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/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26 static struct bio_set *btrfs_bioset;
  27
  28 static inline bool extent_state_in_tree(const struct extent_state *state)
  29 {
  30         return !RB_EMPTY_NODE(&state->rb_node);
  31 }
  32
  33 #ifdef CONFIG_BTRFS_DEBUG
  34 static LIST_HEAD(buffers);
  35 static LIST_HEAD(states);
  36
  37 static DEFINE_SPINLOCK(leak_lock);
  38
  39 static inline
  40 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  41 {
  42         unsigned long flags;
  43
  44         spin_lock_irqsave(&leak_lock, flags);
  45         list_add(new, head);
  46         spin_unlock_irqrestore(&leak_lock, flags);
  47 }
  48
  49 static inline
  50 void btrfs_leak_debug_del(struct list_head *entry)
  51 {
  52         unsigned long flags;
  53
  54         spin_lock_irqsave(&leak_lock, flags);
  55         list_del(entry);
  56         spin_unlock_irqrestore(&leak_lock, flags);
  57 }
  58
  59 static inline
  60 void btrfs_leak_debug_check(void)
  61 {
  62         struct extent_state *state;
  63         struct extent_buffer *eb;
  64
  65         while (!list_empty(&states)) {
  66                 state = list_entry(states.next, struct extent_state, leak_list);
  67                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  68                        state->start, state->end, state->state,
  69                        extent_state_in_tree(state),
  70                        atomic_read(&state->refs));
  71                 list_del(&state->leak_list);
  72                 kmem_cache_free(extent_state_cache, state);
  73         }
  74
  75         while (!list_empty(&buffers)) {
  76                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  77                 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu "
  78                        "refs %d\n",
  79                        eb->start, eb->len, atomic_read(&eb->refs));
  80                 list_del(&eb->leak_list);
  81                 kmem_cache_free(extent_buffer_cache, eb);
  82         }
  83 }
  84
  85 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  86         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  87 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  88                 struct extent_io_tree *tree, u64 start, u64 end)
  89 {
  90         struct inode *inode;
  91         u64 isize;
  92
  93         if (!tree->mapping)
  94                 return;
  95
  96         inode = tree->mapping->host;
  97         isize = i_size_read(inode);
  98         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  99                 printk_ratelimited(KERN_DEBUG
 100                     "BTRFS: %s: ino %llu isize %llu odd range [%llu,%llu]\n",
 101                                 caller, btrfs_ino(inode), isize, start, end);
 102         }
 103 }
 104 #else
 105 #define btrfs_leak_debug_add(new, head) do {} while (0)
 106 #define btrfs_leak_debug_del(entry)     do {} while (0)
 107 #define btrfs_leak_debug_check()        do {} while (0)
 108 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 109 #endif
 110
 111 #define BUFFER_LRU_MAX 64
 112
 113 struct tree_entry {
 114         u64 start;
 115         u64 end;
 116         struct rb_node rb_node;
 117 };
 118
 119 struct extent_page_data {
 120         struct bio *bio;
 121         struct extent_io_tree *tree;
 122         get_extent_t *get_extent;
 123         unsigned long bio_flags;
 124
 125         /* tells writepage not to lock the state bits for this range
 126          * it still does the unlocking
 127          */
 128         unsigned int extent_locked:1;
 129
 130         /* tells the submit_bio code to use a WRITE_SYNC */
 131         unsigned int sync_io:1;
 132 };
 133
 134 static noinline void flush_write_bio(void *data);
 135 static inline struct btrfs_fs_info *
 136 tree_fs_info(struct extent_io_tree *tree)
 137 {
 138         if (!tree->mapping)
 139                 return NULL;
 140         return btrfs_sb(tree->mapping->host->i_sb);
 141 }
 142
 143 int __init extent_io_init(void)
 144 {
 145         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 146                         sizeof(struct extent_state), 0,
 147                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 148         if (!extent_state_cache)
 149                 return -ENOMEM;
 150
 151         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 152                         sizeof(struct extent_buffer), 0,
 153                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 154         if (!extent_buffer_cache)
 155                 goto free_state_cache;
 156
 157         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 158                                      offsetof(struct btrfs_io_bio, bio));
 159         if (!btrfs_bioset)
 160                 goto free_buffer_cache;
 161
 162         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 163                 goto free_bioset;
 164
 165         return 0;
 166
 167 free_bioset:
 168         bioset_free(btrfs_bioset);
 169         btrfs_bioset = NULL;
 170
 171 free_buffer_cache:
 172         kmem_cache_destroy(extent_buffer_cache);
 173         extent_buffer_cache = NULL;
 174
 175 free_state_cache:
 176         kmem_cache_destroy(extent_state_cache);
 177         extent_state_cache = NULL;
 178         return -ENOMEM;
 179 }
 180
 181 void extent_io_exit(void)
 182 {
 183         btrfs_leak_debug_check();
 184
 185         /*
 186          * Make sure all delayed rcu free are flushed before we
 187          * destroy caches.
 188          */
 189         rcu_barrier();
 190         if (extent_state_cache)
 191                 kmem_cache_destroy(extent_state_cache);
 192         if (extent_buffer_cache)
 193                 kmem_cache_destroy(extent_buffer_cache);
 194         if (btrfs_bioset)
 195                 bioset_free(btrfs_bioset);
 196 }
 197
 198 void extent_io_tree_init(struct extent_io_tree *tree,
 199                          struct address_space *mapping)
 200 {
 201         tree->state = RB_ROOT;
 202         tree->ops = NULL;
 203         tree->dirty_bytes = 0;
 204         spin_lock_init(&tree->lock);
 205         tree->mapping = mapping;
 206 }
 207
 208 static struct extent_state *alloc_extent_state(gfp_t mask)
 209 {
 210         struct extent_state *state;
 211
 212         state = kmem_cache_alloc(extent_state_cache, mask);
 213         if (!state)
 214                 return state;
 215         state->state = 0;
 216         state->private = 0;
 217         RB_CLEAR_NODE(&state->rb_node);
 218         btrfs_leak_debug_add(&state->leak_list, &states);
 219         atomic_set(&state->refs, 1);
 220         init_waitqueue_head(&state->wq);
 221         trace_alloc_extent_state(state, mask, _RET_IP_);
 222         return state;
 223 }
 224
 225 void free_extent_state(struct extent_state *state)
 226 {
 227         if (!state)
 228                 return;
 229         if (atomic_dec_and_test(&state->refs)) {
 230                 WARN_ON(extent_state_in_tree(state));
 231                 btrfs_leak_debug_del(&state->leak_list);
 232                 trace_free_extent_state(state, _RET_IP_);
 233                 kmem_cache_free(extent_state_cache, state);
 234         }
 235 }
 236
 237 static struct rb_node *tree_insert(struct rb_root *root,
 238                                    struct rb_node *search_start,
 239                                    u64 offset,
 240                                    struct rb_node *node,
 241                                    struct rb_node ***p_in,
 242                                    struct rb_node **parent_in)
 243 {
 244         struct rb_node **p;
 245         struct rb_node *parent = NULL;
 246         struct tree_entry *entry;
 247
 248         if (p_in && parent_in) {
 249                 p = *p_in;
 250                 parent = *parent_in;
 251                 goto do_insert;
 252         }
 253
 254         p = search_start ? &search_start : &root->rb_node;
 255         while (*p) {
 256                 parent = *p;
 257                 entry = rb_entry(parent, struct tree_entry, rb_node);
 258
 259                 if (offset < entry->start)
 260                         p = &(*p)->rb_left;
 261                 else if (offset > entry->end)
 262                         p = &(*p)->rb_right;
 263                 else
 264                         return parent;
 265         }
 266
 267 do_insert:
 268         rb_link_node(node, parent, p);
 269         rb_insert_color(node, root);
 270         return NULL;
 271 }
 272
 273 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 274                                       struct rb_node **prev_ret,
 275                                       struct rb_node **next_ret,
 276                                       struct rb_node ***p_ret,
 277                                       struct rb_node **parent_ret)
 278 {
 279         struct rb_root *root = &tree->state;
 280         struct rb_node **n = &root->rb_node;
 281         struct rb_node *prev = NULL;
 282         struct rb_node *orig_prev = NULL;
 283         struct tree_entry *entry;
 284         struct tree_entry *prev_entry = NULL;
 285
 286         while (*n) {
 287                 prev = *n;
 288                 entry = rb_entry(prev, struct tree_entry, rb_node);
 289                 prev_entry = entry;
 290
 291                 if (offset < entry->start)
 292                         n = &(*n)->rb_left;
 293                 else if (offset > entry->end)
 294                         n = &(*n)->rb_right;
 295                 else
 296                         return *n;
 297         }
 298
 299         if (p_ret)
 300                 *p_ret = n;
 301         if (parent_ret)
 302                 *parent_ret = prev;
 303
 304         if (prev_ret) {
 305                 orig_prev = prev;
 306                 while (prev && offset > prev_entry->end) {
 307                         prev = rb_next(prev);
 308                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 309                 }
 310                 *prev_ret = prev;
 311                 prev = orig_prev;
 312         }
 313
 314         if (next_ret) {
 315                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 316                 while (prev && offset < prev_entry->start) {
 317                         prev = rb_prev(prev);
 318                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 319                 }
 320                 *next_ret = prev;
 321         }
 322         return NULL;
 323 }
 324
 325 static inline struct rb_node *
 326 tree_search_for_insert(struct extent_io_tree *tree,
 327                        u64 offset,
 328                        struct rb_node ***p_ret,
 329                        struct rb_node **parent_ret)
 330 {
 331         struct rb_node *prev = NULL;
 332         struct rb_node *ret;
 333
 334         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 335         if (!ret)
 336                 return prev;
 337         return ret;
 338 }
 339
 340 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 341                                           u64 offset)
 342 {
 343         return tree_search_for_insert(tree, offset, NULL, NULL);
 344 }
 345
 346 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 347                      struct extent_state *other)
 348 {
 349         if (tree->ops && tree->ops->merge_extent_hook)
 350                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 351                                              other);
 352 }
 353
 354 /*
 355  * utility function to look for merge candidates inside a given range.
 356  * Any extents with matching state are merged together into a single
 357  * extent in the tree.  Extents with EXTENT_IO in their state field
 358  * are not merged because the end_io handlers need to be able to do
 359  * operations on them without sleeping (or doing allocations/splits).
 360  *
 361  * This should be called with the tree lock held.
 362  */
 363 static void merge_state(struct extent_io_tree *tree,
 364                         struct extent_state *state)
 365 {
 366         struct extent_state *other;
 367         struct rb_node *other_node;
 368
 369         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 370                 return;
 371
 372         other_node = rb_prev(&state->rb_node);
 373         if (other_node) {
 374                 other = rb_entry(other_node, struct extent_state, rb_node);
 375                 if (other->end == state->start - 1 &&
 376                     other->state == state->state) {
 377                         merge_cb(tree, state, other);
 378                         state->start = other->start;
 379                         rb_erase(&other->rb_node, &tree->state);
 380                         RB_CLEAR_NODE(&other->rb_node);
 381                         free_extent_state(other);
 382                 }
 383         }
 384         other_node = rb_next(&state->rb_node);
 385         if (other_node) {
 386                 other = rb_entry(other_node, struct extent_state, rb_node);
 387                 if (other->start == state->end + 1 &&
 388                     other->state == state->state) {
 389                         merge_cb(tree, state, other);
 390                         state->end = other->end;
 391                         rb_erase(&other->rb_node, &tree->state);
 392                         RB_CLEAR_NODE(&other->rb_node);
 393                         free_extent_state(other);
 394                 }
 395         }
 396 }
 397
 398 static void set_state_cb(struct extent_io_tree *tree,
 399                          struct extent_state *state, unsigned *bits)
 400 {
 401         if (tree->ops && tree->ops->set_bit_hook)
 402                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 403 }
 404
 405 static void clear_state_cb(struct extent_io_tree *tree,
 406                            struct extent_state *state, unsigned *bits)
 407 {
 408         if (tree->ops && tree->ops->clear_bit_hook)
 409                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 410 }
 411
 412 static void set_state_bits(struct extent_io_tree *tree,
 413                            struct extent_state *state, unsigned *bits);
 414
 415 /*
 416  * insert an extent_state struct into the tree.  'bits' are set on the
 417  * struct before it is inserted.
 418  *
 419  * This may return -EEXIST if the extent is already there, in which case the
 420  * state struct is freed.
 421  *
 422  * The tree lock is not taken internally.  This is a utility function and
 423  * probably isn't what you want to call (see set/clear_extent_bit).
 424  */
 425 static int insert_state(struct extent_io_tree *tree,
 426                         struct extent_state *state, u64 start, u64 end,
 427                         struct rb_node ***p,
 428                         struct rb_node **parent,
 429                         unsigned *bits)
 430 {
 431         struct rb_node *node;
 432
 433         if (end < start)
 434                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 435                        end, start);
 436         state->start = start;
 437         state->end = end;
 438
 439         set_state_bits(tree, state, bits);
 440
 441         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 442         if (node) {
 443                 struct extent_state *found;
 444                 found = rb_entry(node, struct extent_state, rb_node);
 445                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
 446                        "%llu %llu\n",
 447                        found->start, found->end, start, end);
 448                 return -EEXIST;
 449         }
 450         merge_state(tree, state);
 451         return 0;
 452 }
 453
 454 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 455                      u64 split)
 456 {
 457         if (tree->ops && tree->ops->split_extent_hook)
 458                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 459 }
 460
 461 /*
 462  * split a given extent state struct in two, inserting the preallocated
 463  * struct 'prealloc' as the newly created second half.  'split' indicates an
 464  * offset inside 'orig' where it should be split.
 465  *
 466  * Before calling,
 467  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 468  * are two extent state structs in the tree:
 469  * prealloc: [orig->start, split - 1]
 470  * orig: [ split, orig->end ]
 471  *
 472  * The tree locks are not taken by this function. They need to be held
 473  * by the caller.
 474  */
 475 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 476                        struct extent_state *prealloc, u64 split)
 477 {
 478         struct rb_node *node;
 479
 480         split_cb(tree, orig, split);
 481
 482         prealloc->start = orig->start;
 483         prealloc->end = split - 1;
 484         prealloc->state = orig->state;
 485         orig->start = split;
 486
 487         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 488                            &prealloc->rb_node, NULL, NULL);
 489         if (node) {
 490                 free_extent_state(prealloc);
 491                 return -EEXIST;
 492         }
 493         return 0;
 494 }
 495
 496 static struct extent_state *next_state(struct extent_state *state)
 497 {
 498         struct rb_node *next = rb_next(&state->rb_node);
 499         if (next)
 500                 return rb_entry(next, struct extent_state, rb_node);
 501         else
 502                 return NULL;
 503 }
 504
 505 /*
 506  * utility function to clear some bits in an extent state struct.
 507  * it will optionally wake up any one waiting on this state (wake == 1).
 508  *
 509  * If no bits are set on the state struct after clearing things, the
 510  * struct is freed and removed from the tree
 511  */
 512 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 513                                             struct extent_state *state,
 514                                             unsigned *bits, int wake)
 515 {
 516         struct extent_state *next;
 517         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 518
 519         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 520                 u64 range = state->end - state->start + 1;
 521                 WARN_ON(range > tree->dirty_bytes);
 522                 tree->dirty_bytes -= range;
 523         }
 524         clear_state_cb(tree, state, bits);
 525         state->state &= ~bits_to_clear;
 526         if (wake)
 527                 wake_up(&state->wq);
 528         if (state->state == 0) {
 529                 next = next_state(state);
 530                 if (extent_state_in_tree(state)) {
 531                         rb_erase(&state->rb_node, &tree->state);
 532                         RB_CLEAR_NODE(&state->rb_node);
 533                         free_extent_state(state);
 534                 } else {
 535                         WARN_ON(1);
 536                 }
 537         } else {
 538                 merge_state(tree, state);
 539                 next = next_state(state);
 540         }
 541         return next;
 542 }
 543
 544 static struct extent_state *
 545 alloc_extent_state_atomic(struct extent_state *prealloc)
 546 {
 547         if (!prealloc)
 548                 prealloc = alloc_extent_state(GFP_ATOMIC);
 549
 550         return prealloc;
 551 }
 552
 553 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 554 {
 555         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 556                     "Extent tree was modified by another "
 557                     "thread while locked.");
 558 }
 559
 560 /*
 561  * clear some bits on a range in the tree.  This may require splitting
 562  * or inserting elements in the tree, so the gfp mask is used to
 563  * indicate which allocations or sleeping are allowed.
 564  *
 565  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 566  * the given range from the tree regardless of state (ie for truncate).
 567  *
 568  * the range [start, end] is inclusive.
 569  *
 570  * This takes the tree lock, and returns 0 on success and < 0 on error.
 571  */
 572 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 573                      unsigned bits, int wake, int delete,
 574                      struct extent_state **cached_state,
 575                      gfp_t mask)
 576 {
 577         struct extent_state *state;
 578         struct extent_state *cached;
 579         struct extent_state *prealloc = NULL;
 580         struct rb_node *node;
 581         u64 last_end;
 582         int err;
 583         int clear = 0;
 584
 585         btrfs_debug_check_extent_io_range(tree, start, end);
 586
 587         if (bits & EXTENT_DELALLOC)
 588                 bits |= EXTENT_NORESERVE;
 589
 590         if (delete)
 591                 bits |= ~EXTENT_CTLBITS;
 592         bits |= EXTENT_FIRST_DELALLOC;
 593
 594         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 595                 clear = 1;
 596 again:
 597         if (!prealloc && (mask & __GFP_WAIT)) {
 598                 /*
 599                  * Don't care for allocation failure here because we might end
 600                  * up not needing the pre-allocated extent state at all, which
 601                  * is the case if we only have in the tree extent states that
 602                  * cover our input range and don't cover too any other range.
 603                  * If we end up needing a new extent state we allocate it later.
 604                  */
 605                 prealloc = alloc_extent_state(mask);
 606         }
 607
 608         spin_lock(&tree->lock);
 609         if (cached_state) {
 610                 cached = *cached_state;
 611
 612                 if (clear) {
 613                         *cached_state = NULL;
 614                         cached_state = NULL;
 615                 }
 616
 617                 if (cached && extent_state_in_tree(cached) &&
 618                     cached->start <= start && cached->end > start) {
 619                         if (clear)
 620                                 atomic_dec(&cached->refs);
 621                         state = cached;
 622                         goto hit_next;
 623                 }
 624                 if (clear)
 625                         free_extent_state(cached);
 626         }
 627         /*
 628          * this search will find the extents that end after
 629          * our range starts
 630          */
 631         node = tree_search(tree, start);
 632         if (!node)
 633                 goto out;
 634         state = rb_entry(node, struct extent_state, rb_node);
 635 hit_next:
 636         if (state->start > end)
 637                 goto out;
 638         WARN_ON(state->end < start);
 639         last_end = state->end;
 640
 641         /* the state doesn't have the wanted bits, go ahead */
 642         if (!(state->state & bits)) {
 643                 state = next_state(state);
 644                 goto next;
 645         }
 646
 647         /*
 648          *     | ---- desired range ---- |
 649          *  | state | or
 650          *  | ------------- state -------------- |
 651          *
 652          * We need to split the extent we found, and may flip
 653          * bits on second half.
 654          *
 655          * If the extent we found extends past our range, we
 656          * just split and search again.  It'll get split again
 657          * the next time though.
 658          *
 659          * If the extent we found is inside our range, we clear
 660          * the desired bit on it.
 661          */
 662
 663         if (state->start < start) {
 664                 prealloc = alloc_extent_state_atomic(prealloc);
 665                 BUG_ON(!prealloc);
 666                 err = split_state(tree, state, prealloc, start);
 667                 if (err)
 668                         extent_io_tree_panic(tree, err);
 669
 670                 prealloc = NULL;
 671                 if (err)
 672                         goto out;
 673                 if (state->end <= end) {
 674                         state = clear_state_bit(tree, state, &bits, wake);
 675                         goto next;
 676                 }
 677                 goto search_again;
 678         }
 679         /*
 680          * | ---- desired range ---- |
 681          *                        | state |
 682          * We need to split the extent, and clear the bit
 683          * on the first half
 684          */
 685         if (state->start <= end && state->end > end) {
 686                 prealloc = alloc_extent_state_atomic(prealloc);
 687                 BUG_ON(!prealloc);
 688                 err = split_state(tree, state, prealloc, end + 1);
 689                 if (err)
 690                         extent_io_tree_panic(tree, err);
 691
 692                 if (wake)
 693                         wake_up(&state->wq);
 694
 695                 clear_state_bit(tree, prealloc, &bits, wake);
 696
 697                 prealloc = NULL;
 698                 goto out;
 699         }
 700
 701         state = clear_state_bit(tree, state, &bits, wake);
 702 next:
 703         if (last_end == (u64)-1)
 704                 goto out;
 705         start = last_end + 1;
 706         if (start <= end && state && !need_resched())
 707                 goto hit_next;
 708         goto search_again;
 709
 710 out:
 711         spin_unlock(&tree->lock);
 712         if (prealloc)
 713                 free_extent_state(prealloc);
 714
 715         return 0;
 716
 717 search_again:
 718         if (start > end)
 719                 goto out;
 720         spin_unlock(&tree->lock);
 721         if (mask & __GFP_WAIT)
 722                 cond_resched();
 723         goto again;
 724 }
 725
 726 static void wait_on_state(struct extent_io_tree *tree,
 727                           struct extent_state *state)
 728                 __releases(tree->lock)
 729                 __acquires(tree->lock)
 730 {
 731         DEFINE_WAIT(wait);
 732         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 733         spin_unlock(&tree->lock);
 734         schedule();
 735         spin_lock(&tree->lock);
 736         finish_wait(&state->wq, &wait);
 737 }
 738
 739 /*
 740  * waits for one or more bits to clear on a range in the state tree.
 741  * The range [start, end] is inclusive.
 742  * The tree lock is taken by this function
 743  */
 744 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 745                             unsigned long bits)
 746 {
 747         struct extent_state *state;
 748         struct rb_node *node;
 749
 750         btrfs_debug_check_extent_io_range(tree, start, end);
 751
 752         spin_lock(&tree->lock);
 753 again:
 754         while (1) {
 755                 /*
 756                  * this search will find all the extents that end after
 757                  * our range starts
 758                  */
 759                 node = tree_search(tree, start);
 760 process_node:
 761                 if (!node)
 762                         break;
 763
 764                 state = rb_entry(node, struct extent_state, rb_node);
 765
 766                 if (state->start > end)
 767                         goto out;
 768
 769                 if (state->state & bits) {
 770                         start = state->start;
 771                         atomic_inc(&state->refs);
 772                         wait_on_state(tree, state);
 773                         free_extent_state(state);
 774                         goto again;
 775                 }
 776                 start = state->end + 1;
 777
 778                 if (start > end)
 779                         break;
 780
 781                 if (!cond_resched_lock(&tree->lock)) {
 782                         node = rb_next(node);
 783                         goto process_node;
 784                 }
 785         }
 786 out:
 787         spin_unlock(&tree->lock);
 788 }
 789
 790 static void set_state_bits(struct extent_io_tree *tree,
 791                            struct extent_state *state,
 792                            unsigned *bits)
 793 {
 794         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 795
 796         set_state_cb(tree, state, bits);
 797         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 798                 u64 range = state->end - state->start + 1;
 799                 tree->dirty_bytes += range;
 800         }
 801         state->state |= bits_to_set;
 802 }
 803
 804 static void cache_state_if_flags(struct extent_state *state,
 805                                  struct extent_state **cached_ptr,
 806                                  unsigned flags)
 807 {
 808         if (cached_ptr && !(*cached_ptr)) {
 809                 if (!flags || (state->state & flags)) {
 810                         *cached_ptr = state;
 811                         atomic_inc(&state->refs);
 812                 }
 813         }
 814 }
 815
 816 static void cache_state(struct extent_state *state,
 817                         struct extent_state **cached_ptr)
 818 {
 819         return cache_state_if_flags(state, cached_ptr,
 820                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 821 }
 822
 823 /*
 824  * set some bits on a range in the tree.  This may require allocations or
 825  * sleeping, so the gfp mask is used to indicate what is allowed.
 826  *
 827  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 828  * part of the range already has the desired bits set.  The start of the
 829  * existing range is returned in failed_start in this case.
 830  *
 831  * [start, end] is inclusive This takes the tree lock.
 832  */
 833
 834 static int __must_check
 835 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 836                  unsigned bits, unsigned exclusive_bits,
 837                  u64 *failed_start, struct extent_state **cached_state,
 838                  gfp_t mask)
 839 {
 840         struct extent_state *state;
 841         struct extent_state *prealloc = NULL;
 842         struct rb_node *node;
 843         struct rb_node **p;
 844         struct rb_node *parent;
 845         int err = 0;
 846         u64 last_start;
 847         u64 last_end;
 848
 849         btrfs_debug_check_extent_io_range(tree, start, end);
 850
 851         bits |= EXTENT_FIRST_DELALLOC;
 852 again:
 853         if (!prealloc && (mask & __GFP_WAIT)) {
 854                 prealloc = alloc_extent_state(mask);
 855                 BUG_ON(!prealloc);
 856         }
 857
 858         spin_lock(&tree->lock);
 859         if (cached_state && *cached_state) {
 860                 state = *cached_state;
 861                 if (state->start <= start && state->end > start &&
 862                     extent_state_in_tree(state)) {
 863                         node = &state->rb_node;
 864                         goto hit_next;
 865                 }
 866         }
 867         /*
 868          * this search will find all the extents that end after
 869          * our range starts.
 870          */
 871         node = tree_search_for_insert(tree, start, &p, &parent);
 872         if (!node) {
 873                 prealloc = alloc_extent_state_atomic(prealloc);
 874                 BUG_ON(!prealloc);
 875                 err = insert_state(tree, prealloc, start, end,
 876                                    &p, &parent, &bits);
 877                 if (err)
 878                         extent_io_tree_panic(tree, err);
 879
 880                 cache_state(prealloc, cached_state);
 881                 prealloc = NULL;
 882                 goto out;
 883         }
 884         state = rb_entry(node, struct extent_state, rb_node);
 885 hit_next:
 886         last_start = state->start;
 887         last_end = state->end;
 888
 889         /*
 890          * | ---- desired range ---- |
 891          * | state |
 892          *
 893          * Just lock what we found and keep going
 894          */
 895         if (state->start == start && state->end <= end) {
 896                 if (state->state & exclusive_bits) {
 897                         *failed_start = state->start;
 898                         err = -EEXIST;
 899                         goto out;
 900                 }
 901
 902                 set_state_bits(tree, state, &bits);
 903                 cache_state(state, cached_state);
 904                 merge_state(tree, state);
 905                 if (last_end == (u64)-1)
 906                         goto out;
 907                 start = last_end + 1;
 908                 state = next_state(state);
 909                 if (start < end && state && state->start == start &&
 910                     !need_resched())
 911                         goto hit_next;
 912                 goto search_again;
 913         }
 914
 915         /*
 916          *     | ---- desired range ---- |
 917          * | state |
 918          *   or
 919          * | ------------- state -------------- |
 920          *
 921          * We need to split the extent we found, and may flip bits on
 922          * second half.
 923          *
 924          * If the extent we found extends past our
 925          * range, we just split and search again.  It'll get split
 926          * again the next time though.
 927          *
 928          * If the extent we found is inside our range, we set the
 929          * desired bit on it.
 930          */
 931         if (state->start < start) {
 932                 if (state->state & exclusive_bits) {
 933                         *failed_start = start;
 934                         err = -EEXIST;
 935                         goto out;
 936                 }
 937
 938                 prealloc = alloc_extent_state_atomic(prealloc);
 939                 BUG_ON(!prealloc);
 940                 err = split_state(tree, state, prealloc, start);
 941                 if (err)
 942                         extent_io_tree_panic(tree, err);
 943
 944                 prealloc = NULL;
 945                 if (err)
 946                         goto out;
 947                 if (state->end <= end) {
 948                         set_state_bits(tree, state, &bits);
 949                         cache_state(state, cached_state);
 950                         merge_state(tree, state);
 951                         if (last_end == (u64)-1)
 952                                 goto out;
 953                         start = last_end + 1;
 954                         state = next_state(state);
 955                         if (start < end && state && state->start == start &&
 956                             !need_resched())
 957                                 goto hit_next;
 958                 }
 959                 goto search_again;
 960         }
 961         /*
 962          * | ---- desired range ---- |
 963          *     | state | or               | state |
 964          *
 965          * There's a hole, we need to insert something in it and
 966          * ignore the extent we found.
 967          */
 968         if (state->start > start) {
 969                 u64 this_end;
 970                 if (end < last_start)
 971                         this_end = end;
 972                 else
 973                         this_end = last_start - 1;
 974
 975                 prealloc = alloc_extent_state_atomic(prealloc);
 976                 BUG_ON(!prealloc);
 977
 978                 /*
 979                  * Avoid to free 'prealloc' if it can be merged with
 980                  * the later extent.
 981                  */
 982                 err = insert_state(tree, prealloc, start, this_end,
 983                                    NULL, NULL, &bits);
 984                 if (err)
 985                         extent_io_tree_panic(tree, err);
 986
 987                 cache_state(prealloc, cached_state);
 988                 prealloc = NULL;
 989                 start = this_end + 1;
 990                 goto search_again;
 991         }
 992         /*
 993          * | ---- desired range ---- |
 994          *                        | state |
 995          * We need to split the extent, and set the bit
 996          * on the first half
 997          */
 998         if (state->start <= end && state->end > end) {
 999                 if (state->state & exclusive_bits) {
1000                         *failed_start = start;
1001                         err = -EEXIST;
1002                         goto out;
1003                 }
1004
1005                 prealloc = alloc_extent_state_atomic(prealloc);
1006                 BUG_ON(!prealloc);
1007                 err = split_state(tree, state, prealloc, end + 1);
1008                 if (err)
1009                         extent_io_tree_panic(tree, err);
1010
1011                 set_state_bits(tree, prealloc, &bits);
1012                 cache_state(prealloc, cached_state);
1013                 merge_state(tree, prealloc);
1014                 prealloc = NULL;
1015                 goto out;
1016         }
1017
1018         goto search_again;
1019
1020 out:
1021         spin_unlock(&tree->lock);
1022         if (prealloc)
1023                 free_extent_state(prealloc);
1024
1025         return err;
1026
1027 search_again:
1028         if (start > end)
1029                 goto out;
1030         spin_unlock(&tree->lock);
1031         if (mask & __GFP_WAIT)
1032                 cond_resched();
1033         goto again;
1034 }
1035
1036 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1037                    unsigned bits, u64 * failed_start,
1038                    struct extent_state **cached_state, gfp_t mask)
1039 {
1040         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1041                                 cached_state, mask);
1042 }
1043
1044
1045 /**
1046  * convert_extent_bit - convert all bits in a given range from one bit to
1047  *                      another
1048  * @tree:       the io tree to search
1049  * @start:      the start offset in bytes
1050  * @end:        the end offset in bytes (inclusive)
1051  * @bits:       the bits to set in this range
1052  * @clear_bits: the bits to clear in this range
1053  * @cached_state:       state that we're going to cache
1054  * @mask:       the allocation mask
1055  *
1056  * This will go through and set bits for the given range.  If any states exist
1057  * already in this range they are set with the given bit and cleared of the
1058  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1059  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1060  * boundary bits like LOCK.
1061  */
1062 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1063                        unsigned bits, unsigned clear_bits,
1064                        struct extent_state **cached_state, gfp_t mask)
1065 {
1066         struct extent_state *state;
1067         struct extent_state *prealloc = NULL;
1068         struct rb_node *node;
1069         struct rb_node **p;
1070         struct rb_node *parent;
1071         int err = 0;
1072         u64 last_start;
1073         u64 last_end;
1074         bool first_iteration = true;
1075
1076         btrfs_debug_check_extent_io_range(tree, start, end);
1077
1078 again:
1079         if (!prealloc && (mask & __GFP_WAIT)) {
1080                 /*
1081                  * Best effort, don't worry if extent state allocation fails
1082                  * here for the first iteration. We might have a cached state
1083                  * that matches exactly the target range, in which case no
1084                  * extent state allocations are needed. We'll only know this
1085                  * after locking the tree.
1086                  */
1087                 prealloc = alloc_extent_state(mask);
1088                 if (!prealloc && !first_iteration)
1089                         return -ENOMEM;
1090         }
1091
1092         spin_lock(&tree->lock);
1093         if (cached_state && *cached_state) {
1094                 state = *cached_state;
1095                 if (state->start <= start && state->end > start &&
1096                     extent_state_in_tree(state)) {
1097                         node = &state->rb_node;
1098                         goto hit_next;
1099                 }
1100         }
1101
1102         /*
1103          * this search will find all the extents that end after
1104          * our range starts.
1105          */
1106         node = tree_search_for_insert(tree, start, &p, &parent);
1107         if (!node) {
1108                 prealloc = alloc_extent_state_atomic(prealloc);
1109                 if (!prealloc) {
1110                         err = -ENOMEM;
1111                         goto out;
1112                 }
1113                 err = insert_state(tree, prealloc, start, end,
1114                                    &p, &parent, &bits);
1115                 if (err)
1116                         extent_io_tree_panic(tree, err);
1117                 cache_state(prealloc, cached_state);
1118                 prealloc = NULL;
1119                 goto out;
1120         }
1121         state = rb_entry(node, struct extent_state, rb_node);
1122 hit_next:
1123         last_start = state->start;
1124         last_end = state->end;
1125
1126         /*
1127          * | ---- desired range ---- |
1128          * | state |
1129          *
1130          * Just lock what we found and keep going
1131          */
1132         if (state->start == start && state->end <= end) {
1133                 set_state_bits(tree, state, &bits);
1134                 cache_state(state, cached_state);
1135                 state = clear_state_bit(tree, state, &clear_bits, 0);
1136                 if (last_end == (u64)-1)
1137                         goto out;
1138                 start = last_end + 1;
1139                 if (start < end && state && state->start == start &&
1140                     !need_resched())
1141                         goto hit_next;
1142                 goto search_again;
1143         }
1144
1145         /*
1146          *     | ---- desired range ---- |
1147          * | state |
1148          *   or
1149          * | ------------- state -------------- |
1150          *
1151          * We need to split the extent we found, and may flip bits on
1152          * second half.
1153          *
1154          * If the extent we found extends past our
1155          * range, we just split and search again.  It'll get split
1156          * again the next time though.
1157          *
1158          * If the extent we found is inside our range, we set the
1159          * desired bit on it.
1160          */
1161         if (state->start < start) {
1162                 prealloc = alloc_extent_state_atomic(prealloc);
1163                 if (!prealloc) {
1164                         err = -ENOMEM;
1165                         goto out;
1166                 }
1167                 err = split_state(tree, state, prealloc, start);
1168                 if (err)
1169                         extent_io_tree_panic(tree, err);
1170                 prealloc = NULL;
1171                 if (err)
1172                         goto out;
1173                 if (state->end <= end) {
1174                         set_state_bits(tree, state, &bits);
1175                         cache_state(state, cached_state);
1176                         state = clear_state_bit(tree, state, &clear_bits, 0);
1177                         if (last_end == (u64)-1)
1178                                 goto out;
1179                         start = last_end + 1;
1180                         if (start < end && state && state->start == start &&
1181                             !need_resched())
1182                                 goto hit_next;
1183                 }
1184                 goto search_again;
1185         }
1186         /*
1187          * | ---- desired range ---- |
1188          *     | state | or               | state |
1189          *
1190          * There's a hole, we need to insert something in it and
1191          * ignore the extent we found.
1192          */
1193         if (state->start > start) {
1194                 u64 this_end;
1195                 if (end < last_start)
1196                         this_end = end;
1197                 else
1198                         this_end = last_start - 1;
1199
1200                 prealloc = alloc_extent_state_atomic(prealloc);
1201                 if (!prealloc) {
1202                         err = -ENOMEM;
1203                         goto out;
1204                 }
1205
1206                 /*
1207                  * Avoid to free 'prealloc' if it can be merged with
1208                  * the later extent.
1209                  */
1210                 err = insert_state(tree, prealloc, start, this_end,
1211                                    NULL, NULL, &bits);
1212                 if (err)
1213                         extent_io_tree_panic(tree, err);
1214                 cache_state(prealloc, cached_state);
1215                 prealloc = NULL;
1216                 start = this_end + 1;
1217                 goto search_again;
1218         }
1219         /*
1220          * | ---- desired range ---- |
1221          *                        | state |
1222          * We need to split the extent, and set the bit
1223          * on the first half
1224          */
1225         if (state->start <= end && state->end > end) {
1226                 prealloc = alloc_extent_state_atomic(prealloc);
1227                 if (!prealloc) {
1228                         err = -ENOMEM;
1229                         goto out;
1230                 }
1231
1232                 err = split_state(tree, state, prealloc, end + 1);
1233                 if (err)
1234                         extent_io_tree_panic(tree, err);
1235
1236                 set_state_bits(tree, prealloc, &bits);
1237                 cache_state(prealloc, cached_state);
1238                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1239                 prealloc = NULL;
1240                 goto out;
1241         }
1242
1243         goto search_again;
1244
1245 out:
1246         spin_unlock(&tree->lock);
1247         if (prealloc)
1248                 free_extent_state(prealloc);
1249
1250         return err;
1251
1252 search_again:
1253         if (start > end)
1254                 goto out;
1255         spin_unlock(&tree->lock);
1256         if (mask & __GFP_WAIT)
1257                 cond_resched();
1258         first_iteration = false;
1259         goto again;
1260 }
1261
1262 /* wrappers around set/clear extent bit */
1263 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1264                      gfp_t mask)
1265 {
1266         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1267                               NULL, mask);
1268 }
1269
1270 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1271                     unsigned bits, gfp_t mask)
1272 {
1273         return set_extent_bit(tree, start, end, bits, NULL,
1274                               NULL, mask);
1275 }
1276
1277 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1278                       unsigned bits, gfp_t mask)
1279 {
1280         int wake = 0;
1281
1282         if (bits & EXTENT_LOCKED)
1283                 wake = 1;
1284
1285         return clear_extent_bit(tree, start, end, bits, wake, 0, NULL, mask);
1286 }
1287
1288 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1289                         struct extent_state **cached_state, gfp_t mask)
1290 {
1291         return set_extent_bit(tree, start, end,
1292                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1293                               NULL, cached_state, mask);
1294 }
1295
1296 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1297                       struct extent_state **cached_state, gfp_t mask)
1298 {
1299         return set_extent_bit(tree, start, end,
1300                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1301                               NULL, cached_state, mask);
1302 }
1303
1304 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1305                        gfp_t mask)
1306 {
1307         return clear_extent_bit(tree, start, end,
1308                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1309                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1310 }
1311
1312 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1313                      gfp_t mask)
1314 {
1315         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1316                               NULL, mask);
1317 }
1318
1319 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1320                         struct extent_state **cached_state, gfp_t mask)
1321 {
1322         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1323                               cached_state, mask);
1324 }
1325
1326 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1327                           struct extent_state **cached_state, gfp_t mask)
1328 {
1329         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1330                                 cached_state, mask);
1331 }
1332
1333 /*
1334  * either insert or lock state struct between start and end use mask to tell
1335  * us if waiting is desired.
1336  */
1337 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1338                      unsigned bits, struct extent_state **cached_state)
1339 {
1340         int err;
1341         u64 failed_start;
1342
1343         while (1) {
1344                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1345                                        EXTENT_LOCKED, &failed_start,
1346                                        cached_state, GFP_NOFS);
1347                 if (err == -EEXIST) {
1348                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1349                         start = failed_start;
1350                 } else
1351                         break;
1352                 WARN_ON(start > end);
1353         }
1354         return err;
1355 }
1356
1357 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1358 {
1359         return lock_extent_bits(tree, start, end, 0, NULL);
1360 }
1361
1362 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1363 {
1364         int err;
1365         u64 failed_start;
1366
1367         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1368                                &failed_start, NULL, GFP_NOFS);
1369         if (err == -EEXIST) {
1370                 if (failed_start > start)
1371                         clear_extent_bit(tree, start, failed_start - 1,
1372                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1373                 return 0;
1374         }
1375         return 1;
1376 }
1377
1378 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1379                          struct extent_state **cached, gfp_t mask)
1380 {
1381         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1382                                 mask);
1383 }
1384
1385 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1386 {
1387         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1388                                 GFP_NOFS);
1389 }
1390
1391 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1392 {
1393         unsigned long index = start >> PAGE_CACHE_SHIFT;
1394         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1395         struct page *page;
1396
1397         while (index <= end_index) {
1398                 page = find_get_page(inode->i_mapping, index);
1399                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1400                 clear_page_dirty_for_io(page);
1401                 page_cache_release(page);
1402                 index++;
1403         }
1404         return 0;
1405 }
1406
1407 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1408 {
1409         unsigned long index = start >> PAGE_CACHE_SHIFT;
1410         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1411         struct page *page;
1412
1413         while (index <= end_index) {
1414                 page = find_get_page(inode->i_mapping, index);
1415                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1416                 __set_page_dirty_nobuffers(page);
1417                 account_page_redirty(page);
1418                 page_cache_release(page);
1419                 index++;
1420         }
1421         return 0;
1422 }
1423
1424 /*
1425  * helper function to set both pages and extents in the tree writeback
1426  */
1427 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1428 {
1429         unsigned long index = start >> PAGE_CACHE_SHIFT;
1430         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1431         struct page *page;
1432
1433         while (index <= end_index) {
1434                 page = find_get_page(tree->mapping, index);
1435                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1436                 set_page_writeback(page);
1437                 page_cache_release(page);
1438                 index++;
1439         }
1440         return 0;
1441 }
1442
1443 /* find the first state struct with 'bits' set after 'start', and
1444  * return it.  tree->lock must be held.  NULL will returned if
1445  * nothing was found after 'start'
1446  */
1447 static struct extent_state *
1448 find_first_extent_bit_state(struct extent_io_tree *tree,
1449                             u64 start, unsigned bits)
1450 {
1451         struct rb_node *node;
1452         struct extent_state *state;
1453
1454         /*
1455          * this search will find all the extents that end after
1456          * our range starts.
1457          */
1458         node = tree_search(tree, start);
1459         if (!node)
1460                 goto out;
1461
1462         while (1) {
1463                 state = rb_entry(node, struct extent_state, rb_node);
1464                 if (state->end >= start && (state->state & bits))
1465                         return state;
1466
1467                 node = rb_next(node);
1468                 if (!node)
1469                         break;
1470         }
1471 out:
1472         return NULL;
1473 }
1474
1475 /*
1476  * find the first offset in the io tree with 'bits' set. zero is
1477  * returned if we find something, and *start_ret and *end_ret are
1478  * set to reflect the state struct that was found.
1479  *
1480  * If nothing was found, 1 is returned. If found something, return 0.
1481  */
1482 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1483                           u64 *start_ret, u64 *end_ret, unsigned bits,
1484                           struct extent_state **cached_state)
1485 {
1486         struct extent_state *state;
1487         struct rb_node *n;
1488         int ret = 1;
1489
1490         spin_lock(&tree->lock);
1491         if (cached_state && *cached_state) {
1492                 state = *cached_state;
1493                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1494                         n = rb_next(&state->rb_node);
1495                         while (n) {
1496                                 state = rb_entry(n, struct extent_state,
1497                                                  rb_node);
1498                                 if (state->state & bits)
1499                                         goto got_it;
1500                                 n = rb_next(n);
1501                         }
1502                         free_extent_state(*cached_state);
1503                         *cached_state = NULL;
1504                         goto out;
1505                 }
1506                 free_extent_state(*cached_state);
1507                 *cached_state = NULL;
1508         }
1509
1510         state = find_first_extent_bit_state(tree, start, bits);
1511 got_it:
1512         if (state) {
1513                 cache_state_if_flags(state, cached_state, 0);
1514                 *start_ret = state->start;
1515                 *end_ret = state->end;
1516                 ret = 0;
1517         }
1518 out:
1519         spin_unlock(&tree->lock);
1520         return ret;
1521 }
1522
1523 /*
1524  * find a contiguous range of bytes in the file marked as delalloc, not
1525  * more than 'max_bytes'.  start and end are used to return the range,
1526  *
1527  * 1 is returned if we find something, 0 if nothing was in the tree
1528  */
1529 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1530                                         u64 *start, u64 *end, u64 max_bytes,
1531                                         struct extent_state **cached_state)
1532 {
1533         struct rb_node *node;
1534         struct extent_state *state;
1535         u64 cur_start = *start;
1536         u64 found = 0;
1537         u64 total_bytes = 0;
1538
1539         spin_lock(&tree->lock);
1540
1541         /*
1542          * this search will find all the extents that end after
1543          * our range starts.
1544          */
1545         node = tree_search(tree, cur_start);
1546         if (!node) {
1547                 if (!found)
1548                         *end = (u64)-1;
1549                 goto out;
1550         }
1551
1552         while (1) {
1553                 state = rb_entry(node, struct extent_state, rb_node);
1554                 if (found && (state->start != cur_start ||
1555                               (state->state & EXTENT_BOUNDARY))) {
1556                         goto out;
1557                 }
1558                 if (!(state->state & EXTENT_DELALLOC)) {
1559                         if (!found)
1560                                 *end = state->end;
1561                         goto out;
1562                 }
1563                 if (!found) {
1564                         *start = state->start;
1565                         *cached_state = state;
1566                         atomic_inc(&state->refs);
1567                 }
1568                 found++;
1569                 *end = state->end;
1570                 cur_start = state->end + 1;
1571                 node = rb_next(node);
1572                 total_bytes += state->end - state->start + 1;
1573                 if (total_bytes >= max_bytes)
1574                         break;
1575                 if (!node)
1576                         break;
1577         }
1578 out:
1579         spin_unlock(&tree->lock);
1580         return found;
1581 }
1582
1583 static noinline void __unlock_for_delalloc(struct inode *inode,
1584                                            struct page *locked_page,
1585                                            u64 start, u64 end)
1586 {
1587         int ret;
1588         struct page *pages[16];
1589         unsigned long index = start >> PAGE_CACHE_SHIFT;
1590         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1591         unsigned long nr_pages = end_index - index + 1;
1592         int i;
1593
1594         if (index == locked_page->index && end_index == index)
1595                 return;
1596
1597         while (nr_pages > 0) {
1598                 ret = find_get_pages_contig(inode->i_mapping, index,
1599                                      min_t(unsigned long, nr_pages,
1600                                      ARRAY_SIZE(pages)), pages);
1601                 for (i = 0; i < ret; i++) {
1602                         if (pages[i] != locked_page)
1603                                 unlock_page(pages[i]);
1604                         page_cache_release(pages[i]);
1605                 }
1606                 nr_pages -= ret;
1607                 index += ret;
1608                 cond_resched();
1609         }
1610 }
1611
1612 static noinline int lock_delalloc_pages(struct inode *inode,
1613                                         struct page *locked_page,
1614                                         u64 delalloc_start,
1615                                         u64 delalloc_end)
1616 {
1617         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1618         unsigned long start_index = index;
1619         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1620         unsigned long pages_locked = 0;
1621         struct page *pages[16];
1622         unsigned long nrpages;
1623         int ret;
1624         int i;
1625
1626         /* the caller is responsible for locking the start index */
1627         if (index == locked_page->index && index == end_index)
1628                 return 0;
1629
1630         /* skip the page at the start index */
1631         nrpages = end_index - index + 1;
1632         while (nrpages > 0) {
1633                 ret = find_get_pages_contig(inode->i_mapping, index,
1634                                      min_t(unsigned long,
1635                                      nrpages, ARRAY_SIZE(pages)), pages);
1636                 if (ret == 0) {
1637                         ret = -EAGAIN;
1638                         goto done;
1639                 }
1640                 /* now we have an array of pages, lock them all */
1641                 for (i = 0; i < ret; i++) {
1642                         /*
1643                          * the caller is taking responsibility for
1644                          * locked_page
1645                          */
1646                         if (pages[i] != locked_page) {
1647                                 lock_page(pages[i]);
1648                                 if (!PageDirty(pages[i]) ||
1649                                     pages[i]->mapping != inode->i_mapping) {
1650                                         ret = -EAGAIN;
1651                                         unlock_page(pages[i]);
1652                                         page_cache_release(pages[i]);
1653                                         goto done;
1654                                 }
1655                         }
1656                         page_cache_release(pages[i]);
1657                         pages_locked++;
1658                 }
1659                 nrpages -= ret;
1660                 index += ret;
1661                 cond_resched();
1662         }
1663         ret = 0;
1664 done:
1665         if (ret && pages_locked) {
1666                 __unlock_for_delalloc(inode, locked_page,
1667                               delalloc_start,
1668                               ((u64)(start_index + pages_locked - 1)) <<
1669                               PAGE_CACHE_SHIFT);
1670         }
1671         return ret;
1672 }
1673
1674 /*
1675  * find a contiguous range of bytes in the file marked as delalloc, not
1676  * more than 'max_bytes'.  start and end are used to return the range,
1677  *
1678  * 1 is returned if we find something, 0 if nothing was in the tree
1679  */
1680 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1681                                     struct extent_io_tree *tree,
1682                                     struct page *locked_page, u64 *start,
1683                                     u64 *end, u64 max_bytes)
1684 {
1685         u64 delalloc_start;
1686         u64 delalloc_end;
1687         u64 found;
1688         struct extent_state *cached_state = NULL;
1689         int ret;
1690         int loops = 0;
1691
1692 again:
1693         /* step one, find a bunch of delalloc bytes starting at start */
1694         delalloc_start = *start;
1695         delalloc_end = 0;
1696         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1697                                     max_bytes, &cached_state);
1698         if (!found || delalloc_end <= *start) {
1699                 *start = delalloc_start;
1700                 *end = delalloc_end;
1701                 free_extent_state(cached_state);
1702                 return 0;
1703         }
1704
1705         /*
1706          * start comes from the offset of locked_page.  We have to lock
1707          * pages in order, so we can't process delalloc bytes before
1708          * locked_page
1709          */
1710         if (delalloc_start < *start)
1711                 delalloc_start = *start;
1712
1713         /*
1714          * make sure to limit the number of pages we try to lock down
1715          */
1716         if (delalloc_end + 1 - delalloc_start > max_bytes)
1717                 delalloc_end = delalloc_start + max_bytes - 1;
1718
1719         /* step two, lock all the pages after the page that has start */
1720         ret = lock_delalloc_pages(inode, locked_page,
1721                                   delalloc_start, delalloc_end);
1722         if (ret == -EAGAIN) {
1723                 /* some of the pages are gone, lets avoid looping by
1724                  * shortening the size of the delalloc range we're searching
1725                  */
1726                 free_extent_state(cached_state);
1727                 cached_state = NULL;
1728                 if (!loops) {
1729                         max_bytes = PAGE_CACHE_SIZE;
1730                         loops = 1;
1731                         goto again;
1732                 } else {
1733                         found = 0;
1734                         goto out_failed;
1735                 }
1736         }
1737         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1738
1739         /* step three, lock the state bits for the whole range */
1740         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1741
1742         /* then test to make sure it is all still delalloc */
1743         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1744                              EXTENT_DELALLOC, 1, cached_state);
1745         if (!ret) {
1746                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1747                                      &cached_state, GFP_NOFS);
1748                 __unlock_for_delalloc(inode, locked_page,
1749                               delalloc_start, delalloc_end);
1750                 cond_resched();
1751                 goto again;
1752         }
1753         free_extent_state(cached_state);
1754         *start = delalloc_start;
1755         *end = delalloc_end;
1756 out_failed:
1757         return found;
1758 }
1759
1760 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1761                                  struct page *locked_page,
1762                                  unsigned clear_bits,
1763                                  unsigned long page_ops)
1764 {
1765         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1766         int ret;
1767         struct page *pages[16];
1768         unsigned long index = start >> PAGE_CACHE_SHIFT;
1769         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1770         unsigned long nr_pages = end_index - index + 1;
1771         int i;
1772
1773         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1774         if (page_ops == 0)
1775                 return 0;
1776
1777         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1778                 mapping_set_error(inode->i_mapping, -EIO);
1779
1780         while (nr_pages > 0) {
1781                 ret = find_get_pages_contig(inode->i_mapping, index,
1782                                      min_t(unsigned long,
1783                                      nr_pages, ARRAY_SIZE(pages)), pages);
1784                 for (i = 0; i < ret; i++) {
1785
1786                         if (page_ops & PAGE_SET_PRIVATE2)
1787                                 SetPagePrivate2(pages[i]);
1788
1789                         if (pages[i] == locked_page) {
1790                                 page_cache_release(pages[i]);
1791                                 continue;
1792                         }
1793                         if (page_ops & PAGE_CLEAR_DIRTY)
1794                                 clear_page_dirty_for_io(pages[i]);
1795                         if (page_ops & PAGE_SET_WRITEBACK)
1796                                 set_page_writeback(pages[i]);
1797                         if (page_ops & PAGE_SET_ERROR)
1798                                 SetPageError(pages[i]);
1799                         if (page_ops & PAGE_END_WRITEBACK)
1800                                 end_page_writeback(pages[i]);
1801                         if (page_ops & PAGE_UNLOCK)
1802                                 unlock_page(pages[i]);
1803                         page_cache_release(pages[i]);
1804                 }
1805                 nr_pages -= ret;
1806                 index += ret;
1807                 cond_resched();
1808         }
1809         return 0;
1810 }
1811
1812 /*
1813  * count the number of bytes in the tree that have a given bit(s)
1814  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1815  * cached.  The total number found is returned.
1816  */
1817 u64 count_range_bits(struct extent_io_tree *tree,
1818                      u64 *start, u64 search_end, u64 max_bytes,
1819                      unsigned bits, int contig)
1820 {
1821         struct rb_node *node;
1822         struct extent_state *state;
1823         u64 cur_start = *start;
1824         u64 total_bytes = 0;
1825         u64 last = 0;
1826         int found = 0;
1827
1828         if (WARN_ON(search_end <= cur_start))
1829                 return 0;
1830
1831         spin_lock(&tree->lock);
1832         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1833                 total_bytes = tree->dirty_bytes;
1834                 goto out;
1835         }
1836         /*
1837          * this search will find all the extents that end after
1838          * our range starts.
1839          */
1840         node = tree_search(tree, cur_start);
1841         if (!node)
1842                 goto out;
1843
1844         while (1) {
1845                 state = rb_entry(node, struct extent_state, rb_node);
1846                 if (state->start > search_end)
1847                         break;
1848                 if (contig && found && state->start > last + 1)
1849                         break;
1850                 if (state->end >= cur_start && (state->state & bits) == bits) {
1851                         total_bytes += min(search_end, state->end) + 1 -
1852                                        max(cur_start, state->start);
1853                         if (total_bytes >= max_bytes)
1854                                 break;
1855                         if (!found) {
1856                                 *start = max(cur_start, state->start);
1857                                 found = 1;
1858                         }
1859                         last = state->end;
1860                 } else if (contig && found) {
1861                         break;
1862                 }
1863                 node = rb_next(node);
1864                 if (!node)
1865                         break;
1866         }
1867 out:
1868         spin_unlock(&tree->lock);
1869         return total_bytes;
1870 }
1871
1872 /*
1873  * set the private field for a given byte offset in the tree.  If there isn't
1874  * an extent_state there already, this does nothing.
1875  */
1876 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1877 {
1878         struct rb_node *node;
1879         struct extent_state *state;
1880         int ret = 0;
1881
1882         spin_lock(&tree->lock);
1883         /*
1884          * this search will find all the extents that end after
1885          * our range starts.
1886          */
1887         node = tree_search(tree, start);
1888         if (!node) {
1889                 ret = -ENOENT;
1890                 goto out;
1891         }
1892         state = rb_entry(node, struct extent_state, rb_node);
1893         if (state->start != start) {
1894                 ret = -ENOENT;
1895                 goto out;
1896         }
1897         state->private = private;
1898 out:
1899         spin_unlock(&tree->lock);
1900         return ret;
1901 }
1902
1903 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1904 {
1905         struct rb_node *node;
1906         struct extent_state *state;
1907         int ret = 0;
1908
1909         spin_lock(&tree->lock);
1910         /*
1911          * this search will find all the extents that end after
1912          * our range starts.
1913          */
1914         node = tree_search(tree, start);
1915         if (!node) {
1916                 ret = -ENOENT;
1917                 goto out;
1918         }
1919         state = rb_entry(node, struct extent_state, rb_node);
1920         if (state->start != start) {
1921                 ret = -ENOENT;
1922                 goto out;
1923         }
1924         *private = state->private;
1925 out:
1926         spin_unlock(&tree->lock);
1927         return ret;
1928 }
1929
1930 /*
1931  * searches a range in the state tree for a given mask.
1932  * If 'filled' == 1, this returns 1 only if every extent in the tree
1933  * has the bits set.  Otherwise, 1 is returned if any bit in the
1934  * range is found set.
1935  */
1936 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1937                    unsigned bits, int filled, struct extent_state *cached)
1938 {
1939         struct extent_state *state = NULL;
1940         struct rb_node *node;
1941         int bitset = 0;
1942
1943         spin_lock(&tree->lock);
1944         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1945             cached->end > start)
1946                 node = &cached->rb_node;
1947         else
1948                 node = tree_search(tree, start);
1949         while (node && start <= end) {
1950                 state = rb_entry(node, struct extent_state, rb_node);
1951
1952                 if (filled && state->start > start) {
1953                         bitset = 0;
1954                         break;
1955                 }
1956
1957                 if (state->start > end)
1958                         break;
1959
1960                 if (state->state & bits) {
1961                         bitset = 1;
1962                         if (!filled)
1963                                 break;
1964                 } else if (filled) {
1965                         bitset = 0;
1966                         break;
1967                 }
1968
1969                 if (state->end == (u64)-1)
1970                         break;
1971
1972                 start = state->end + 1;
1973                 if (start > end)
1974                         break;
1975                 node = rb_next(node);
1976                 if (!node) {
1977                         if (filled)
1978                                 bitset = 0;
1979                         break;
1980                 }
1981         }
1982         spin_unlock(&tree->lock);
1983         return bitset;
1984 }
1985
1986 /*
1987  * helper function to set a given page up to date if all the
1988  * extents in the tree for that page are up to date
1989  */
1990 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1991 {
1992         u64 start = page_offset(page);
1993         u64 end = start + PAGE_CACHE_SIZE - 1;
1994         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1995                 SetPageUptodate(page);
1996 }
1997
1998 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1999 {
2000         int ret;
2001         int err = 0;
2002         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2003
2004         set_state_private(failure_tree, rec->start, 0);
2005         ret = clear_extent_bits(failure_tree, rec->start,
2006                                 rec->start + rec->len - 1,
2007                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2008         if (ret)
2009                 err = ret;
2010
2011         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
2012                                 rec->start + rec->len - 1,
2013                                 EXTENT_DAMAGED, GFP_NOFS);
2014         if (ret && !err)
2015                 err = ret;
2016
2017         kfree(rec);
2018         return err;
2019 }
2020
2021 /*
2022  * this bypasses the standard btrfs submit functions deliberately, as
2023  * the standard behavior is to write all copies in a raid setup. here we only
2024  * want to write the one bad copy. so we do the mapping for ourselves and issue
2025  * submit_bio directly.
2026  * to avoid any synchronization issues, wait for the data after writing, which
2027  * actually prevents the read that triggered the error from finishing.
2028  * currently, there can be no more than two copies of every data bit. thus,
2029  * exactly one rewrite is required.
2030  */
2031 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2032                       struct page *page, unsigned int pg_offset, int mirror_num)
2033 {
2034         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2035         struct bio *bio;
2036         struct btrfs_device *dev;
2037         u64 map_length = 0;
2038         u64 sector;
2039         struct btrfs_bio *bbio = NULL;
2040         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2041         int ret;
2042
2043         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2044         BUG_ON(!mirror_num);
2045
2046         /* we can't repair anything in raid56 yet */
2047         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2048                 return 0;
2049
2050         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2051         if (!bio)
2052                 return -EIO;
2053         bio->bi_iter.bi_size = 0;
2054         map_length = length;
2055
2056         ret = btrfs_map_block(fs_info, WRITE, logical,
2057                               &map_length, &bbio, mirror_num);
2058         if (ret) {
2059                 bio_put(bio);
2060                 return -EIO;
2061         }
2062         BUG_ON(mirror_num != bbio->mirror_num);
2063         sector = bbio->stripes[mirror_num-1].physical >> 9;
2064         bio->bi_iter.bi_sector = sector;
2065         dev = bbio->stripes[mirror_num-1].dev;
2066         btrfs_put_bbio(bbio);
2067         if (!dev || !dev->bdev || !dev->writeable) {
2068                 bio_put(bio);
2069                 return -EIO;
2070         }
2071         bio->bi_bdev = dev->bdev;
2072         bio_add_page(bio, page, length, pg_offset);
2073
2074         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2075                 /* try to remap that extent elsewhere? */
2076                 bio_put(bio);
2077                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2078                 return -EIO;
2079         }
2080
2081         printk_ratelimited_in_rcu(KERN_INFO
2082                                   "BTRFS: read error corrected: ino %llu off %llu (dev %s sector %llu)\n",
2083                                   btrfs_ino(inode), start,
2084                                   rcu_str_deref(dev->name), sector);
2085         bio_put(bio);
2086         return 0;
2087 }
2088
2089 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2090                          int mirror_num)
2091 {
2092         u64 start = eb->start;
2093         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2094         int ret = 0;
2095
2096         if (root->fs_info->sb->s_flags & MS_RDONLY)
2097                 return -EROFS;
2098
2099         for (i = 0; i < num_pages; i++) {
2100                 struct page *p = eb->pages[i];
2101
2102                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2103                                         PAGE_CACHE_SIZE, start, p,
2104                                         start - page_offset(p), mirror_num);
2105                 if (ret)
2106                         break;
2107                 start += PAGE_CACHE_SIZE;
2108         }
2109
2110         return ret;
2111 }
2112
2113 /*
2114  * each time an IO finishes, we do a fast check in the IO failure tree
2115  * to see if we need to process or clean up an io_failure_record
2116  */
2117 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2118                      unsigned int pg_offset)
2119 {
2120         u64 private;
2121         u64 private_failure;
2122         struct io_failure_record *failrec;
2123         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2124         struct extent_state *state;
2125         int num_copies;
2126         int ret;
2127
2128         private = 0;
2129         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2130                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2131         if (!ret)
2132                 return 0;
2133
2134         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2135                                 &private_failure);
2136         if (ret)
2137                 return 0;
2138
2139         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2140         BUG_ON(!failrec->this_mirror);
2141
2142         if (failrec->in_validation) {
2143                 /* there was no real error, just free the record */
2144                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2145                          failrec->start);
2146                 goto out;
2147         }
2148         if (fs_info->sb->s_flags & MS_RDONLY)
2149                 goto out;
2150
2151         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2152         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2153                                             failrec->start,
2154                                             EXTENT_LOCKED);
2155         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2156
2157         if (state && state->start <= failrec->start &&
2158             state->end >= failrec->start + failrec->len - 1) {
2159                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2160                                               failrec->len);
2161                 if (num_copies > 1)  {
2162                         repair_io_failure(inode, start, failrec->len,
2163                                           failrec->logical, page,
2164                                           pg_offset, failrec->failed_mirror);
2165                 }
2166         }
2167
2168 out:
2169         free_io_failure(inode, failrec);
2170
2171         return 0;
2172 }
2173
2174 /*
2175  * Can be called when
2176  * - hold extent lock
2177  * - under ordered extent
2178  * - the inode is freeing
2179  */
2180 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2181 {
2182         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2183         struct io_failure_record *failrec;
2184         struct extent_state *state, *next;
2185
2186         if (RB_EMPTY_ROOT(&failure_tree->state))
2187                 return;
2188
2189         spin_lock(&failure_tree->lock);
2190         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2191         while (state) {
2192                 if (state->start > end)
2193                         break;
2194
2195                 ASSERT(state->end <= end);
2196
2197                 next = next_state(state);
2198
2199                 failrec = (struct io_failure_record *)(unsigned long)state->private;
2200                 free_extent_state(state);
2201                 kfree(failrec);
2202
2203                 state = next;
2204         }
2205         spin_unlock(&failure_tree->lock);
2206 }
2207
2208 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2209                                 struct io_failure_record **failrec_ret)
2210 {
2211         struct io_failure_record *failrec;
2212         u64 private;
2213         struct extent_map *em;
2214         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2215         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2216         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2217         int ret;
2218         u64 logical;
2219
2220         ret = get_state_private(failure_tree, start, &private);
2221         if (ret) {
2222                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2223                 if (!failrec)
2224                         return -ENOMEM;
2225
2226                 failrec->start = start;
2227                 failrec->len = end - start + 1;
2228                 failrec->this_mirror = 0;
2229                 failrec->bio_flags = 0;
2230                 failrec->in_validation = 0;
2231
2232                 read_lock(&em_tree->lock);
2233                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2234                 if (!em) {
2235                         read_unlock(&em_tree->lock);
2236                         kfree(failrec);
2237                         return -EIO;
2238                 }
2239
2240                 if (em->start > start || em->start + em->len <= start) {
2241                         free_extent_map(em);
2242                         em = NULL;
2243                 }
2244                 read_unlock(&em_tree->lock);
2245                 if (!em) {
2246                         kfree(failrec);
2247                         return -EIO;
2248                 }
2249
2250                 logical = start - em->start;
2251                 logical = em->block_start + logical;
2252                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2253                         logical = em->block_start;
2254                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2255                         extent_set_compress_type(&failrec->bio_flags,
2256                                                  em->compress_type);
2257                 }
2258
2259                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2260                          logical, start, failrec->len);
2261
2262                 failrec->logical = logical;
2263                 free_extent_map(em);
2264
2265                 /* set the bits in the private failure tree */
2266                 ret = set_extent_bits(failure_tree, start, end,
2267                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2268                 if (ret >= 0)
2269                         ret = set_state_private(failure_tree, start,
2270                                                 (u64)(unsigned long)failrec);
2271                 /* set the bits in the inode's tree */
2272                 if (ret >= 0)
2273                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2274                                                 GFP_NOFS);
2275                 if (ret < 0) {
2276                         kfree(failrec);
2277                         return ret;
2278                 }
2279         } else {
2280                 failrec = (struct io_failure_record *)(unsigned long)private;
2281                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2282                          failrec->logical, failrec->start, failrec->len,
2283                          failrec->in_validation);
2284                 /*
2285                  * when data can be on disk more than twice, add to failrec here
2286                  * (e.g. with a list for failed_mirror) to make
2287                  * clean_io_failure() clean all those errors at once.
2288                  */
2289         }
2290
2291         *failrec_ret = failrec;
2292
2293         return 0;
2294 }
2295
2296 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2297                            struct io_failure_record *failrec, int failed_mirror)
2298 {
2299         int num_copies;
2300
2301         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2302                                       failrec->logical, failrec->len);
2303         if (num_copies == 1) {
2304                 /*
2305                  * we only have a single copy of the data, so don't bother with
2306                  * all the retry and error correction code that follows. no
2307                  * matter what the error is, it is very likely to persist.
2308                  */
2309                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2310                          num_copies, failrec->this_mirror, failed_mirror);
2311                 return 0;
2312         }
2313
2314         /*
2315          * there are two premises:
2316          *      a) deliver good data to the caller
2317          *      b) correct the bad sectors on disk
2318          */
2319         if (failed_bio->bi_vcnt > 1) {
2320                 /*
2321                  * to fulfill b), we need to know the exact failing sectors, as
2322                  * we don't want to rewrite any more than the failed ones. thus,
2323                  * we need separate read requests for the failed bio
2324                  *
2325                  * if the following BUG_ON triggers, our validation request got
2326                  * merged. we need separate requests for our algorithm to work.
2327                  */
2328                 BUG_ON(failrec->in_validation);
2329                 failrec->in_validation = 1;
2330                 failrec->this_mirror = failed_mirror;
2331         } else {
2332                 /*
2333                  * we're ready to fulfill a) and b) alongside. get a good copy
2334                  * of the failed sector and if we succeed, we have setup
2335                  * everything for repair_io_failure to do the rest for us.
2336                  */
2337                 if (failrec->in_validation) {
2338                         BUG_ON(failrec->this_mirror != failed_mirror);
2339                         failrec->in_validation = 0;
2340                         failrec->this_mirror = 0;
2341                 }
2342                 failrec->failed_mirror = failed_mirror;
2343                 failrec->this_mirror++;
2344                 if (failrec->this_mirror == failed_mirror)
2345                         failrec->this_mirror++;
2346         }
2347
2348         if (failrec->this_mirror > num_copies) {
2349                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2350                          num_copies, failrec->this_mirror, failed_mirror);
2351                 return 0;
2352         }
2353
2354         return 1;
2355 }
2356
2357
2358 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2359                                     struct io_failure_record *failrec,
2360                                     struct page *page, int pg_offset, int icsum,
2361                                     bio_end_io_t *endio_func, void *data)
2362 {
2363         struct bio *bio;
2364         struct btrfs_io_bio *btrfs_failed_bio;
2365         struct btrfs_io_bio *btrfs_bio;
2366
2367         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2368         if (!bio)
2369                 return NULL;
2370
2371         bio->bi_end_io = endio_func;
2372         bio->bi_iter.bi_sector = failrec->logical >> 9;
2373         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2374         bio->bi_iter.bi_size = 0;
2375         bio->bi_private = data;
2376
2377         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2378         if (btrfs_failed_bio->csum) {
2379                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2380                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2381
2382                 btrfs_bio = btrfs_io_bio(bio);
2383                 btrfs_bio->csum = btrfs_bio->csum_inline;
2384                 icsum *= csum_size;
2385                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2386                        csum_size);
2387         }
2388
2389         bio_add_page(bio, page, failrec->len, pg_offset);
2390
2391         return bio;
2392 }
2393
2394 /*
2395  * this is a generic handler for readpage errors (default
2396  * readpage_io_failed_hook). if other copies exist, read those and write back
2397  * good data to the failed position. does not investigate in remapping the
2398  * failed extent elsewhere, hoping the device will be smart enough to do this as
2399  * needed
2400  */
2401
2402 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2403                               struct page *page, u64 start, u64 end,
2404                               int failed_mirror)
2405 {
2406         struct io_failure_record *failrec;
2407         struct inode *inode = page->mapping->host;
2408         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2409         struct bio *bio;
2410         int read_mode;
2411         int ret;
2412
2413         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2414
2415         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2416         if (ret)
2417                 return ret;
2418
2419         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2420         if (!ret) {
2421                 free_io_failure(inode, failrec);
2422                 return -EIO;
2423         }
2424
2425         if (failed_bio->bi_vcnt > 1)
2426                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2427         else
2428                 read_mode = READ_SYNC;
2429
2430         phy_offset >>= inode->i_sb->s_blocksize_bits;
2431         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2432                                       start - page_offset(page),
2433                                       (int)phy_offset, failed_bio->bi_end_io,
2434                                       NULL);
2435         if (!bio) {
2436                 free_io_failure(inode, failrec);
2437                 return -EIO;
2438         }
2439
2440         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2441                  read_mode, failrec->this_mirror, failrec->in_validation);
2442
2443         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2444                                          failrec->this_mirror,
2445                                          failrec->bio_flags, 0);
2446         if (ret) {
2447                 free_io_failure(inode, failrec);
2448                 bio_put(bio);
2449         }
2450
2451         return ret;
2452 }
2453
2454 /* lots and lots of room for performance fixes in the end_bio funcs */
2455
2456 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2457 {
2458         int uptodate = (err == 0);
2459         struct extent_io_tree *tree;
2460         int ret = 0;
2461
2462         tree = &BTRFS_I(page->mapping->host)->io_tree;
2463
2464         if (tree->ops && tree->ops->writepage_end_io_hook) {
2465                 ret = tree->ops->writepage_end_io_hook(page, start,
2466                                                end, NULL, uptodate);
2467                 if (ret)
2468                         uptodate = 0;
2469         }
2470
2471         if (!uptodate) {
2472                 ClearPageUptodate(page);
2473                 SetPageError(page);
2474                 ret = ret < 0 ? ret : -EIO;
2475                 mapping_set_error(page->mapping, ret);
2476         }
2477         return 0;
2478 }
2479
2480 /*
2481  * after a writepage IO is done, we need to:
2482  * clear the uptodate bits on error
2483  * clear the writeback bits in the extent tree for this IO
2484  * end_page_writeback if the page has no more pending IO
2485  *
2486  * Scheduling is not allowed, so the extent state tree is expected
2487  * to have one and only one object corresponding to this IO.
2488  */
2489 static void end_bio_extent_writepage(struct bio *bio, int err)
2490 {
2491         struct bio_vec *bvec;
2492         u64 start;
2493         u64 end;
2494         int i;
2495
2496         bio_for_each_segment_all(bvec, bio, i) {
2497                 struct page *page = bvec->bv_page;
2498
2499                 /* We always issue full-page reads, but if some block
2500                  * in a page fails to read, blk_update_request() will
2501                  * advance bv_offset and adjust bv_len to compensate.
2502                  * Print a warning for nonzero offsets, and an error
2503                  * if they don't add up to a full page.  */
2504                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2505                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2506                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2507                                    "partial page write in btrfs with offset %u and length %u",
2508                                         bvec->bv_offset, bvec->bv_len);
2509                         else
2510                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2511                                    "incomplete page write in btrfs with offset %u and "
2512                                    "length %u",
2513                                         bvec->bv_offset, bvec->bv_len);
2514                 }
2515
2516                 start = page_offset(page);
2517                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2518
2519                 if (end_extent_writepage(page, err, start, end))
2520                         continue;
2521
2522                 end_page_writeback(page);
2523         }
2524
2525         bio_put(bio);
2526 }
2527
2528 static void
2529 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2530                               int uptodate)
2531 {
2532         struct extent_state *cached = NULL;
2533         u64 end = start + len - 1;
2534
2535         if (uptodate && tree->track_uptodate)
2536                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2537         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2538 }
2539
2540 /*
2541  * after a readpage IO is done, we need to:
2542  * clear the uptodate bits on error
2543  * set the uptodate bits if things worked
2544  * set the page up to date if all extents in the tree are uptodate
2545  * clear the lock bit in the extent tree
2546  * unlock the page if there are no other extents locked for it
2547  *
2548  * Scheduling is not allowed, so the extent state tree is expected
2549  * to have one and only one object corresponding to this IO.
2550  */
2551 static void end_bio_extent_readpage(struct bio *bio, int err)
2552 {
2553         struct bio_vec *bvec;
2554         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2555         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2556         struct extent_io_tree *tree;
2557         u64 offset = 0;
2558         u64 start;
2559         u64 end;
2560         u64 len;
2561         u64 extent_start = 0;
2562         u64 extent_len = 0;
2563         int mirror;
2564         int ret;
2565         int i;
2566
2567         if (err)
2568                 uptodate = 0;
2569
2570         bio_for_each_segment_all(bvec, bio, i) {
2571                 struct page *page = bvec->bv_page;
2572                 struct inode *inode = page->mapping->host;
2573
2574                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2575                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector, err,
2576                          io_bio->mirror_num);
2577                 tree = &BTRFS_I(inode)->io_tree;
2578
2579                 /* We always issue full-page reads, but if some block
2580                  * in a page fails to read, blk_update_request() will
2581                  * advance bv_offset and adjust bv_len to compensate.
2582                  * Print a warning for nonzero offsets, and an error
2583                  * if they don't add up to a full page.  */
2584                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2585                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2586                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2587                                    "partial page read in btrfs with offset %u and length %u",
2588                                         bvec->bv_offset, bvec->bv_len);
2589                         else
2590                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2591                                    "incomplete page read in btrfs with offset %u and "
2592                                    "length %u",
2593                                         bvec->bv_offset, bvec->bv_len);
2594                 }
2595
2596                 start = page_offset(page);
2597                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2598                 len = bvec->bv_len;
2599
2600                 mirror = io_bio->mirror_num;
2601                 if (likely(uptodate && tree->ops &&
2602                            tree->ops->readpage_end_io_hook)) {
2603                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2604                                                               page, start, end,
2605                                                               mirror);
2606                         if (ret)
2607                                 uptodate = 0;
2608                         else
2609                                 clean_io_failure(inode, start, page, 0);
2610                 }
2611
2612                 if (likely(uptodate))
2613                         goto readpage_ok;
2614
2615                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2616                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2617                         if (!ret && !err &&
2618                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2619                                 uptodate = 1;
2620                 } else {
2621                         /*
2622                          * The generic bio_readpage_error handles errors the
2623                          * following way: If possible, new read requests are
2624                          * created and submitted and will end up in
2625                          * end_bio_extent_readpage as well (if we're lucky, not
2626                          * in the !uptodate case). In that case it returns 0 and
2627                          * we just go on with the next page in our bio. If it
2628                          * can't handle the error it will return -EIO and we
2629                          * remain responsible for that page.
2630                          */
2631                         ret = bio_readpage_error(bio, offset, page, start, end,
2632                                                  mirror);
2633                         if (ret == 0) {
2634                                 uptodate =
2635                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2636                                 if (err)
2637                                         uptodate = 0;
2638                                 offset += len;
2639                                 continue;
2640                         }
2641                 }
2642 readpage_ok:
2643                 if (likely(uptodate)) {
2644                         loff_t i_size = i_size_read(inode);
2645                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2646                         unsigned off;
2647
2648                         /* Zero out the end if this page straddles i_size */
2649                         off = i_size & (PAGE_CACHE_SIZE-1);
2650                         if (page->index == end_index && off)
2651                                 zero_user_segment(page, off, PAGE_CACHE_SIZE);
2652                         SetPageUptodate(page);
2653                 } else {
2654                         ClearPageUptodate(page);
2655                         SetPageError(page);
2656                 }
2657                 unlock_page(page);
2658                 offset += len;
2659
2660                 if (unlikely(!uptodate)) {
2661                         if (extent_len) {
2662                                 endio_readpage_release_extent(tree,
2663                                                               extent_start,
2664                                                               extent_len, 1);
2665                                 extent_start = 0;
2666                                 extent_len = 0;
2667                         }
2668                         endio_readpage_release_extent(tree, start,
2669                                                       end - start + 1, 0);
2670                 } else if (!extent_len) {
2671                         extent_start = start;
2672                         extent_len = end + 1 - start;
2673                 } else if (extent_start + extent_len == start) {
2674                         extent_len += end + 1 - start;
2675                 } else {
2676                         endio_readpage_release_extent(tree, extent_start,
2677                                                       extent_len, uptodate);
2678                         extent_start = start;
2679                         extent_len = end + 1 - start;
2680                 }
2681         }
2682
2683         if (extent_len)
2684                 endio_readpage_release_extent(tree, extent_start, extent_len,
2685                                               uptodate);
2686         if (io_bio->end_io)
2687                 io_bio->end_io(io_bio, err);
2688         bio_put(bio);
2689 }
2690
2691 /*
2692  * this allocates from the btrfs_bioset.  We're returning a bio right now
2693  * but you can call btrfs_io_bio for the appropriate container_of magic
2694  */
2695 struct bio *
2696 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2697                 gfp_t gfp_flags)
2698 {
2699         struct btrfs_io_bio *btrfs_bio;
2700         struct bio *bio;
2701
2702         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2703
2704         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2705                 while (!bio && (nr_vecs /= 2)) {
2706                         bio = bio_alloc_bioset(gfp_flags,
2707                                                nr_vecs, btrfs_bioset);
2708                 }
2709         }
2710
2711         if (bio) {
2712                 bio->bi_bdev = bdev;
2713                 bio->bi_iter.bi_sector = first_sector;
2714                 btrfs_bio = btrfs_io_bio(bio);
2715                 btrfs_bio->csum = NULL;
2716                 btrfs_bio->csum_allocated = NULL;
2717                 btrfs_bio->end_io = NULL;
2718         }
2719         return bio;
2720 }
2721
2722 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2723 {
2724         struct btrfs_io_bio *btrfs_bio;
2725         struct bio *new;
2726
2727         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2728         if (new) {
2729                 btrfs_bio = btrfs_io_bio(new);
2730                 btrfs_bio->csum = NULL;
2731                 btrfs_bio->csum_allocated = NULL;
2732                 btrfs_bio->end_io = NULL;
2733         }
2734         return new;
2735 }
2736
2737 /* this also allocates from the btrfs_bioset */
2738 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2739 {
2740         struct btrfs_io_bio *btrfs_bio;
2741         struct bio *bio;
2742
2743         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2744         if (bio) {
2745                 btrfs_bio = btrfs_io_bio(bio);
2746                 btrfs_bio->csum = NULL;
2747                 btrfs_bio->csum_allocated = NULL;
2748                 btrfs_bio->end_io = NULL;
2749         }
2750         return bio;
2751 }
2752
2753
2754 static int __must_check submit_one_bio(int rw, struct bio *bio,
2755                                        int mirror_num, unsigned long bio_flags)
2756 {
2757         int ret = 0;
2758         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2759         struct page *page = bvec->bv_page;
2760         struct extent_io_tree *tree = bio->bi_private;
2761         u64 start;
2762
2763         start = page_offset(page) + bvec->bv_offset;
2764
2765         bio->bi_private = NULL;
2766
2767         bio_get(bio);
2768
2769         if (tree->ops && tree->ops->submit_bio_hook)
2770                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2771                                            mirror_num, bio_flags, start);
2772         else
2773                 btrfsic_submit_bio(rw, bio);
2774
2775         bio_put(bio);
2776         return ret;
2777 }
2778
2779 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2780                      unsigned long offset, size_t size, struct bio *bio,
2781                      unsigned long bio_flags)
2782 {
2783         int ret = 0;
2784         if (tree->ops && tree->ops->merge_bio_hook)
2785                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2786                                                 bio_flags);
2787         BUG_ON(ret < 0);
2788         return ret;
2789
2790 }
2791
2792 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2793                               struct page *page, sector_t sector,
2794                               size_t size, unsigned long offset,
2795                               struct block_device *bdev,
2796                               struct bio **bio_ret,
2797                               unsigned long max_pages,
2798                               bio_end_io_t end_io_func,
2799                               int mirror_num,
2800                               unsigned long prev_bio_flags,
2801                               unsigned long bio_flags)
2802 {
2803         int ret = 0;
2804         struct bio *bio;
2805         int nr;
2806         int contig = 0;
2807         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2808         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2809         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2810
2811         if (bio_ret && *bio_ret) {
2812                 bio = *bio_ret;
2813                 if (old_compressed)
2814                         contig = bio->bi_iter.bi_sector == sector;
2815                 else
2816                         contig = bio_end_sector(bio) == sector;
2817
2818                 if (prev_bio_flags != bio_flags || !contig ||
2819                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2820                     bio_add_page(bio, page, page_size, offset) < page_size) {
2821                         ret = submit_one_bio(rw, bio, mirror_num,
2822                                              prev_bio_flags);
2823                         if (ret < 0) {
2824                                 *bio_ret = NULL;
2825                                 return ret;
2826                         }
2827                         bio = NULL;
2828                 } else {
2829                         return 0;
2830                 }
2831         }
2832         if (this_compressed)
2833                 nr = BIO_MAX_PAGES;
2834         else
2835                 nr = bio_get_nr_vecs(bdev);
2836
2837         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2838         if (!bio)
2839                 return -ENOMEM;
2840
2841         bio_add_page(bio, page, page_size, offset);
2842         bio->bi_end_io = end_io_func;
2843         bio->bi_private = tree;
2844
2845         if (bio_ret)
2846                 *bio_ret = bio;
2847         else
2848                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2849
2850         return ret;
2851 }
2852
2853 static void attach_extent_buffer_page(struct extent_buffer *eb,
2854                                       struct page *page)
2855 {
2856         if (!PagePrivate(page)) {
2857                 SetPagePrivate(page);
2858                 page_cache_get(page);
2859                 set_page_private(page, (unsigned long)eb);
2860         } else {
2861                 WARN_ON(page->private != (unsigned long)eb);
2862         }
2863 }
2864
2865 void set_page_extent_mapped(struct page *page)
2866 {
2867         if (!PagePrivate(page)) {
2868                 SetPagePrivate(page);
2869                 page_cache_get(page);
2870                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2871         }
2872 }
2873
2874 static struct extent_map *
2875 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2876                  u64 start, u64 len, get_extent_t *get_extent,
2877                  struct extent_map **em_cached)
2878 {
2879         struct extent_map *em;
2880
2881         if (em_cached && *em_cached) {
2882                 em = *em_cached;
2883                 if (extent_map_in_tree(em) && start >= em->start &&
2884                     start < extent_map_end(em)) {
2885                         atomic_inc(&em->refs);
2886                         return em;
2887                 }
2888
2889                 free_extent_map(em);
2890                 *em_cached = NULL;
2891         }
2892
2893         em = get_extent(inode, page, pg_offset, start, len, 0);
2894         if (em_cached && !IS_ERR_OR_NULL(em)) {
2895                 BUG_ON(*em_cached);
2896                 atomic_inc(&em->refs);
2897                 *em_cached = em;
2898         }
2899         return em;
2900 }
2901 /*
2902  * basic readpage implementation.  Locked extent state structs are inserted
2903  * into the tree that are removed when the IO is done (by the end_io
2904  * handlers)
2905  * XXX JDM: This needs looking at to ensure proper page locking
2906  */
2907 static int __do_readpage(struct extent_io_tree *tree,
2908                          struct page *page,
2909                          get_extent_t *get_extent,
2910                          struct extent_map **em_cached,
2911                          struct bio **bio, int mirror_num,
2912                          unsigned long *bio_flags, int rw)
2913 {
2914         struct inode *inode = page->mapping->host;
2915         u64 start = page_offset(page);
2916         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2917         u64 end;
2918         u64 cur = start;
2919         u64 extent_offset;
2920         u64 last_byte = i_size_read(inode);
2921         u64 block_start;
2922         u64 cur_end;
2923         sector_t sector;
2924         struct extent_map *em;
2925         struct block_device *bdev;
2926         int ret;
2927         int nr = 0;
2928         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2929         size_t pg_offset = 0;
2930         size_t iosize;
2931         size_t disk_io_size;
2932         size_t blocksize = inode->i_sb->s_blocksize;
2933         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2934
2935         set_page_extent_mapped(page);
2936
2937         end = page_end;
2938         if (!PageUptodate(page)) {
2939                 if (cleancache_get_page(page) == 0) {
2940                         BUG_ON(blocksize != PAGE_SIZE);
2941                         unlock_extent(tree, start, end);
2942                         goto out;
2943                 }
2944         }
2945
2946         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2947                 char *userpage;
2948                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2949
2950                 if (zero_offset) {
2951                         iosize = PAGE_CACHE_SIZE - zero_offset;
2952                         userpage = kmap_atomic(page);
2953                         memset(userpage + zero_offset, 0, iosize);
2954                         flush_dcache_page(page);
2955                         kunmap_atomic(userpage);
2956                 }
2957         }
2958         while (cur <= end) {
2959                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2960
2961                 if (cur >= last_byte) {
2962                         char *userpage;
2963                         struct extent_state *cached = NULL;
2964
2965                         iosize = PAGE_CACHE_SIZE - pg_offset;
2966                         userpage = kmap_atomic(page);
2967                         memset(userpage + pg_offset, 0, iosize);
2968                         flush_dcache_page(page);
2969                         kunmap_atomic(userpage);
2970                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2971                                             &cached, GFP_NOFS);
2972                         if (!parent_locked)
2973                                 unlock_extent_cached(tree, cur,
2974                                                      cur + iosize - 1,
2975                                                      &cached, GFP_NOFS);
2976                         break;
2977                 }
2978                 em = __get_extent_map(inode, page, pg_offset, cur,
2979                                       end - cur + 1, get_extent, em_cached);
2980                 if (IS_ERR_OR_NULL(em)) {
2981                         SetPageError(page);
2982                         if (!parent_locked)
2983                                 unlock_extent(tree, cur, end);
2984                         break;
2985                 }
2986                 extent_offset = cur - em->start;
2987                 BUG_ON(extent_map_end(em) <= cur);
2988                 BUG_ON(end < cur);
2989
2990                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2991                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2992                         extent_set_compress_type(&this_bio_flag,
2993                                                  em->compress_type);
2994                 }
2995
2996                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2997                 cur_end = min(extent_map_end(em) - 1, end);
2998                 iosize = ALIGN(iosize, blocksize);
2999                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
3000                         disk_io_size = em->block_len;
3001                         sector = em->block_start >> 9;
3002                 } else {
3003                         sector = (em->block_start + extent_offset) >> 9;
3004                         disk_io_size = iosize;
3005                 }
3006                 bdev = em->bdev;
3007                 block_start = em->block_start;
3008                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3009                         block_start = EXTENT_MAP_HOLE;
3010                 free_extent_map(em);
3011                 em = NULL;
3012
3013                 /* we've found a hole, just zero and go on */
3014                 if (block_start == EXTENT_MAP_HOLE) {
3015                         char *userpage;
3016                         struct extent_state *cached = NULL;
3017
3018                         userpage = kmap_atomic(page);
3019                         memset(userpage + pg_offset, 0, iosize);
3020                         flush_dcache_page(page);
3021                         kunmap_atomic(userpage);
3022
3023                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3024                                             &cached, GFP_NOFS);
3025                         unlock_extent_cached(tree, cur, cur + iosize - 1,
3026                                              &cached, GFP_NOFS);
3027                         cur = cur + iosize;
3028                         pg_offset += iosize;
3029                         continue;
3030                 }
3031                 /* the get_extent function already copied into the page */
3032                 if (test_range_bit(tree, cur, cur_end,
3033                                    EXTENT_UPTODATE, 1, NULL)) {
3034                         check_page_uptodate(tree, page);
3035                         if (!parent_locked)
3036                                 unlock_extent(tree, cur, cur + iosize - 1);
3037                         cur = cur + iosize;
3038                         pg_offset += iosize;
3039                         continue;
3040                 }
3041                 /* we have an inline extent but it didn't get marked up
3042                  * to date.  Error out
3043                  */
3044                 if (block_start == EXTENT_MAP_INLINE) {
3045                         SetPageError(page);
3046                         if (!parent_locked)
3047                                 unlock_extent(tree, cur, cur + iosize - 1);
3048                         cur = cur + iosize;
3049                         pg_offset += iosize;
3050                         continue;
3051                 }
3052
3053                 pnr -= page->index;
3054                 ret = submit_extent_page(rw, tree, page,
3055                                          sector, disk_io_size, pg_offset,
3056                                          bdev, bio, pnr,
3057                                          end_bio_extent_readpage, mirror_num,
3058                                          *bio_flags,
3059                                          this_bio_flag);
3060                 if (!ret) {
3061                         nr++;
3062                         *bio_flags = this_bio_flag;
3063                 } else {
3064                         SetPageError(page);
3065                         if (!parent_locked)
3066                                 unlock_extent(tree, cur, cur + iosize - 1);
3067                 }
3068                 cur = cur + iosize;
3069                 pg_offset += iosize;
3070         }
3071 out:
3072         if (!nr) {
3073                 if (!PageError(page))
3074                         SetPageUptodate(page);
3075                 unlock_page(page);
3076         }
3077         return 0;
3078 }
3079
3080 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3081                                              struct page *pages[], int nr_pages,
3082                                              u64 start, u64 end,
3083                                              get_extent_t *get_extent,
3084                                              struct extent_map **em_cached,
3085                                              struct bio **bio, int mirror_num,
3086                                              unsigned long *bio_flags, int rw)
3087 {
3088         struct inode *inode;
3089         struct btrfs_ordered_extent *ordered;
3090         int index;
3091
3092         inode = pages[0]->mapping->host;
3093         while (1) {
3094                 lock_extent(tree, start, end);
3095                 ordered = btrfs_lookup_ordered_range(inode, start,
3096                                                      end - start + 1);
3097                 if (!ordered)
3098                         break;
3099                 unlock_extent(tree, start, end);
3100                 btrfs_start_ordered_extent(inode, ordered, 1);
3101                 btrfs_put_ordered_extent(ordered);
3102         }
3103
3104         for (index = 0; index < nr_pages; index++) {
3105                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3106                               mirror_num, bio_flags, rw);
3107                 page_cache_release(pages[index]);
3108         }
3109 }
3110
3111 static void __extent_readpages(struct extent_io_tree *tree,
3112                                struct page *pages[],
3113                                int nr_pages, get_extent_t *get_extent,
3114                                struct extent_map **em_cached,
3115                                struct bio **bio, int mirror_num,
3116                                unsigned long *bio_flags, int rw)
3117 {
3118         u64 start = 0;
3119         u64 end = 0;
3120         u64 page_start;
3121         int index;
3122         int first_index = 0;
3123
3124         for (index = 0; index < nr_pages; index++) {
3125                 page_start = page_offset(pages[index]);
3126                 if (!end) {
3127                         start = page_start;
3128                         end = start + PAGE_CACHE_SIZE - 1;
3129                         first_index = index;
3130                 } else if (end + 1 == page_start) {
3131                         end += PAGE_CACHE_SIZE;
3132                 } else {
3133                         __do_contiguous_readpages(tree, &pages[first_index],
3134                                                   index - first_index, start,
3135                                                   end, get_extent, em_cached,
3136                                                   bio, mirror_num, bio_flags,
3137                                                   rw);
3138                         start = page_start;
3139                         end = start + PAGE_CACHE_SIZE - 1;
3140                         first_index = index;
3141                 }
3142         }
3143
3144         if (end)
3145                 __do_contiguous_readpages(tree, &pages[first_index],
3146                                           index - first_index, start,
3147                                           end, get_extent, em_cached, bio,
3148                                           mirror_num, bio_flags, rw);
3149 }
3150
3151 static int __extent_read_full_page(struct extent_io_tree *tree,
3152                                    struct page *page,
3153                                    get_extent_t *get_extent,
3154                                    struct bio **bio, int mirror_num,
3155                                    unsigned long *bio_flags, int rw)
3156 {
3157         struct inode *inode = page->mapping->host;
3158         struct btrfs_ordered_extent *ordered;
3159         u64 start = page_offset(page);
3160         u64 end = start + PAGE_CACHE_SIZE - 1;
3161         int ret;
3162
3163         while (1) {
3164                 lock_extent(tree, start, end);
3165                 ordered = btrfs_lookup_ordered_extent(inode, start);
3166                 if (!ordered)
3167                         break;
3168                 unlock_extent(tree, start, end);
3169                 btrfs_start_ordered_extent(inode, ordered, 1);
3170                 btrfs_put_ordered_extent(ordered);
3171         }
3172
3173         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3174                             bio_flags, rw);
3175         return ret;
3176 }
3177
3178 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3179                             get_extent_t *get_extent, int mirror_num)
3180 {
3181         struct bio *bio = NULL;
3182         unsigned long bio_flags = 0;
3183         int ret;
3184
3185         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3186                                       &bio_flags, READ);
3187         if (bio)
3188                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3189         return ret;
3190 }
3191
3192 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3193                                  get_extent_t *get_extent, int mirror_num)
3194 {
3195         struct bio *bio = NULL;
3196         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3197         int ret;
3198
3199         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3200                                       &bio_flags, READ);
3201         if (bio)
3202                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3203         return ret;
3204 }
3205
3206 static noinline void update_nr_written(struct page *page,
3207                                       struct writeback_control *wbc,
3208                                       unsigned long nr_written)
3209 {
3210         wbc->nr_to_write -= nr_written;
3211         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3212             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3213                 page->mapping->writeback_index = page->index + nr_written;
3214 }
3215
3216 /*
3217  * helper for __extent_writepage, doing all of the delayed allocation setup.
3218  *
3219  * This returns 1 if our fill_delalloc function did all the work required
3220  * to write the page (copy into inline extent).  In this case the IO has
3221  * been started and the page is already unlocked.
3222  *
3223  * This returns 0 if all went well (page still locked)
3224  * This returns < 0 if there were errors (page still locked)
3225  */
3226 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3227                               struct page *page, struct writeback_control *wbc,
3228                               struct extent_page_data *epd,
3229                               u64 delalloc_start,
3230                               unsigned long *nr_written)
3231 {
3232         struct extent_io_tree *tree = epd->tree;
3233         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3234         u64 nr_delalloc;
3235         u64 delalloc_to_write = 0;
3236         u64 delalloc_end = 0;
3237         int ret;
3238         int page_started = 0;
3239
3240         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3241                 return 0;
3242
3243         while (delalloc_end < page_end) {
3244                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3245                                                page,
3246                                                &delalloc_start,
3247                                                &delalloc_end,
3248                                                BTRFS_MAX_EXTENT_SIZE);
3249                 if (nr_delalloc == 0) {
3250                         delalloc_start = delalloc_end + 1;
3251                         continue;
3252                 }
3253                 ret = tree->ops->fill_delalloc(inode, page,
3254                                                delalloc_start,
3255                                                delalloc_end,
3256                                                &page_started,
3257                                                nr_written);
3258                 /* File system has been set read-only */
3259                 if (ret) {
3260                         SetPageError(page);
3261                         /* fill_delalloc should be return < 0 for error
3262                          * but just in case, we use > 0 here meaning the
3263                          * IO is started, so we don't want to return > 0
3264                          * unless things are going well.
3265                          */
3266                         ret = ret < 0 ? ret : -EIO;
3267                         goto done;
3268                 }
3269                 /*
3270                  * delalloc_end is already one less than the total
3271                  * length, so we don't subtract one from
3272                  * PAGE_CACHE_SIZE
3273                  */
3274                 delalloc_to_write += (delalloc_end - delalloc_start +
3275                                       PAGE_CACHE_SIZE) >>
3276                                       PAGE_CACHE_SHIFT;
3277                 delalloc_start = delalloc_end + 1;
3278         }
3279         if (wbc->nr_to_write < delalloc_to_write) {
3280                 int thresh = 8192;
3281
3282                 if (delalloc_to_write < thresh * 2)
3283                         thresh = delalloc_to_write;
3284                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3285                                          thresh);
3286         }
3287
3288         /* did the fill delalloc function already unlock and start
3289          * the IO?
3290          */
3291         if (page_started) {
3292                 /*
3293                  * we've unlocked the page, so we can't update
3294                  * the mapping's writeback index, just update
3295                  * nr_to_write.
3296                  */
3297                 wbc->nr_to_write -= *nr_written;
3298                 return 1;
3299         }
3300
3301         ret = 0;
3302
3303 done:
3304         return ret;
3305 }
3306
3307 /*
3308  * helper for __extent_writepage.  This calls the writepage start hooks,
3309  * and does the loop to map the page into extents and bios.
3310  *
3311  * We return 1 if the IO is started and the page is unlocked,
3312  * 0 if all went well (page still locked)
3313  * < 0 if there were errors (page still locked)
3314  */
3315 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3316                                  struct page *page,
3317                                  struct writeback_control *wbc,
3318                                  struct extent_page_data *epd,
3319                                  loff_t i_size,
3320                                  unsigned long nr_written,
3321                                  int write_flags, int *nr_ret)
3322 {
3323         struct extent_io_tree *tree = epd->tree;
3324         u64 start = page_offset(page);
3325         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3326         u64 end;
3327         u64 cur = start;
3328         u64 extent_offset;
3329         u64 block_start;
3330         u64 iosize;
3331         sector_t sector;
3332         struct extent_state *cached_state = NULL;
3333         struct extent_map *em;
3334         struct block_device *bdev;
3335         size_t pg_offset = 0;
3336         size_t blocksize;
3337         int ret = 0;
3338         int nr = 0;
3339         bool compressed;
3340
3341         if (tree->ops && tree->ops->writepage_start_hook) {
3342                 ret = tree->ops->writepage_start_hook(page, start,
3343                                                       page_end);
3344                 if (ret) {
3345                         /* Fixup worker will requeue */
3346                         if (ret == -EBUSY)
3347                                 wbc->pages_skipped++;
3348                         else
3349                                 redirty_page_for_writepage(wbc, page);
3350
3351                         update_nr_written(page, wbc, nr_written);
3352                         unlock_page(page);
3353                         ret = 1;
3354                         goto done_unlocked;
3355                 }
3356         }
3357
3358         /*
3359          * we don't want to touch the inode after unlocking the page,
3360          * so we update the mapping writeback index now
3361          */
3362         update_nr_written(page, wbc, nr_written + 1);
3363
3364         end = page_end;
3365         if (i_size <= start) {
3366                 if (tree->ops && tree->ops->writepage_end_io_hook)
3367                         tree->ops->writepage_end_io_hook(page, start,
3368                                                          page_end, NULL, 1);
3369                 goto done;
3370         }
3371
3372         blocksize = inode->i_sb->s_blocksize;
3373
3374         while (cur <= end) {
3375                 u64 em_end;
3376                 if (cur >= i_size) {
3377                         if (tree->ops && tree->ops->writepage_end_io_hook)
3378                                 tree->ops->writepage_end_io_hook(page, cur,
3379                                                          page_end, NULL, 1);
3380                         break;
3381                 }
3382                 em = epd->get_extent(inode, page, pg_offset, cur,
3383                                      end - cur + 1, 1);
3384                 if (IS_ERR_OR_NULL(em)) {
3385                         SetPageError(page);
3386                         ret = PTR_ERR_OR_ZERO(em);
3387                         break;
3388                 }
3389
3390                 extent_offset = cur - em->start;
3391                 em_end = extent_map_end(em);
3392                 BUG_ON(em_end <= cur);
3393                 BUG_ON(end < cur);
3394                 iosize = min(em_end - cur, end - cur + 1);
3395                 iosize = ALIGN(iosize, blocksize);
3396                 sector = (em->block_start + extent_offset) >> 9;
3397                 bdev = em->bdev;
3398                 block_start = em->block_start;
3399                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3400                 free_extent_map(em);
3401                 em = NULL;
3402
3403                 /*
3404                  * compressed and inline extents are written through other
3405                  * paths in the FS
3406                  */
3407                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3408                     block_start == EXTENT_MAP_INLINE) {
3409                         /*
3410                          * end_io notification does not happen here for
3411                          * compressed extents
3412                          */
3413                         if (!compressed && tree->ops &&
3414                             tree->ops->writepage_end_io_hook)
3415                                 tree->ops->writepage_end_io_hook(page, cur,
3416                                                          cur + iosize - 1,
3417                                                          NULL, 1);
3418                         else if (compressed) {
3419                                 /* we don't want to end_page_writeback on
3420                                  * a compressed extent.  this happens
3421                                  * elsewhere
3422                                  */
3423                                 nr++;
3424                         }
3425
3426                         cur += iosize;
3427                         pg_offset += iosize;
3428                         continue;
3429                 }
3430
3431                 if (tree->ops && tree->ops->writepage_io_hook) {
3432                         ret = tree->ops->writepage_io_hook(page, cur,
3433                                                 cur + iosize - 1);
3434                 } else {
3435                         ret = 0;
3436                 }
3437                 if (ret) {
3438                         SetPageError(page);
3439                 } else {
3440                         unsigned long max_nr = (i_size >> PAGE_CACHE_SHIFT) + 1;
3441
3442                         set_range_writeback(tree, cur, cur + iosize - 1);
3443                         if (!PageWriteback(page)) {
3444                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3445                                            "page %lu not writeback, cur %llu end %llu",
3446                                        page->index, cur, end);
3447                         }
3448
3449                         ret = submit_extent_page(write_flags, tree, page,
3450                                                  sector, iosize, pg_offset,
3451                                                  bdev, &epd->bio, max_nr,
3452                                                  end_bio_extent_writepage,
3453                                                  0, 0, 0);
3454                         if (ret)
3455                                 SetPageError(page);
3456                 }
3457                 cur = cur + iosize;
3458                 pg_offset += iosize;
3459                 nr++;
3460         }
3461 done:
3462         *nr_ret = nr;
3463
3464 done_unlocked:
3465
3466         /* drop our reference on any cached states */
3467         free_extent_state(cached_state);
3468         return ret;
3469 }
3470
3471 /*
3472  * the writepage semantics are similar to regular writepage.  extent
3473  * records are inserted to lock ranges in the tree, and as dirty areas
3474  * are found, they are marked writeback.  Then the lock bits are removed
3475  * and the end_io handler clears the writeback ranges
3476  */
3477 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3478                               void *data)
3479 {
3480         struct inode *inode = page->mapping->host;
3481         struct extent_page_data *epd = data;
3482         u64 start = page_offset(page);
3483         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3484         int ret;
3485         int nr = 0;
3486         size_t pg_offset = 0;
3487         loff_t i_size = i_size_read(inode);
3488         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3489         int write_flags;
3490         unsigned long nr_written = 0;
3491
3492         if (wbc->sync_mode == WB_SYNC_ALL)
3493                 write_flags = WRITE_SYNC;
3494         else
3495                 write_flags = WRITE;
3496
3497         trace___extent_writepage(page, inode, wbc);
3498
3499         WARN_ON(!PageLocked(page));
3500
3501         ClearPageError(page);
3502
3503         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3504         if (page->index > end_index ||
3505            (page->index == end_index && !pg_offset)) {
3506                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3507                 unlock_page(page);
3508                 return 0;
3509         }
3510
3511         if (page->index == end_index) {
3512                 char *userpage;
3513
3514                 userpage = kmap_atomic(page);
3515                 memset(userpage + pg_offset, 0,
3516                        PAGE_CACHE_SIZE - pg_offset);
3517                 kunmap_atomic(userpage);
3518                 flush_dcache_page(page);
3519         }
3520
3521         pg_offset = 0;
3522
3523         set_page_extent_mapped(page);
3524
3525         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3526         if (ret == 1)
3527                 goto done_unlocked;
3528         if (ret)
3529                 goto done;
3530
3531         ret = __extent_writepage_io(inode, page, wbc, epd,
3532                                     i_size, nr_written, write_flags, &nr);
3533         if (ret == 1)
3534                 goto done_unlocked;
3535
3536 done:
3537         if (nr == 0) {
3538                 /* make sure the mapping tag for page dirty gets cleared */
3539                 set_page_writeback(page);
3540                 end_page_writeback(page);
3541         }
3542         if (PageError(page)) {
3543                 ret = ret < 0 ? ret : -EIO;
3544                 end_extent_writepage(page, ret, start, page_end);
3545         }
3546         unlock_page(page);
3547         return ret;
3548
3549 done_unlocked:
3550         return 0;
3551 }
3552
3553 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3554 {
3555         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3556                        TASK_UNINTERRUPTIBLE);
3557 }
3558
3559 static noinline_for_stack int
3560 lock_extent_buffer_for_io(struct extent_buffer *eb,
3561                           struct btrfs_fs_info *fs_info,
3562                           struct extent_page_data *epd)
3563 {
3564         unsigned long i, num_pages;
3565         int flush = 0;
3566         int ret = 0;
3567
3568         if (!btrfs_try_tree_write_lock(eb)) {
3569                 flush = 1;
3570                 flush_write_bio(epd);
3571                 btrfs_tree_lock(eb);
3572         }
3573
3574         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3575                 btrfs_tree_unlock(eb);
3576                 if (!epd->sync_io)
3577                         return 0;
3578                 if (!flush) {
3579                         flush_write_bio(epd);
3580                         flush = 1;
3581                 }
3582                 while (1) {
3583                         wait_on_extent_buffer_writeback(eb);
3584                         btrfs_tree_lock(eb);
3585                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3586                                 break;
3587                         btrfs_tree_unlock(eb);
3588                 }
3589         }
3590
3591         /*
3592          * We need to do this to prevent races in people who check if the eb is
3593          * under IO since we can end up having no IO bits set for a short period
3594          * of time.
3595          */
3596         spin_lock(&eb->refs_lock);
3597         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3598                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3599                 spin_unlock(&eb->refs_lock);
3600                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3601                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3602                                      -eb->len,
3603                                      fs_info->dirty_metadata_batch);
3604                 ret = 1;
3605         } else {
3606                 spin_unlock(&eb->refs_lock);
3607         }
3608
3609         btrfs_tree_unlock(eb);
3610
3611         if (!ret)
3612                 return ret;
3613
3614         num_pages = num_extent_pages(eb->start, eb->len);
3615         for (i = 0; i < num_pages; i++) {
3616                 struct page *p = eb->pages[i];
3617
3618                 if (!trylock_page(p)) {
3619                         if (!flush) {
3620                                 flush_write_bio(epd);
3621                                 flush = 1;
3622                         }
3623                         lock_page(p);
3624                 }
3625         }
3626
3627         return ret;
3628 }
3629
3630 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3631 {
3632         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3633         smp_mb__after_atomic();
3634         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3635 }
3636
3637 static void set_btree_ioerr(struct page *page)
3638 {
3639         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3640         struct btrfs_inode *btree_ino = BTRFS_I(eb->fs_info->btree_inode);
3641
3642         SetPageError(page);
3643         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3644                 return;
3645
3646         /*
3647          * If writeback for a btree extent that doesn't belong to a log tree
3648          * failed, increment the counter transaction->eb_write_errors.
3649          * We do this because while the transaction is running and before it's
3650          * committing (when we call filemap_fdata[write|wait]_range against
3651          * the btree inode), we might have
3652          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3653          * returns an error or an error happens during writeback, when we're
3654          * committing the transaction we wouldn't know about it, since the pages
3655          * can be no longer dirty nor marked anymore for writeback (if a
3656          * subsequent modification to the extent buffer didn't happen before the
3657          * transaction commit), which makes filemap_fdata[write|wait]_range not
3658          * able to find the pages tagged with SetPageError at transaction
3659          * commit time. So if this happens we must abort the transaction,
3660          * otherwise we commit a super block with btree roots that point to
3661          * btree nodes/leafs whose content on disk is invalid - either garbage
3662          * or the content of some node/leaf from a past generation that got
3663          * cowed or deleted and is no longer valid.
3664          *
3665          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3666          * not be enough - we need to distinguish between log tree extents vs
3667          * non-log tree extents, and the next filemap_fdatawait_range() call
3668          * will catch and clear such errors in the mapping - and that call might
3669          * be from a log sync and not from a transaction commit. Also, checking
3670          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3671          * not done and would not be reliable - the eb might have been released
3672          * from memory and reading it back again means that flag would not be
3673          * set (since it's a runtime flag, not persisted on disk).
3674          *
3675          * Using the flags below in the btree inode also makes us achieve the
3676          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3677          * writeback for all dirty pages and before filemap_fdatawait_range()
3678          * is called, the writeback for all dirty pages had already finished
3679          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3680          * filemap_fdatawait_range() would return success, as it could not know
3681          * that writeback errors happened (the pages were no longer tagged for
3682          * writeback).
3683          */
3684         switch (eb->log_index) {
3685         case -1:
3686                 set_bit(BTRFS_INODE_BTREE_ERR, &btree_ino->runtime_flags);
3687                 break;
3688         case 0:
3689                 set_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
3690                 break;
3691         case 1:
3692                 set_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
3693                 break;
3694         default:
3695                 BUG(); /* unexpected, logic error */
3696         }
3697 }
3698
3699 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3700 {
3701         struct bio_vec *bvec;
3702         struct extent_buffer *eb;
3703         int i, done;
3704
3705         bio_for_each_segment_all(bvec, bio, i) {
3706                 struct page *page = bvec->bv_page;
3707
3708                 eb = (struct extent_buffer *)page->private;
3709                 BUG_ON(!eb);
3710                 done = atomic_dec_and_test(&eb->io_pages);
3711
3712                 if (err || test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3713                         ClearPageUptodate(page);
3714                         set_btree_ioerr(page);
3715                 }
3716
3717                 end_page_writeback(page);
3718
3719                 if (!done)
3720                         continue;
3721
3722                 end_extent_buffer_writeback(eb);
3723         }
3724
3725         bio_put(bio);
3726 }
3727
3728 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3729                         struct btrfs_fs_info *fs_info,
3730                         struct writeback_control *wbc,
3731                         struct extent_page_data *epd)
3732 {
3733         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3734         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3735         u64 offset = eb->start;
3736         unsigned long i, num_pages;
3737         unsigned long bio_flags = 0;
3738         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3739         int ret = 0;
3740
3741         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3742         num_pages = num_extent_pages(eb->start, eb->len);
3743         atomic_set(&eb->io_pages, num_pages);
3744         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3745                 bio_flags = EXTENT_BIO_TREE_LOG;
3746
3747         for (i = 0; i < num_pages; i++) {
3748                 struct page *p = eb->pages[i];
3749
3750                 clear_page_dirty_for_io(p);
3751                 set_page_writeback(p);
3752                 ret = submit_extent_page(rw, tree, p, offset >> 9,
3753                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3754                                          -1, end_bio_extent_buffer_writepage,
3755                                          0, epd->bio_flags, bio_flags);
3756                 epd->bio_flags = bio_flags;
3757                 if (ret) {
3758                         set_btree_ioerr(p);
3759                         end_page_writeback(p);
3760                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3761                                 end_extent_buffer_writeback(eb);
3762                         ret = -EIO;
3763                         break;
3764                 }
3765                 offset += PAGE_CACHE_SIZE;
3766                 update_nr_written(p, wbc, 1);
3767                 unlock_page(p);
3768         }
3769
3770         if (unlikely(ret)) {
3771                 for (; i < num_pages; i++) {
3772                         struct page *p = eb->pages[i];
3773                         clear_page_dirty_for_io(p);
3774                         unlock_page(p);
3775                 }
3776         }
3777
3778         return ret;
3779 }
3780
3781 int btree_write_cache_pages(struct address_space *mapping,
3782                                    struct writeback_control *wbc)
3783 {
3784         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3785         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3786         struct extent_buffer *eb, *prev_eb = NULL;
3787         struct extent_page_data epd = {
3788                 .bio = NULL,
3789                 .tree = tree,
3790                 .extent_locked = 0,
3791                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3792                 .bio_flags = 0,
3793         };
3794         int ret = 0;
3795         int done = 0;
3796         int nr_to_write_done = 0;
3797         struct pagevec pvec;
3798         int nr_pages;
3799         pgoff_t index;
3800         pgoff_t end;            /* Inclusive */
3801         int scanned = 0;
3802         int tag;
3803
3804         pagevec_init(&pvec, 0);
3805         if (wbc->range_cyclic) {
3806                 index = mapping->writeback_index; /* Start from prev offset */
3807                 end = -1;
3808         } else {
3809                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3810                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3811                 scanned = 1;
3812         }
3813         if (wbc->sync_mode == WB_SYNC_ALL)
3814                 tag = PAGECACHE_TAG_TOWRITE;
3815         else
3816                 tag = PAGECACHE_TAG_DIRTY;
3817 retry:
3818         if (wbc->sync_mode == WB_SYNC_ALL)
3819                 tag_pages_for_writeback(mapping, index, end);
3820         while (!done && !nr_to_write_done && (index <= end) &&
3821                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3822                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3823                 unsigned i;
3824
3825                 scanned = 1;
3826                 for (i = 0; i < nr_pages; i++) {
3827                         struct page *page = pvec.pages[i];
3828
3829                         if (!PagePrivate(page))
3830                                 continue;
3831
3832                         if (!wbc->range_cyclic && page->index > end) {
3833                                 done = 1;
3834                                 break;
3835                         }
3836
3837                         spin_lock(&mapping->private_lock);
3838                         if (!PagePrivate(page)) {
3839                                 spin_unlock(&mapping->private_lock);
3840                                 continue;
3841                         }
3842
3843                         eb = (struct extent_buffer *)page->private;
3844
3845                         /*
3846                          * Shouldn't happen and normally this would be a BUG_ON
3847                          * but no sense in crashing the users box for something
3848                          * we can survive anyway.
3849                          */
3850                         if (WARN_ON(!eb)) {
3851                                 spin_unlock(&mapping->private_lock);
3852                                 continue;
3853                         }
3854
3855                         if (eb == prev_eb) {
3856                                 spin_unlock(&mapping->private_lock);
3857                                 continue;
3858                         }
3859
3860                         ret = atomic_inc_not_zero(&eb->refs);
3861                         spin_unlock(&mapping->private_lock);
3862                         if (!ret)
3863                                 continue;
3864
3865                         prev_eb = eb;
3866                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3867                         if (!ret) {
3868                                 free_extent_buffer(eb);
3869                                 continue;
3870                         }
3871
3872                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3873                         if (ret) {
3874                                 done = 1;
3875                                 free_extent_buffer(eb);
3876                                 break;
3877                         }
3878                         free_extent_buffer(eb);
3879
3880                         /*
3881                          * the filesystem may choose to bump up nr_to_write.
3882                          * We have to make sure to honor the new nr_to_write
3883                          * at any time
3884                          */
3885                         nr_to_write_done = wbc->nr_to_write <= 0;
3886                 }
3887                 pagevec_release(&pvec);
3888                 cond_resched();
3889         }
3890         if (!scanned && !done) {
3891                 /*
3892                  * We hit the last page and there is more work to be done: wrap
3893                  * back to the start of the file
3894                  */
3895                 scanned = 1;
3896                 index = 0;
3897                 goto retry;
3898         }
3899         flush_write_bio(&epd);
3900         return ret;
3901 }
3902
3903 /**
3904  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3905  * @mapping: address space structure to write
3906  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3907  * @writepage: function called for each page
3908  * @data: data passed to writepage function
3909  *
3910  * If a page is already under I/O, write_cache_pages() skips it, even
3911  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3912  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3913  * and msync() need to guarantee that all the data which was dirty at the time
3914  * the call was made get new I/O started against them.  If wbc->sync_mode is
3915  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3916  * existing IO to complete.
3917  */
3918 static int extent_write_cache_pages(struct extent_io_tree *tree,
3919                              struct address_space *mapping,
3920                              struct writeback_control *wbc,
3921                              writepage_t writepage, void *data,
3922                              void (*flush_fn)(void *))
3923 {
3924         struct inode *inode = mapping->host;
3925         int ret = 0;
3926         int done = 0;
3927         int err = 0;
3928         int nr_to_write_done = 0;
3929         struct pagevec pvec;
3930         int nr_pages;
3931         pgoff_t index;
3932         pgoff_t end;            /* Inclusive */
3933         int scanned = 0;
3934         int tag;
3935
3936         /*
3937          * We have to hold onto the inode so that ordered extents can do their
3938          * work when the IO finishes.  The alternative to this is failing to add
3939          * an ordered extent if the igrab() fails there and that is a huge pain
3940          * to deal with, so instead just hold onto the inode throughout the
3941          * writepages operation.  If it fails here we are freeing up the inode
3942          * anyway and we'd rather not waste our time writing out stuff that is
3943          * going to be truncated anyway.
3944          */
3945         if (!igrab(inode))
3946                 return 0;
3947
3948         pagevec_init(&pvec, 0);
3949         if (wbc->range_cyclic) {
3950                 index = mapping->writeback_index; /* Start from prev offset */
3951                 end = -1;
3952         } else {
3953                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3954                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3955                 scanned = 1;
3956         }
3957         if (wbc->sync_mode == WB_SYNC_ALL)
3958                 tag = PAGECACHE_TAG_TOWRITE;
3959         else
3960                 tag = PAGECACHE_TAG_DIRTY;
3961 retry:
3962         if (wbc->sync_mode == WB_SYNC_ALL)
3963                 tag_pages_for_writeback(mapping, index, end);
3964         while (!done && !nr_to_write_done && (index <= end) &&
3965                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3966                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3967                 unsigned i;
3968
3969                 scanned = 1;
3970                 for (i = 0; i < nr_pages; i++) {
3971                         struct page *page = pvec.pages[i];
3972
3973                         /*
3974                          * At this point we hold neither mapping->tree_lock nor
3975                          * lock on the page itself: the page may be truncated or
3976                          * invalidated (changing page->mapping to NULL), or even
3977                          * swizzled back from swapper_space to tmpfs file
3978                          * mapping
3979                          */
3980                         if (!trylock_page(page)) {
3981                                 flush_fn(data);
3982                                 lock_page(page);
3983                         }
3984
3985                         if (unlikely(page->mapping != mapping)) {
3986                                 unlock_page(page);
3987                                 continue;
3988                         }
3989
3990                         if (!wbc->range_cyclic && page->index > end) {
3991                                 done = 1;
3992                                 unlock_page(page);
3993                                 continue;
3994                         }
3995
3996                         if (wbc->sync_mode != WB_SYNC_NONE) {
3997                                 if (PageWriteback(page))
3998                                         flush_fn(data);
3999                                 wait_on_page_writeback(page);
4000                         }
4001
4002                         if (PageWriteback(page) ||
4003                             !clear_page_dirty_for_io(page)) {
4004                                 unlock_page(page);
4005                                 continue;
4006                         }
4007
4008                         ret = (*writepage)(page, wbc, data);
4009
4010                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4011                                 unlock_page(page);
4012                                 ret = 0;
4013                         }
4014                         if (!err && ret < 0)
4015                                 err = ret;
4016
4017                         /*
4018                          * the filesystem may choose to bump up nr_to_write.
4019                          * We have to make sure to honor the new nr_to_write
4020                          * at any time
4021                          */
4022                         nr_to_write_done = wbc->nr_to_write <= 0;
4023                 }
4024                 pagevec_release(&pvec);
4025                 cond_resched();
4026         }
4027         if (!scanned && !done && !err) {
4028                 /*
4029                  * We hit the last page and there is more work to be done: wrap
4030                  * back to the start of the file
4031                  */
4032                 scanned = 1;
4033                 index = 0;
4034                 goto retry;
4035         }
4036         btrfs_add_delayed_iput(inode);
4037         return err;
4038 }
4039
4040 static void flush_epd_write_bio(struct extent_page_data *epd)
4041 {
4042         if (epd->bio) {
4043                 int rw = WRITE;
4044                 int ret;
4045
4046                 if (epd->sync_io)
4047                         rw = WRITE_SYNC;
4048
4049                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
4050                 BUG_ON(ret < 0); /* -ENOMEM */
4051                 epd->bio = NULL;
4052         }
4053 }
4054
4055 static noinline void flush_write_bio(void *data)
4056 {
4057         struct extent_page_data *epd = data;
4058         flush_epd_write_bio(epd);
4059 }
4060
4061 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4062                           get_extent_t *get_extent,
4063                           struct writeback_control *wbc)
4064 {
4065         int ret;
4066         struct extent_page_data epd = {
4067                 .bio = NULL,
4068                 .tree = tree,
4069                 .get_extent = get_extent,
4070                 .extent_locked = 0,
4071                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4072                 .bio_flags = 0,
4073         };
4074
4075         ret = __extent_writepage(page, wbc, &epd);
4076
4077         flush_epd_write_bio(&epd);
4078         return ret;
4079 }
4080
4081 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4082                               u64 start, u64 end, get_extent_t *get_extent,
4083                               int mode)
4084 {
4085         int ret = 0;
4086         struct address_space *mapping = inode->i_mapping;
4087         struct page *page;
4088         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
4089                 PAGE_CACHE_SHIFT;
4090
4091         struct extent_page_data epd = {
4092                 .bio = NULL,
4093                 .tree = tree,
4094                 .get_extent = get_extent,
4095                 .extent_locked = 1,
4096                 .sync_io = mode == WB_SYNC_ALL,
4097                 .bio_flags = 0,
4098         };
4099         struct writeback_control wbc_writepages = {
4100                 .sync_mode      = mode,
4101                 .nr_to_write    = nr_pages * 2,
4102                 .range_start    = start,
4103                 .range_end      = end + 1,
4104         };
4105
4106         while (start <= end) {
4107                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
4108                 if (clear_page_dirty_for_io(page))
4109                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4110                 else {
4111                         if (tree->ops && tree->ops->writepage_end_io_hook)
4112                                 tree->ops->writepage_end_io_hook(page, start,
4113                                                  start + PAGE_CACHE_SIZE - 1,
4114                                                  NULL, 1);
4115                         unlock_page(page);
4116                 }
4117                 page_cache_release(page);
4118                 start += PAGE_CACHE_SIZE;
4119         }
4120
4121         flush_epd_write_bio(&epd);
4122         return ret;
4123 }
4124
4125 int extent_writepages(struct extent_io_tree *tree,
4126                       struct address_space *mapping,
4127                       get_extent_t *get_extent,
4128                       struct writeback_control *wbc)
4129 {
4130         int ret = 0;
4131         struct extent_page_data epd = {
4132                 .bio = NULL,
4133                 .tree = tree,
4134                 .get_extent = get_extent,
4135                 .extent_locked = 0,
4136                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4137                 .bio_flags = 0,
4138         };
4139
4140         ret = extent_write_cache_pages(tree, mapping, wbc,
4141                                        __extent_writepage, &epd,
4142                                        flush_write_bio);
4143         flush_epd_write_bio(&epd);
4144         return ret;
4145 }
4146
4147 int extent_readpages(struct extent_io_tree *tree,
4148                      struct address_space *mapping,
4149                      struct list_head *pages, unsigned nr_pages,
4150                      get_extent_t get_extent)
4151 {
4152         struct bio *bio = NULL;
4153         unsigned page_idx;
4154         unsigned long bio_flags = 0;
4155         struct page *pagepool[16];
4156         struct page *page;
4157         struct extent_map *em_cached = NULL;
4158         int nr = 0;
4159
4160         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4161                 page = list_entry(pages->prev, struct page, lru);
4162
4163                 prefetchw(&page->flags);
4164                 list_del(&page->lru);
4165                 if (add_to_page_cache_lru(page, mapping,
4166                                         page->index, GFP_NOFS)) {
4167                         page_cache_release(page);
4168                         continue;
4169                 }
4170
4171                 pagepool[nr++] = page;
4172                 if (nr < ARRAY_SIZE(pagepool))
4173                         continue;
4174                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4175                                    &bio, 0, &bio_flags, READ);
4176                 nr = 0;
4177         }
4178         if (nr)
4179                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4180                                    &bio, 0, &bio_flags, READ);
4181
4182         if (em_cached)
4183                 free_extent_map(em_cached);
4184
4185         BUG_ON(!list_empty(pages));
4186         if (bio)
4187                 return submit_one_bio(READ, bio, 0, bio_flags);
4188         return 0;
4189 }
4190
4191 /*
4192  * basic invalidatepage code, this waits on any locked or writeback
4193  * ranges corresponding to the page, and then deletes any extent state
4194  * records from the tree
4195  */
4196 int extent_invalidatepage(struct extent_io_tree *tree,
4197                           struct page *page, unsigned long offset)
4198 {
4199         struct extent_state *cached_state = NULL;
4200         u64 start = page_offset(page);
4201         u64 end = start + PAGE_CACHE_SIZE - 1;
4202         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4203
4204         start += ALIGN(offset, blocksize);
4205         if (start > end)
4206                 return 0;
4207
4208         lock_extent_bits(tree, start, end, 0, &cached_state);
4209         wait_on_page_writeback(page);
4210         clear_extent_bit(tree, start, end,
4211                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4212                          EXTENT_DO_ACCOUNTING,
4213                          1, 1, &cached_state, GFP_NOFS);
4214         return 0;
4215 }
4216
4217 /*
4218  * a helper for releasepage, this tests for areas of the page that
4219  * are locked or under IO and drops the related state bits if it is safe
4220  * to drop the page.
4221  */
4222 static int try_release_extent_state(struct extent_map_tree *map,
4223                                     struct extent_io_tree *tree,
4224                                     struct page *page, gfp_t mask)
4225 {
4226         u64 start = page_offset(page);
4227         u64 end = start + PAGE_CACHE_SIZE - 1;
4228         int ret = 1;
4229
4230         if (test_range_bit(tree, start, end,
4231                            EXTENT_IOBITS, 0, NULL))
4232                 ret = 0;
4233         else {
4234                 if ((mask & GFP_NOFS) == GFP_NOFS)
4235                         mask = GFP_NOFS;
4236                 /*
4237                  * at this point we can safely clear everything except the
4238                  * locked bit and the nodatasum bit
4239                  */
4240                 ret = clear_extent_bit(tree, start, end,
4241                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4242                                  0, 0, NULL, mask);
4243
4244                 /* if clear_extent_bit failed for enomem reasons,
4245                  * we can't allow the release to continue.
4246                  */
4247                 if (ret < 0)
4248                         ret = 0;
4249                 else
4250                         ret = 1;
4251         }
4252         return ret;
4253 }
4254
4255 /*
4256  * a helper for releasepage.  As long as there are no locked extents
4257  * in the range corresponding to the page, both state records and extent
4258  * map records are removed
4259  */
4260 int try_release_extent_mapping(struct extent_map_tree *map,
4261                                struct extent_io_tree *tree, struct page *page,
4262                                gfp_t mask)
4263 {
4264         struct extent_map *em;
4265         u64 start = page_offset(page);
4266         u64 end = start + PAGE_CACHE_SIZE - 1;
4267
4268         if ((mask & __GFP_WAIT) &&
4269             page->mapping->host->i_size > 16 * 1024 * 1024) {
4270                 u64 len;
4271                 while (start <= end) {
4272                         len = end - start + 1;
4273                         write_lock(&map->lock);
4274                         em = lookup_extent_mapping(map, start, len);
4275                         if (!em) {
4276                                 write_unlock(&map->lock);
4277                                 break;
4278                         }
4279                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4280                             em->start != start) {
4281                                 write_unlock(&map->lock);
4282                                 free_extent_map(em);
4283                                 break;
4284                         }
4285                         if (!test_range_bit(tree, em->start,
4286                                             extent_map_end(em) - 1,
4287                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4288                                             0, NULL)) {
4289                                 remove_extent_mapping(map, em);
4290                                 /* once for the rb tree */
4291                                 free_extent_map(em);
4292                         }
4293                         start = extent_map_end(em);
4294                         write_unlock(&map->lock);
4295
4296                         /* once for us */
4297                         free_extent_map(em);
4298                 }
4299         }
4300         return try_release_extent_state(map, tree, page, mask);
4301 }
4302
4303 /*
4304  * helper function for fiemap, which doesn't want to see any holes.
4305  * This maps until we find something past 'last'
4306  */
4307 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4308                                                 u64 offset,
4309                                                 u64 last,
4310                                                 get_extent_t *get_extent)
4311 {
4312         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4313         struct extent_map *em;
4314         u64 len;
4315
4316         if (offset >= last)
4317                 return NULL;
4318
4319         while (1) {
4320                 len = last - offset;
4321                 if (len == 0)
4322                         break;
4323                 len = ALIGN(len, sectorsize);
4324                 em = get_extent(inode, NULL, 0, offset, len, 0);
4325                 if (IS_ERR_OR_NULL(em))
4326                         return em;
4327
4328                 /* if this isn't a hole return it */
4329                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4330                     em->block_start != EXTENT_MAP_HOLE) {
4331                         return em;
4332                 }
4333
4334                 /* this is a hole, advance to the next extent */
4335                 offset = extent_map_end(em);
4336                 free_extent_map(em);
4337                 if (offset >= last)
4338                         break;
4339         }
4340         return NULL;
4341 }
4342
4343 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4344                 __u64 start, __u64 len, get_extent_t *get_extent)
4345 {
4346         int ret = 0;
4347         u64 off = start;
4348         u64 max = start + len;
4349         u32 flags = 0;
4350         u32 found_type;
4351         u64 last;
4352         u64 last_for_get_extent = 0;
4353         u64 disko = 0;
4354         u64 isize = i_size_read(inode);
4355         struct btrfs_key found_key;
4356         struct extent_map *em = NULL;
4357         struct extent_state *cached_state = NULL;
4358         struct btrfs_path *path;
4359         struct btrfs_root *root = BTRFS_I(inode)->root;
4360         int end = 0;
4361         u64 em_start = 0;
4362         u64 em_len = 0;
4363         u64 em_end = 0;
4364
4365         if (len == 0)
4366                 return -EINVAL;
4367
4368         path = btrfs_alloc_path();
4369         if (!path)
4370                 return -ENOMEM;
4371         path->leave_spinning = 1;
4372
4373         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4374         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4375
4376         /*
4377          * lookup the last file extent.  We're not using i_size here
4378          * because there might be preallocation past i_size
4379          */
4380         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
4381                                        0);
4382         if (ret < 0) {
4383                 btrfs_free_path(path);
4384                 return ret;
4385         }
4386         WARN_ON(!ret);
4387         path->slots[0]--;
4388         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4389         found_type = found_key.type;
4390
4391         /* No extents, but there might be delalloc bits */
4392         if (found_key.objectid != btrfs_ino(inode) ||
4393             found_type != BTRFS_EXTENT_DATA_KEY) {
4394                 /* have to trust i_size as the end */
4395                 last = (u64)-1;
4396                 last_for_get_extent = isize;
4397         } else {
4398                 /*
4399                  * remember the start of the last extent.  There are a
4400                  * bunch of different factors that go into the length of the
4401                  * extent, so its much less complex to remember where it started
4402                  */
4403                 last = found_key.offset;
4404                 last_for_get_extent = last + 1;
4405         }
4406         btrfs_release_path(path);
4407
4408         /*
4409          * we might have some extents allocated but more delalloc past those
4410          * extents.  so, we trust isize unless the start of the last extent is
4411          * beyond isize
4412          */
4413         if (last < isize) {
4414                 last = (u64)-1;
4415                 last_for_get_extent = isize;
4416         }
4417
4418         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4419                          &cached_state);
4420
4421         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4422                                    get_extent);
4423         if (!em)
4424                 goto out;
4425         if (IS_ERR(em)) {
4426                 ret = PTR_ERR(em);
4427                 goto out;
4428         }
4429
4430         while (!end) {
4431                 u64 offset_in_extent = 0;
4432
4433                 /* break if the extent we found is outside the range */
4434                 if (em->start >= max || extent_map_end(em) < off)
4435                         break;
4436
4437                 /*
4438                  * get_extent may return an extent that starts before our
4439                  * requested range.  We have to make sure the ranges
4440                  * we return to fiemap always move forward and don't
4441                  * overlap, so adjust the offsets here
4442                  */
4443                 em_start = max(em->start, off);
4444
4445                 /*
4446                  * record the offset from the start of the extent
4447                  * for adjusting the disk offset below.  Only do this if the
4448                  * extent isn't compressed since our in ram offset may be past
4449                  * what we have actually allocated on disk.
4450                  */
4451                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4452                         offset_in_extent = em_start - em->start;
4453                 em_end = extent_map_end(em);
4454                 em_len = em_end - em_start;
4455                 disko = 0;
4456                 flags = 0;
4457
4458                 /*
4459                  * bump off for our next call to get_extent
4460                  */
4461                 off = extent_map_end(em);
4462                 if (off >= max)
4463                         end = 1;
4464
4465                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4466                         end = 1;
4467                         flags |= FIEMAP_EXTENT_LAST;
4468                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4469                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4470                                   FIEMAP_EXTENT_NOT_ALIGNED);
4471                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4472                         flags |= (FIEMAP_EXTENT_DELALLOC |
4473                                   FIEMAP_EXTENT_UNKNOWN);
4474                 } else if (fieinfo->fi_extents_max) {
4475                         u64 bytenr = em->block_start -
4476                                 (em->start - em->orig_start);
4477
4478                         disko = em->block_start + offset_in_extent;
4479
4480                         /*
4481                          * As btrfs supports shared space, this information
4482                          * can be exported to userspace tools via
4483                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4484                          * then we're just getting a count and we can skip the
4485                          * lookup stuff.
4486                          */
4487                         ret = btrfs_check_shared(NULL, root->fs_info,
4488                                                  root->objectid,
4489                                                  btrfs_ino(inode), bytenr);
4490                         if (ret < 0)
4491                                 goto out_free;
4492                         if (ret)
4493                                 flags |= FIEMAP_EXTENT_SHARED;
4494                         ret = 0;
4495                 }
4496                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4497                         flags |= FIEMAP_EXTENT_ENCODED;
4498                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4499                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4500
4501                 free_extent_map(em);
4502                 em = NULL;
4503                 if ((em_start >= last) || em_len == (u64)-1 ||
4504                    (last == (u64)-1 && isize <= em_end)) {
4505                         flags |= FIEMAP_EXTENT_LAST;
4506                         end = 1;
4507                 }
4508
4509                 /* now scan forward to see if this is really the last extent. */
4510                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4511                                            get_extent);
4512                 if (IS_ERR(em)) {
4513                         ret = PTR_ERR(em);
4514                         goto out;
4515                 }
4516                 if (!em) {
4517                         flags |= FIEMAP_EXTENT_LAST;
4518                         end = 1;
4519                 }
4520                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4521                                               em_len, flags);
4522                 if (ret) {
4523                         if (ret == 1)
4524                                 ret = 0;
4525                         goto out_free;
4526                 }
4527         }
4528 out_free:
4529         free_extent_map(em);
4530 out:
4531         btrfs_free_path(path);
4532         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4533                              &cached_state, GFP_NOFS);
4534         return ret;
4535 }
4536
4537 static void __free_extent_buffer(struct extent_buffer *eb)
4538 {
4539         btrfs_leak_debug_del(&eb->leak_list);
4540         kmem_cache_free(extent_buffer_cache, eb);
4541 }
4542
4543 int extent_buffer_under_io(struct extent_buffer *eb)
4544 {
4545         return (atomic_read(&eb->io_pages) ||
4546                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4547                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4548 }
4549
4550 /*
4551  * Helper for releasing extent buffer page.
4552  */
4553 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4554 {
4555         unsigned long index;
4556         struct page *page;
4557         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4558
4559         BUG_ON(extent_buffer_under_io(eb));
4560
4561         index = num_extent_pages(eb->start, eb->len);
4562         if (index == 0)
4563                 return;
4564
4565         do {
4566                 index--;
4567                 page = eb->pages[index];
4568                 if (!page)
4569                         continue;
4570                 if (mapped)
4571                         spin_lock(&page->mapping->private_lock);
4572                 /*
4573                  * We do this since we'll remove the pages after we've
4574                  * removed the eb from the radix tree, so we could race
4575                  * and have this page now attached to the new eb.  So
4576                  * only clear page_private if it's still connected to
4577                  * this eb.
4578                  */
4579                 if (PagePrivate(page) &&
4580                     page->private == (unsigned long)eb) {
4581                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4582                         BUG_ON(PageDirty(page));
4583                         BUG_ON(PageWriteback(page));
4584                         /*
4585                          * We need to make sure we haven't be attached
4586                          * to a new eb.
4587                          */
4588                         ClearPagePrivate(page);
4589                         set_page_private(page, 0);
4590                         /* One for the page private */
4591                         page_cache_release(page);
4592                 }
4593
4594                 if (mapped)
4595                         spin_unlock(&page->mapping->private_lock);
4596
4597                 /* One for when we alloced the page */
4598                 page_cache_release(page);
4599         } while (index != 0);
4600 }
4601
4602 /*
4603  * Helper for releasing the extent buffer.
4604  */
4605 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4606 {
4607         btrfs_release_extent_buffer_page(eb);
4608         __free_extent_buffer(eb);
4609 }
4610
4611 static struct extent_buffer *
4612 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4613                       unsigned long len)
4614 {
4615         struct extent_buffer *eb = NULL;
4616
4617         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS);
4618         if (eb == NULL)
4619                 return NULL;
4620         eb->start = start;
4621         eb->len = len;
4622         eb->fs_info = fs_info;
4623         eb->bflags = 0;
4624         rwlock_init(&eb->lock);
4625         atomic_set(&eb->write_locks, 0);
4626         atomic_set(&eb->read_locks, 0);
4627         atomic_set(&eb->blocking_readers, 0);
4628         atomic_set(&eb->blocking_writers, 0);
4629         atomic_set(&eb->spinning_readers, 0);
4630         atomic_set(&eb->spinning_writers, 0);
4631         eb->lock_nested = 0;
4632         init_waitqueue_head(&eb->write_lock_wq);
4633         init_waitqueue_head(&eb->read_lock_wq);
4634
4635         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4636
4637         spin_lock_init(&eb->refs_lock);
4638         atomic_set(&eb->refs, 1);
4639         atomic_set(&eb->io_pages, 0);
4640
4641         /*
4642          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4643          */
4644         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4645                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4646         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4647
4648         return eb;
4649 }
4650
4651 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4652 {
4653         unsigned long i;
4654         struct page *p;
4655         struct extent_buffer *new;
4656         unsigned long num_pages = num_extent_pages(src->start, src->len);
4657
4658         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4659         if (new == NULL)
4660                 return NULL;
4661
4662         for (i = 0; i < num_pages; i++) {
4663                 p = alloc_page(GFP_NOFS);
4664                 if (!p) {
4665                         btrfs_release_extent_buffer(new);
4666                         return NULL;
4667                 }
4668                 attach_extent_buffer_page(new, p);
4669                 WARN_ON(PageDirty(p));
4670                 SetPageUptodate(p);
4671                 new->pages[i] = p;
4672         }
4673
4674         copy_extent_buffer(new, src, 0, 0, src->len);
4675         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4676         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4677
4678         return new;
4679 }
4680
4681 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4682                                                 u64 start)
4683 {
4684         struct extent_buffer *eb;
4685         unsigned long len;
4686         unsigned long num_pages;
4687         unsigned long i;
4688
4689         if (!fs_info) {
4690                 /*
4691                  * Called only from tests that don't always have a fs_info
4692                  * available, but we know that nodesize is 4096
4693                  */
4694                 len = 4096;
4695         } else {
4696                 len = fs_info->tree_root->nodesize;
4697         }
4698         num_pages = num_extent_pages(0, len);
4699
4700         eb = __alloc_extent_buffer(fs_info, start, len);
4701         if (!eb)
4702                 return NULL;
4703
4704         for (i = 0; i < num_pages; i++) {
4705                 eb->pages[i] = alloc_page(GFP_NOFS);
4706                 if (!eb->pages[i])
4707                         goto err;
4708         }
4709         set_extent_buffer_uptodate(eb);
4710         btrfs_set_header_nritems(eb, 0);
4711         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4712
4713         return eb;
4714 err:
4715         for (; i > 0; i--)
4716                 __free_page(eb->pages[i - 1]);
4717         __free_extent_buffer(eb);
4718         return NULL;
4719 }
4720
4721 static void check_buffer_tree_ref(struct extent_buffer *eb)
4722 {
4723         int refs;
4724         /* the ref bit is tricky.  We have to make sure it is set
4725          * if we have the buffer dirty.   Otherwise the
4726          * code to free a buffer can end up dropping a dirty
4727          * page
4728          *
4729          * Once the ref bit is set, it won't go away while the
4730          * buffer is dirty or in writeback, and it also won't
4731          * go away while we have the reference count on the
4732          * eb bumped.
4733          *
4734          * We can't just set the ref bit without bumping the
4735          * ref on the eb because free_extent_buffer might
4736          * see the ref bit and try to clear it.  If this happens
4737          * free_extent_buffer might end up dropping our original
4738          * ref by mistake and freeing the page before we are able
4739          * to add one more ref.
4740          *
4741          * So bump the ref count first, then set the bit.  If someone
4742          * beat us to it, drop the ref we added.
4743          */
4744         refs = atomic_read(&eb->refs);
4745         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4746                 return;
4747
4748         spin_lock(&eb->refs_lock);
4749         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4750                 atomic_inc(&eb->refs);
4751         spin_unlock(&eb->refs_lock);
4752 }
4753
4754 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4755                 struct page *accessed)
4756 {
4757         unsigned long num_pages, i;
4758
4759         check_buffer_tree_ref(eb);
4760
4761         num_pages = num_extent_pages(eb->start, eb->len);
4762         for (i = 0; i < num_pages; i++) {
4763                 struct page *p = eb->pages[i];
4764
4765                 if (p != accessed)
4766                         mark_page_accessed(p);
4767         }
4768 }
4769
4770 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4771                                          u64 start)
4772 {
4773         struct extent_buffer *eb;
4774
4775         rcu_read_lock();
4776         eb = radix_tree_lookup(&fs_info->buffer_radix,
4777                                start >> PAGE_CACHE_SHIFT);
4778         if (eb && atomic_inc_not_zero(&eb->refs)) {
4779                 rcu_read_unlock();
4780                 /*
4781                  * Lock our eb's refs_lock to avoid races with
4782                  * free_extent_buffer. When we get our eb it might be flagged
4783                  * with EXTENT_BUFFER_STALE and another task running
4784                  * free_extent_buffer might have seen that flag set,
4785                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4786                  * writeback flags not set) and it's still in the tree (flag
4787                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4788                  * of decrementing the extent buffer's reference count twice.
4789                  * So here we could race and increment the eb's reference count,
4790                  * clear its stale flag, mark it as dirty and drop our reference
4791                  * before the other task finishes executing free_extent_buffer,
4792                  * which would later result in an attempt to free an extent
4793                  * buffer that is dirty.
4794                  */
4795                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4796                         spin_lock(&eb->refs_lock);
4797                         spin_unlock(&eb->refs_lock);
4798                 }
4799                 mark_extent_buffer_accessed(eb, NULL);
4800                 return eb;
4801         }
4802         rcu_read_unlock();
4803
4804         return NULL;
4805 }
4806
4807 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4808 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4809                                                u64 start)
4810 {
4811         struct extent_buffer *eb, *exists = NULL;
4812         int ret;
4813
4814         eb = find_extent_buffer(fs_info, start);
4815         if (eb)
4816                 return eb;
4817         eb = alloc_dummy_extent_buffer(fs_info, start);
4818         if (!eb)
4819                 return NULL;
4820         eb->fs_info = fs_info;
4821 again:
4822         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4823         if (ret)
4824                 goto free_eb;
4825         spin_lock(&fs_info->buffer_lock);
4826         ret = radix_tree_insert(&fs_info->buffer_radix,
4827                                 start >> PAGE_CACHE_SHIFT, eb);
4828         spin_unlock(&fs_info->buffer_lock);
4829         radix_tree_preload_end();
4830         if (ret == -EEXIST) {
4831                 exists = find_extent_buffer(fs_info, start);
4832                 if (exists)
4833                         goto free_eb;
4834                 else
4835                         goto again;
4836         }
4837         check_buffer_tree_ref(eb);
4838         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4839
4840         /*
4841          * We will free dummy extent buffer's if they come into
4842          * free_extent_buffer with a ref count of 2, but if we are using this we
4843          * want the buffers to stay in memory until we're done with them, so
4844          * bump the ref count again.
4845          */
4846         atomic_inc(&eb->refs);
4847         return eb;
4848 free_eb:
4849         btrfs_release_extent_buffer(eb);
4850         return exists;
4851 }
4852 #endif
4853
4854 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4855                                           u64 start)
4856 {
4857         unsigned long len = fs_info->tree_root->nodesize;
4858         unsigned long num_pages = num_extent_pages(start, len);
4859         unsigned long i;
4860         unsigned long index = start >> PAGE_CACHE_SHIFT;
4861         struct extent_buffer *eb;
4862         struct extent_buffer *exists = NULL;
4863         struct page *p;
4864         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4865         int uptodate = 1;
4866         int ret;
4867
4868         eb = find_extent_buffer(fs_info, start);
4869         if (eb)
4870                 return eb;
4871
4872         eb = __alloc_extent_buffer(fs_info, start, len);
4873         if (!eb)
4874                 return NULL;
4875
4876         for (i = 0; i < num_pages; i++, index++) {
4877                 p = find_or_create_page(mapping, index, GFP_NOFS);
4878                 if (!p)
4879                         goto free_eb;
4880
4881                 spin_lock(&mapping->private_lock);
4882                 if (PagePrivate(p)) {
4883                         /*
4884                          * We could have already allocated an eb for this page
4885                          * and attached one so lets see if we can get a ref on
4886                          * the existing eb, and if we can we know it's good and
4887                          * we can just return that one, else we know we can just
4888                          * overwrite page->private.
4889                          */
4890                         exists = (struct extent_buffer *)p->private;
4891                         if (atomic_inc_not_zero(&exists->refs)) {
4892                                 spin_unlock(&mapping->private_lock);
4893                                 unlock_page(p);
4894                                 page_cache_release(p);
4895                                 mark_extent_buffer_accessed(exists, p);
4896                                 goto free_eb;
4897                         }
4898                         exists = NULL;
4899
4900                         /*
4901                          * Do this so attach doesn't complain and we need to
4902                          * drop the ref the old guy had.
4903                          */
4904                         ClearPagePrivate(p);
4905                         WARN_ON(PageDirty(p));
4906                         page_cache_release(p);
4907                 }
4908                 attach_extent_buffer_page(eb, p);
4909                 spin_unlock(&mapping->private_lock);
4910                 WARN_ON(PageDirty(p));
4911                 eb->pages[i] = p;
4912                 if (!PageUptodate(p))
4913                         uptodate = 0;
4914
4915                 /*
4916                  * see below about how we avoid a nasty race with release page
4917                  * and why we unlock later
4918                  */
4919         }
4920         if (uptodate)
4921                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4922 again:
4923         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4924         if (ret)
4925                 goto free_eb;
4926
4927         spin_lock(&fs_info->buffer_lock);
4928         ret = radix_tree_insert(&fs_info->buffer_radix,
4929                                 start >> PAGE_CACHE_SHIFT, eb);
4930         spin_unlock(&fs_info->buffer_lock);
4931         radix_tree_preload_end();
4932         if (ret == -EEXIST) {
4933                 exists = find_extent_buffer(fs_info, start);
4934                 if (exists)
4935                         goto free_eb;
4936                 else
4937                         goto again;
4938         }
4939         /* add one reference for the tree */
4940         check_buffer_tree_ref(eb);
4941         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4942
4943         /*
4944          * there is a race where release page may have
4945          * tried to find this extent buffer in the radix
4946          * but failed.  It will tell the VM it is safe to
4947          * reclaim the, and it will clear the page private bit.
4948          * We must make sure to set the page private bit properly
4949          * after the extent buffer is in the radix tree so
4950          * it doesn't get lost
4951          */
4952         SetPageChecked(eb->pages[0]);
4953         for (i = 1; i < num_pages; i++) {
4954                 p = eb->pages[i];
4955                 ClearPageChecked(p);
4956                 unlock_page(p);
4957         }
4958         unlock_page(eb->pages[0]);
4959         return eb;
4960
4961 free_eb:
4962         WARN_ON(!atomic_dec_and_test(&eb->refs));
4963         for (i = 0; i < num_pages; i++) {
4964                 if (eb->pages[i])
4965                         unlock_page(eb->pages[i]);
4966         }
4967
4968         btrfs_release_extent_buffer(eb);
4969         return exists;
4970 }
4971
4972 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4973 {
4974         struct extent_buffer *eb =
4975                         container_of(head, struct extent_buffer, rcu_head);
4976
4977         __free_extent_buffer(eb);
4978 }
4979
4980 /* Expects to have eb->eb_lock already held */
4981 static int release_extent_buffer(struct extent_buffer *eb)
4982 {
4983         WARN_ON(atomic_read(&eb->refs) == 0);
4984         if (atomic_dec_and_test(&eb->refs)) {
4985                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4986                         struct btrfs_fs_info *fs_info = eb->fs_info;
4987
4988                         spin_unlock(&eb->refs_lock);
4989
4990                         spin_lock(&fs_info->buffer_lock);
4991                         radix_tree_delete(&fs_info->buffer_radix,
4992                                           eb->start >> PAGE_CACHE_SHIFT);
4993                         spin_unlock(&fs_info->buffer_lock);
4994                 } else {
4995                         spin_unlock(&eb->refs_lock);
4996                 }
4997
4998                 /* Should be safe to release our pages at this point */
4999                 btrfs_release_extent_buffer_page(eb);
5000 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5001                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5002                         __free_extent_buffer(eb);
5003                         return 1;
5004                 }
5005 #endif
5006                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5007                 return 1;
5008         }
5009         spin_unlock(&eb->refs_lock);
5010
5011         return 0;
5012 }
5013
5014 void free_extent_buffer(struct extent_buffer *eb)
5015 {
5016         int refs;
5017         int old;
5018         if (!eb)
5019                 return;
5020
5021         while (1) {
5022                 refs = atomic_read(&eb->refs);
5023                 if (refs <= 3)
5024                         break;
5025                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5026                 if (old == refs)
5027                         return;
5028         }
5029
5030         spin_lock(&eb->refs_lock);
5031         if (atomic_read(&eb->refs) == 2 &&
5032             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5033                 atomic_dec(&eb->refs);
5034
5035         if (atomic_read(&eb->refs) == 2 &&
5036             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5037             !extent_buffer_under_io(eb) &&
5038             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5039                 atomic_dec(&eb->refs);
5040
5041         /*
5042          * I know this is terrible, but it's temporary until we stop tracking
5043          * the uptodate bits and such for the extent buffers.
5044          */
5045         release_extent_buffer(eb);
5046 }
5047
5048 void free_extent_buffer_stale(struct extent_buffer *eb)
5049 {
5050         if (!eb)
5051                 return;
5052
5053         spin_lock(&eb->refs_lock);
5054         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5055
5056         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5057             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5058                 atomic_dec(&eb->refs);
5059         release_extent_buffer(eb);
5060 }
5061
5062 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5063 {
5064         unsigned long i;
5065         unsigned long num_pages;
5066         struct page *page;
5067
5068         num_pages = num_extent_pages(eb->start, eb->len);
5069
5070         for (i = 0; i < num_pages; i++) {
5071                 page = eb->pages[i];
5072                 if (!PageDirty(page))
5073                         continue;
5074
5075                 lock_page(page);
5076                 WARN_ON(!PagePrivate(page));
5077
5078                 clear_page_dirty_for_io(page);
5079                 spin_lock_irq(&page->mapping->tree_lock);
5080                 if (!PageDirty(page)) {
5081                         radix_tree_tag_clear(&page->mapping->page_tree,
5082                                                 page_index(page),
5083                                                 PAGECACHE_TAG_DIRTY);
5084                 }
5085                 spin_unlock_irq(&page->mapping->tree_lock);
5086                 ClearPageError(page);
5087                 unlock_page(page);
5088         }
5089         WARN_ON(atomic_read(&eb->refs) == 0);
5090 }
5091
5092 int set_extent_buffer_dirty(struct extent_buffer *eb)
5093 {
5094         unsigned long i;
5095         unsigned long num_pages;
5096         int was_dirty = 0;
5097
5098         check_buffer_tree_ref(eb);
5099
5100         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5101
5102         num_pages = num_extent_pages(eb->start, eb->len);
5103         WARN_ON(atomic_read(&eb->refs) == 0);
5104         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5105
5106         for (i = 0; i < num_pages; i++)
5107                 set_page_dirty(eb->pages[i]);
5108         return was_dirty;
5109 }
5110
5111 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
5112 {
5113         unsigned long i;
5114         struct page *page;
5115         unsigned long num_pages;
5116
5117         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5118         num_pages = num_extent_pages(eb->start, eb->len);
5119         for (i = 0; i < num_pages; i++) {
5120                 page = eb->pages[i];
5121                 if (page)
5122                         ClearPageUptodate(page);
5123         }
5124         return 0;
5125 }
5126
5127 int set_extent_buffer_uptodate(struct extent_buffer *eb)
5128 {
5129         unsigned long i;
5130         struct page *page;
5131         unsigned long num_pages;
5132
5133         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5134         num_pages = num_extent_pages(eb->start, eb->len);
5135         for (i = 0; i < num_pages; i++) {
5136                 page = eb->pages[i];
5137                 SetPageUptodate(page);
5138         }
5139         return 0;
5140 }
5141
5142 int extent_buffer_uptodate(struct extent_buffer *eb)
5143 {
5144         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5145 }
5146
5147 int read_extent_buffer_pages(struct extent_io_tree *tree,
5148                              struct extent_buffer *eb, u64 start, int wait,
5149                              get_extent_t *get_extent, int mirror_num)
5150 {
5151         unsigned long i;
5152         unsigned long start_i;
5153         struct page *page;
5154         int err;
5155         int ret = 0;
5156         int locked_pages = 0;
5157         int all_uptodate = 1;
5158         unsigned long num_pages;
5159         unsigned long num_reads = 0;
5160         struct bio *bio = NULL;
5161         unsigned long bio_flags = 0;
5162
5163         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5164                 return 0;
5165
5166         if (start) {
5167                 WARN_ON(start < eb->start);
5168                 start_i = (start >> PAGE_CACHE_SHIFT) -
5169                         (eb->start >> PAGE_CACHE_SHIFT);
5170         } else {
5171                 start_i = 0;
5172         }
5173
5174         num_pages = num_extent_pages(eb->start, eb->len);
5175         for (i = start_i; i < num_pages; i++) {
5176                 page = eb->pages[i];
5177                 if (wait == WAIT_NONE) {
5178                         if (!trylock_page(page))
5179                                 goto unlock_exit;
5180                 } else {
5181                         lock_page(page);
5182                 }
5183                 locked_pages++;
5184                 if (!PageUptodate(page)) {
5185                         num_reads++;
5186                         all_uptodate = 0;
5187                 }
5188         }
5189         if (all_uptodate) {
5190                 if (start_i == 0)
5191                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5192                 goto unlock_exit;
5193         }
5194
5195         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5196         eb->read_mirror = 0;
5197         atomic_set(&eb->io_pages, num_reads);
5198         for (i = start_i; i < num_pages; i++) {
5199                 page = eb->pages[i];
5200                 if (!PageUptodate(page)) {
5201                         ClearPageError(page);
5202                         err = __extent_read_full_page(tree, page,
5203                                                       get_extent, &bio,
5204                                                       mirror_num, &bio_flags,
5205                                                       READ | REQ_META);
5206                         if (err)
5207                                 ret = err;
5208                 } else {
5209                         unlock_page(page);
5210                 }
5211         }
5212
5213         if (bio) {
5214                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
5215                                      bio_flags);
5216                 if (err)
5217                         return err;
5218         }
5219
5220         if (ret || wait != WAIT_COMPLETE)
5221                 return ret;
5222
5223         for (i = start_i; i < num_pages; i++) {
5224                 page = eb->pages[i];
5225                 wait_on_page_locked(page);
5226                 if (!PageUptodate(page))
5227                         ret = -EIO;
5228         }
5229
5230         return ret;
5231
5232 unlock_exit:
5233         i = start_i;
5234         while (locked_pages > 0) {
5235                 page = eb->pages[i];
5236                 i++;
5237                 unlock_page(page);
5238                 locked_pages--;
5239         }
5240         return ret;
5241 }
5242
5243 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5244                         unsigned long start,
5245                         unsigned long len)
5246 {
5247         size_t cur;
5248         size_t offset;
5249         struct page *page;
5250         char *kaddr;
5251         char *dst = (char *)dstv;
5252         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5253         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5254
5255         WARN_ON(start > eb->len);
5256         WARN_ON(start + len > eb->start + eb->len);
5257
5258         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5259
5260         while (len > 0) {
5261                 page = eb->pages[i];
5262
5263                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5264                 kaddr = page_address(page);
5265                 memcpy(dst, kaddr + offset, cur);
5266
5267                 dst += cur;
5268                 len -= cur;
5269                 offset = 0;
5270                 i++;
5271         }
5272 }
5273
5274 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5275                         unsigned long start,
5276                         unsigned long len)
5277 {
5278         size_t cur;
5279         size_t offset;
5280         struct page *page;
5281         char *kaddr;
5282         char __user *dst = (char __user *)dstv;
5283         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5284         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5285         int ret = 0;
5286
5287         WARN_ON(start > eb->len);
5288         WARN_ON(start + len > eb->start + eb->len);
5289
5290         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5291
5292         while (len > 0) {
5293                 page = eb->pages[i];
5294
5295                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5296                 kaddr = page_address(page);
5297                 if (copy_to_user(dst, kaddr + offset, cur)) {
5298                         ret = -EFAULT;
5299                         break;
5300                 }
5301
5302                 dst += cur;
5303                 len -= cur;
5304                 offset = 0;
5305                 i++;
5306         }
5307
5308         return ret;
5309 }
5310
5311 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5312                                unsigned long min_len, char **map,
5313                                unsigned long *map_start,
5314                                unsigned long *map_len)
5315 {
5316         size_t offset = start & (PAGE_CACHE_SIZE - 1);
5317         char *kaddr;
5318         struct page *p;
5319         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5320         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5321         unsigned long end_i = (start_offset + start + min_len - 1) >>
5322                 PAGE_CACHE_SHIFT;
5323
5324         if (i != end_i)
5325                 return -EINVAL;
5326
5327         if (i == 0) {
5328                 offset = start_offset;
5329                 *map_start = 0;
5330         } else {
5331                 offset = 0;
5332                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
5333         }
5334
5335         if (start + min_len > eb->len) {
5336                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5337                        "wanted %lu %lu\n",
5338                        eb->start, eb->len, start, min_len);
5339                 return -EINVAL;
5340         }
5341
5342         p = eb->pages[i];
5343         kaddr = page_address(p);
5344         *map = kaddr + offset;
5345         *map_len = PAGE_CACHE_SIZE - offset;
5346         return 0;
5347 }
5348
5349 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5350                           unsigned long start,
5351                           unsigned long len)
5352 {
5353         size_t cur;
5354         size_t offset;
5355         struct page *page;
5356         char *kaddr;
5357         char *ptr = (char *)ptrv;
5358         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5359         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5360         int ret = 0;
5361
5362         WARN_ON(start > eb->len);
5363         WARN_ON(start + len > eb->start + eb->len);
5364
5365         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5366
5367         while (len > 0) {
5368                 page = eb->pages[i];
5369
5370                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5371
5372                 kaddr = page_address(page);
5373                 ret = memcmp(ptr, kaddr + offset, cur);
5374                 if (ret)
5375                         break;
5376
5377                 ptr += cur;
5378                 len -= cur;
5379                 offset = 0;
5380                 i++;
5381         }
5382         return ret;
5383 }
5384
5385 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5386                          unsigned long start, unsigned long len)
5387 {
5388         size_t cur;
5389         size_t offset;
5390         struct page *page;
5391         char *kaddr;
5392         char *src = (char *)srcv;
5393         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5394         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5395
5396         WARN_ON(start > eb->len);
5397         WARN_ON(start + len > eb->start + eb->len);
5398
5399         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5400
5401         while (len > 0) {
5402                 page = eb->pages[i];
5403                 WARN_ON(!PageUptodate(page));
5404
5405                 cur = min(len, PAGE_CACHE_SIZE - offset);
5406                 kaddr = page_address(page);
5407                 memcpy(kaddr + offset, src, cur);
5408
5409                 src += cur;
5410                 len -= cur;
5411                 offset = 0;
5412                 i++;
5413         }
5414 }
5415
5416 void memset_extent_buffer(struct extent_buffer *eb, char c,
5417                           unsigned long start, unsigned long len)
5418 {
5419         size_t cur;
5420         size_t offset;
5421         struct page *page;
5422         char *kaddr;
5423         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5424         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5425
5426         WARN_ON(start > eb->len);
5427         WARN_ON(start + len > eb->start + eb->len);
5428
5429         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5430
5431         while (len > 0) {
5432                 page = eb->pages[i];
5433                 WARN_ON(!PageUptodate(page));
5434
5435                 cur = min(len, PAGE_CACHE_SIZE - offset);
5436                 kaddr = page_address(page);
5437                 memset(kaddr + offset, c, cur);
5438
5439                 len -= cur;
5440                 offset = 0;
5441                 i++;
5442         }
5443 }
5444
5445 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5446                         unsigned long dst_offset, unsigned long src_offset,
5447                         unsigned long len)
5448 {
5449         u64 dst_len = dst->len;
5450         size_t cur;
5451         size_t offset;
5452         struct page *page;
5453         char *kaddr;
5454         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5455         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5456
5457         WARN_ON(src->len != dst_len);
5458
5459         offset = (start_offset + dst_offset) &
5460                 (PAGE_CACHE_SIZE - 1);
5461
5462         while (len > 0) {
5463                 page = dst->pages[i];
5464                 WARN_ON(!PageUptodate(page));
5465
5466                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5467
5468                 kaddr = page_address(page);
5469                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5470
5471                 src_offset += cur;
5472                 len -= cur;
5473                 offset = 0;
5474                 i++;
5475         }
5476 }
5477
5478 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5479 {
5480         unsigned long distance = (src > dst) ? src - dst : dst - src;
5481         return distance < len;
5482 }
5483
5484 static void copy_pages(struct page *dst_page, struct page *src_page,
5485                        unsigned long dst_off, unsigned long src_off,
5486                        unsigned long len)
5487 {
5488         char *dst_kaddr = page_address(dst_page);
5489         char *src_kaddr;
5490         int must_memmove = 0;
5491
5492         if (dst_page != src_page) {
5493                 src_kaddr = page_address(src_page);
5494         } else {
5495                 src_kaddr = dst_kaddr;
5496                 if (areas_overlap(src_off, dst_off, len))
5497                         must_memmove = 1;
5498         }
5499
5500         if (must_memmove)
5501                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5502         else
5503                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5504 }
5505
5506 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5507                            unsigned long src_offset, unsigned long len)
5508 {
5509         size_t cur;
5510         size_t dst_off_in_page;
5511         size_t src_off_in_page;
5512         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5513         unsigned long dst_i;
5514         unsigned long src_i;
5515
5516         if (src_offset + len > dst->len) {
5517                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5518                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5519                 BUG_ON(1);
5520         }
5521         if (dst_offset + len > dst->len) {
5522                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5523                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5524                 BUG_ON(1);
5525         }
5526
5527         while (len > 0) {
5528                 dst_off_in_page = (start_offset + dst_offset) &
5529                         (PAGE_CACHE_SIZE - 1);
5530                 src_off_in_page = (start_offset + src_offset) &
5531                         (PAGE_CACHE_SIZE - 1);
5532
5533                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5534                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5535
5536                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5537                                                src_off_in_page));
5538                 cur = min_t(unsigned long, cur,
5539                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5540
5541                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5542                            dst_off_in_page, src_off_in_page, cur);
5543
5544                 src_offset += cur;
5545                 dst_offset += cur;
5546                 len -= cur;
5547         }
5548 }
5549
5550 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5551                            unsigned long src_offset, unsigned long len)
5552 {
5553         size_t cur;
5554         size_t dst_off_in_page;
5555         size_t src_off_in_page;
5556         unsigned long dst_end = dst_offset + len - 1;
5557         unsigned long src_end = src_offset + len - 1;
5558         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5559         unsigned long dst_i;
5560         unsigned long src_i;
5561
5562         if (src_offset + len > dst->len) {
5563                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5564                        "len %lu len %lu\n", src_offset, len, dst->len);
5565                 BUG_ON(1);
5566         }
5567         if (dst_offset + len > dst->len) {
5568                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5569                        "len %lu len %lu\n", dst_offset, len, dst->len);
5570                 BUG_ON(1);
5571         }
5572         if (dst_offset < src_offset) {
5573                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5574                 return;
5575         }
5576         while (len > 0) {
5577                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5578                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5579
5580                 dst_off_in_page = (start_offset + dst_end) &
5581                         (PAGE_CACHE_SIZE - 1);
5582                 src_off_in_page = (start_offset + src_end) &
5583                         (PAGE_CACHE_SIZE - 1);
5584
5585                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5586                 cur = min(cur, dst_off_in_page + 1);
5587                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5588                            dst_off_in_page - cur + 1,
5589                            src_off_in_page - cur + 1, cur);
5590
5591                 dst_end -= cur;
5592                 src_end -= cur;
5593                 len -= cur;
5594         }
5595 }
5596
5597 int try_release_extent_buffer(struct page *page)
5598 {
5599         struct extent_buffer *eb;
5600
5601         /*
5602          * We need to make sure noboody is attaching this page to an eb right
5603          * now.
5604          */
5605         spin_lock(&page->mapping->private_lock);
5606         if (!PagePrivate(page)) {
5607                 spin_unlock(&page->mapping->private_lock);
5608                 return 1;
5609         }
5610
5611         eb = (struct extent_buffer *)page->private;
5612         BUG_ON(!eb);
5613
5614         /*
5615          * This is a little awful but should be ok, we need to make sure that
5616          * the eb doesn't disappear out from under us while we're looking at
5617          * this page.
5618          */
5619         spin_lock(&eb->refs_lock);
5620         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5621                 spin_unlock(&eb->refs_lock);
5622                 spin_unlock(&page->mapping->private_lock);
5623                 return 0;
5624         }
5625         spin_unlock(&page->mapping->private_lock);
5626
5627         /*
5628          * If tree ref isn't set then we know the ref on this eb is a real ref,
5629          * so just return, this page will likely be freed soon anyway.
5630          */
5631         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5632                 spin_unlock(&eb->refs_lock);
5633                 return 0;
5634         }
5635
5636         return release_extent_buffer(eb);
5637 }