned long end_i = (start_offset + start + min_len - 1) >>
4552                 PAGE_CACHE_SHIFT;
4553
4554         if (i != end_i)
4555                 return -EINVAL;
4556
4557         if (i == 0) {
4558                 offset = start_offset;
4559                 *map_start = 0;
4560         } else {
4561                 offset = 0;
4562                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4563         }
4564
4565         if (start + min_len > eb->len) {
4566                 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4567                        "wanted %lu %lu\n", (unsigned long long)eb->start,
4568                        eb->len, start, min_len);
4569                 WARN_ON(1);
4570                 return -EINVAL;
4571         }
4572
4573         p = extent_buffer_page(eb, i);
4574         kaddr = page_address(p);
4575         *map = kaddr + offset;
4576         *map_len = PAGE_CACHE_SIZE - offset;
4577         return 0;
4578 }
4579
4580 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4581                           unsigned long start,
4582                           unsigned long len)
4583 {
4584         size_t cur;
4585         size_t offset;
4586         struct page *page;
4587         char *kaddr;
4588         char *ptr = (char *)ptrv;
4589         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4590         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4591         int ret = 0;
4592
4593         WARN_ON(start > eb->len);
4594         WARN_ON(start + len > eb->start + eb->len);
4595
4596         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4597
4598         while (len > 0) {
4599                 page = extent_buffer_page(eb, i);
4600
4601                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4602
4603                 kaddr = page_address(page);
4604                 ret = memcmp(ptr, kaddr + offset, cur);
4605                 if (ret)
4606                         break;
4607
4608                 ptr += cur;
4609                 len -= cur;
4610                 offset = 0;
4611                 i++;
4612         }
4613         return ret;
4614 }
4615
4616 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4617                          unsigned long start, unsigned long len)
4618 {
4619         size_t cur;
4620         size_t offset;
4621         struct page *page;
4622         char *kaddr;
4623         char *src = (char *)srcv;
4624         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4625         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4626
4627         WARN_ON(start > eb->len);
4628         WARN_ON(start + len > eb->start + eb->len);
4629
4630         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4631
4632         while (len > 0) {
4633                 page = extent_buffer_page(eb, i);
4634                 WARN_ON(!PageUptodate(page));
4635
4636                 cur = min(len, PAGE_CACHE_SIZE - offset);
4637                 kaddr = page_address(page);
4638                 memcpy(kaddr + offset, src, cur);
4639
4640                 src += cur;
4641                 len -= cur;
4642                 offset = 0;
4643                 i++;
4644         }
4645 }
4646
4647 void memset_extent_buffer(struct extent_buffer *eb, char c,
4648                           unsigned long start, unsigned long len)
4649 {
4650         size_t cur;
4651         size_t offset;
4652         struct page *page;
4653         char *kaddr;
4654         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4655         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4656
4657         WARN_ON(start > eb->len);
4658         WARN_ON(start + len > eb->start + eb->len);
4659
4660         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4661
4662         while (len > 0) {
4663                 page = extent_buffer_page(eb, i);
4664                 WARN_ON(!PageUptodate(page));
4665
4666                 cur = min(len, PAGE_CACHE_SIZE - offset);
4667                 kaddr = page_address(page);
4668                 memset(kaddr + offset, c, cur);
4669
4670                 len -= cur;
4671                 offset = 0;
4672                 i++;
4673         }
4674 }
4675
4676 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
4677                         unsigned long dst_offset, unsigned long src_offset,
4678                         unsigned long len)
4679 {
4680         u64 dst_len = dst->len;
4681         size_t cur;
4682         size_t offset;
4683         struct page *page;
4684         char *kaddr;
4685         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4686         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4687
4688         WARN_ON(src->len != dst_len);
4689
4690         offset = (start_offset + dst_offset) &
4691                 ((unsigned long)PAGE_CACHE_SIZE - 1);
4692
4693         while (len > 0) {
4694                 page = extent_buffer_page(dst, i);
4695                 WARN_ON(!PageUptodate(page));
4696
4697                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
4698
4699                 kaddr = page_address(page);
4700                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4701
4702                 src_offset += cur;
4703                 len -= cur;
4704                 offset = 0;
4705                 i++;
4706         }
4707 }
4708
4709 static void move_pages(struct page *dst_page, struct page *src_page,
4710                        unsigned long dst_off, unsigned long src_off,
4711                        unsigned long len)
4712 {
4713         char *dst_kaddr = page_address(dst_page);
4714         if (dst_page == src_page) {
4715                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
4716         } else {
4717                 char *src_kaddr = page_address(src_page);
4718                 char *p = dst_kaddr + dst_off + len;
4719                 char *s = src_kaddr + src_off + len;
4720
4721                 while (len--)
4722                         *--p = *--s;
4723         }
4724 }
4725
4726 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4727 {
4728         unsigned long distance = (src > dst) ? src - dst : dst - src;
4729         return distance < len;
4730 }
4731
4732 static void copy_pages(struct page *dst_page, struct page *src_page,
4733                        unsigned long dst_off, unsigned long src_off,
4734                        unsigned long len)
4735 {
4736         char *dst_kaddr = page_address(dst_page);
4737         char *src_kaddr;
4738         int must_memmove = 0;
4739
4740         if (dst_page != src_page) {
4741                 src_kaddr = page_address(src_page);
4742         } else {
4743                 src_kaddr = dst_kaddr;
4744                 if (areas_overlap(src_off, dst_off, len))
4745                         must_memmove = 1;
4746         }
4747
4748         if (must_memmove)
4749                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
4750         else
4751                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
4752 }
4753
4754 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4755                            unsigned long src_offset, unsigned long len)
4756 {
4757         size_t cur;
4758         size_t dst_off_in_page;
4759         size_t src_off_in_page;
4760         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4761         unsigned long dst_i;
4762         unsigned long src_i;
4763
4764         if (src_offset + len > dst->len) {
4765                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4766                        "len %lu dst len %lu\n", src_offset, len, dst->len);
4767                 BUG_ON(1);
4768         }
4769         if (dst_offset + len > dst->len) {
4770                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4771                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
4772                 BUG_ON(1);
4773         }
4774
4775         while (len > 0) {
4776                 dst_off_in_page = (start_offset + dst_offset) &
4777                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4778                 src_off_in_page = (start_offset + src_offset) &
4779                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4780
4781                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4782                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
4783
4784                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
4785                                                src_off_in_page));
4786                 cur = min_t(unsigned long, cur,
4787                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
4788
4789                 copy_pages(extent_buffer_page(dst, dst_i),
4790                            extent_buffer_page(dst, src_i),
4791                            dst_off_in_page, src_off_in_page, cur);
4792
4793                 src_offset += cur;
4794                 dst_offset += cur;
4795                 len -= cur;
4796         }
4797 }
4798
4799 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4800                            unsigned long src_offset, unsigned long len)
4801 {
4802         size_t cur;
4803         size_t dst_off_in_page;
4804         size_t src_off_in_page;
4805         unsigned long dst_end = dst_offset + len - 1;
4806         unsigned long src_end = src_offset + len - 1;
4807         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4808         unsigned long dst_i;
4809         unsigned long src_i;
4810
4811         if (src_offset + len > dst->len) {
4812                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4813                        "len %lu len %lu\n", src_offset, len, dst->len);
4814                 BUG_ON(1);
4815         }
4816         if (dst_offset + len > dst->len) {
4817                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4818                        "len %lu len %lu\n", dst_offset, len, dst->len);
4819                 BUG_ON(1);
4820         }
4821         if (dst_offset < src_offset) {
4822                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4823                 return;
4824         }
4825         while (len > 0) {
4826                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4827                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4828
4829                 dst_off_in_page = (start_offset + dst_end) &
4830                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4831                 src_off_in_page = (start_offset + src_end) &
4832                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4833
4834                 cur = min_t(unsigned long, len, src_off_in_page + 1);
4835                 cur = min(cur, dst_off_in_page + 1);
4836                 move_pages(extent_buffer_page(dst, dst_i),
4837                            extent_buffer_page(dst, src_i),
4838                            dst_off_in_page - cur + 1,
4839                            src_off_in_page - cur + 1, cur);
4840
4841                 dst_end -= cur;
4842                 src_end -= cur;
4843                 len -= cur;
4844         }
4845 }
4846
4847 int try_release_extent_buffer(struct page *page, gfp_t mask)
4848 {
4849         struct extent_buffer *eb;
4850
4851         /*
4852          * We need to make sure noboody is attaching this page to an eb right
4853          * now.
4854          */
4855         spin_lock(&page->mapping->private_lock);
4856         if (!PagePrivate(page)) {
4857                 spin_unlock(&page->mapping->private_lock);
4858                 return 1;
4859         }
4860
4861         eb = (struct extent_buffer *)page->private;
4862         BUG_ON(!eb);
4863
4864         /*
4865          * This is a little awful but should be ok, we need to make sure that
4866          * the eb doesn't disappear out from under us while we're looking at
4867          * this page.
4868          */
4869         spin_lock(&eb->refs_lock);
4870         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4871                 spin_unlock(&eb->refs_lock);
4872                 spin_unlock(&page->mapping->private_lock);
4873                 return 0;
4874         }
4875         spin_unlock(&page->mapping->private_lock);
4876
4877         if ((mask & GFP_NOFS) == GFP_NOFS)
4878                 mask = GFP_NOFS;
4879
4880         /*
4881          * If tree ref isn't set then we know the ref on this eb is a real ref,
4882          * so just return, this page will likely be freed soon anyway.
4883          */
4884         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4885                 spin_unlock(&eb->refs_lock);
4886                 return 0;
4887         }
4888         release_extent_buffer(eb, mask);
4889
4890         return 1;
4891 }