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         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1281 }
1282
1283 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1284                         struct extent_state **cached_state, gfp_t mask)
1285 {
1286         return set_extent_bit(tree, start, end,
1287                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1288                               NULL, cached_state, mask);
1289 }
1290
1291 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1292                       struct extent_state **cached_state, gfp_t mask)
1293 {
1294         return set_extent_bit(tree, start, end,
1295                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1296                               NULL, cached_state, mask);
1297 }
1298
1299 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1300                        gfp_t mask)
1301 {
1302         return clear_extent_bit(tree, start, end,
1303                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1304                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1305 }
1306
1307 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1308                      gfp_t mask)
1309 {
1310         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1311                               NULL, mask);
1312 }
1313
1314 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1315                         struct extent_state **cached_state, gfp_t mask)
1316 {
1317         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1318                               cached_state, mask);
1319 }
1320
1321 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1322                           struct extent_state **cached_state, gfp_t mask)
1323 {
1324         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1325                                 cached_state, mask);
1326 }
1327
1328 /*
1329  * either insert or lock state struct between start and end use mask to tell
1330  * us if waiting is desired.
1331  */
1332 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1333                      unsigned bits, struct extent_state **cached_state)
1334 {
1335         int err;
1336         u64 failed_start;
1337
1338         while (1) {
1339                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1340                                        EXTENT_LOCKED, &failed_start,
1341                                        cached_state, GFP_NOFS);
1342                 if (err == -EEXIST) {
1343                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1344                         start = failed_start;
1345                 } else
1346                         break;
1347                 WARN_ON(start > end);
1348         }
1349         return err;
1350 }
1351
1352 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1353 {
1354         return lock_extent_bits(tree, start, end, 0, NULL);
1355 }
1356
1357 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1358 {
1359         int err;
1360         u64 failed_start;
1361
1362         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1363                                &failed_start, NULL, GFP_NOFS);
1364         if (err == -EEXIST) {
1365                 if (failed_start > start)
1366                         clear_extent_bit(tree, start, failed_start - 1,
1367                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1368                 return 0;
1369         }
1370         return 1;
1371 }
1372
1373 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1374                          struct extent_state **cached, gfp_t mask)
1375 {
1376         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1377                                 mask);
1378 }
1379
1380 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1381 {
1382         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1383                                 GFP_NOFS);
1384 }
1385
1386 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1387 {
1388         unsigned long index = start >> PAGE_CACHE_SHIFT;
1389         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1390         struct page *page;
1391
1392         while (index <= end_index) {
1393                 page = find_get_page(inode->i_mapping, index);
1394                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1395                 clear_page_dirty_for_io(page);
1396                 page_cache_release(page);
1397                 index++;
1398         }
1399         return 0;
1400 }
1401
1402 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1403 {
1404         unsigned long index = start >> PAGE_CACHE_SHIFT;
1405         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1406         struct page *page;
1407
1408         while (index <= end_index) {
1409                 page = find_get_page(inode->i_mapping, index);
1410                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1411                 __set_page_dirty_nobuffers(page);
1412                 account_page_redirty(page);
1413                 page_cache_release(page);
1414                 index++;
1415         }
1416         return 0;
1417 }
1418
1419 /*
1420  * helper function to set both pages and extents in the tree writeback
1421  */
1422 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1423 {
1424         unsigned long index = start >> PAGE_CACHE_SHIFT;
1425         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1426         struct page *page;
1427
1428         while (index <= end_index) {
1429                 page = find_get_page(tree->mapping, index);
1430                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1431                 set_page_writeback(page);
1432                 page_cache_release(page);
1433                 index++;
1434         }
1435         return 0;
1436 }
1437
1438 /* find the first state struct with 'bits' set after 'start', and
1439  * return it.  tree->lock must be held.  NULL will returned if
1440  * nothing was found after 'start'
1441  */
1442 static struct extent_state *
1443 find_first_extent_bit_state(struct extent_io_tree *tree,
1444                             u64 start, unsigned bits)
1445 {
1446         struct rb_node *node;
1447         struct extent_state *state;
1448
1449         /*
1450          * this search will find all the extents that end after
1451          * our range starts.
1452          */
1453         node = tree_search(tree, start);
1454         if (!node)
1455                 goto out;
1456
1457         while (1) {
1458                 state = rb_entry(node, struct extent_state, rb_node);
1459                 if (state->end >= start && (state->state & bits))
1460                         return state;
1461
1462                 node = rb_next(node);
1463                 if (!node)
1464                         break;
1465         }
1466 out:
1467         return NULL;
1468 }
1469
1470 /*
1471  * find the first offset in the io tree with 'bits' set. zero is
1472  * returned if we find something, and *start_ret and *end_ret are
1473  * set to reflect the state struct that was found.
1474  *
1475  * If nothing was found, 1 is returned. If found something, return 0.
1476  */
1477 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1478                           u64 *start_ret, u64 *end_ret, unsigned bits,
1479                           struct extent_state **cached_state)
1480 {
1481         struct extent_state *state;
1482         struct rb_node *n;
1483         int ret = 1;
1484
1485         spin_lock(&tree->lock);
1486         if (cached_state && *cached_state) {
1487                 state = *cached_state;
1488                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1489                         n = rb_next(&state->rb_node);
1490                         while (n) {
1491                                 state = rb_entry(n, struct extent_state,
1492                                                  rb_node);
1493                                 if (state->state & bits)
1494                                         goto got_it;
1495                                 n = rb_next(n);
1496                         }
1497                         free_extent_state(*cached_state);
1498                         *cached_state = NULL;
1499                         goto out;
1500                 }
1501                 free_extent_state(*cached_state);
1502                 *cached_state = NULL;
1503         }
1504
1505         state = find_first_extent_bit_state(tree, start, bits);
1506 got_it:
1507         if (state) {
1508                 cache_state_if_flags(state, cached_state, 0);
1509                 *start_ret = state->start;
1510                 *end_ret = state->end;
1511                 ret = 0;
1512         }
1513 out:
1514         spin_unlock(&tree->lock);
1515         return ret;
1516 }
1517
1518 /*
1519  * find a contiguous range of bytes in the file marked as delalloc, not
1520  * more than 'max_bytes'.  start and end are used to return the range,
1521  *
1522  * 1 is returned if we find something, 0 if nothing was in the tree
1523  */
1524 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1525                                         u64 *start, u64 *end, u64 max_bytes,
1526                                         struct extent_state **cached_state)
1527 {
1528         struct rb_node *node;
1529         struct extent_state *state;
1530         u64 cur_start = *start;
1531         u64 found = 0;
1532         u64 total_bytes = 0;
1533
1534         spin_lock(&tree->lock);
1535
1536         /*
1537          * this search will find all the extents that end after
1538          * our range starts.
1539          */
1540         node = tree_search(tree, cur_start);
1541         if (!node) {
1542                 if (!found)
1543                         *end = (u64)-1;
1544                 goto out;
1545         }
1546
1547         while (1) {
1548                 state = rb_entry(node, struct extent_state, rb_node);
1549                 if (found && (state->start != cur_start ||
1550                               (state->state & EXTENT_BOUNDARY))) {
1551                         goto out;
1552                 }
1553                 if (!(state->state & EXTENT_DELALLOC)) {
1554                         if (!found)
1555                                 *end = state->end;
1556                         goto out;
1557                 }
1558                 if (!found) {
1559                         *start = state->start;
1560                         *cached_state = state;
1561                         atomic_inc(&state->refs);
1562                 }
1563                 found++;
1564                 *end = state->end;
1565                 cur_start = state->end + 1;
1566                 node = rb_next(node);
1567                 total_bytes += state->end - state->start + 1;
1568                 if (total_bytes >= max_bytes)
1569                         break;
1570                 if (!node)
1571                         break;
1572         }
1573 out:
1574         spin_unlock(&tree->lock);
1575         return found;
1576 }
1577
1578 static noinline void __unlock_for_delalloc(struct inode *inode,
1579                                            struct page *locked_page,
1580                                            u64 start, u64 end)
1581 {
1582         int ret;
1583         struct page *pages[16];
1584         unsigned long index = start >> PAGE_CACHE_SHIFT;
1585         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1586         unsigned long nr_pages = end_index - index + 1;
1587         int i;
1588
1589         if (index == locked_page->index && end_index == index)
1590                 return;
1591
1592         while (nr_pages > 0) {
1593                 ret = find_get_pages_contig(inode->i_mapping, index,
1594                                      min_t(unsigned long, nr_pages,
1595                                      ARRAY_SIZE(pages)), pages);
1596                 for (i = 0; i < ret; i++) {
1597                         if (pages[i] != locked_page)
1598                                 unlock_page(pages[i]);
1599                         page_cache_release(pages[i]);
1600                 }
1601                 nr_pages -= ret;
1602                 index += ret;
1603                 cond_resched();
1604         }
1605 }
1606
1607 static noinline int lock_delalloc_pages(struct inode *inode,
1608                                         struct page *locked_page,
1609                                         u64 delalloc_start,
1610                                         u64 delalloc_end)
1611 {
1612         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1613         unsigned long start_index = index;
1614         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1615         unsigned long pages_locked = 0;
1616         struct page *pages[16];
1617         unsigned long nrpages;
1618         int ret;
1619         int i;
1620
1621         /* the caller is responsible for locking the start index */
1622         if (index == locked_page->index && index == end_index)
1623                 return 0;
1624
1625         /* skip the page at the start index */
1626         nrpages = end_index - index + 1;
1627         while (nrpages > 0) {
1628                 ret = find_get_pages_contig(inode->i_mapping, index,
1629                                      min_t(unsigned long,
1630                                      nrpages, ARRAY_SIZE(pages)), pages);
1631                 if (ret == 0) {
1632                         ret = -EAGAIN;
1633                         goto done;
1634                 }
1635                 /* now we have an array of pages, lock them all */
1636                 for (i = 0; i < ret; i++) {
1637                         /*
1638                          * the caller is taking responsibility for
1639                          * locked_page
1640                          */
1641                         if (pages[i] != locked_page) {
1642                                 lock_page(pages[i]);
1643                                 if (!PageDirty(pages[i]) ||
1644                                     pages[i]->mapping != inode->i_mapping) {
1645                                         ret = -EAGAIN;
1646                                         unlock_page(pages[i]);
1647                                         page_cache_release(pages[i]);
1648                                         goto done;
1649                                 }
1650                         }
1651                         page_cache_release(pages[i]);
1652                         pages_locked++;
1653                 }
1654                 nrpages -= ret;
1655                 index += ret;
1656                 cond_resched();
1657         }
1658         ret = 0;
1659 done:
1660         if (ret && pages_locked) {
1661                 __unlock_for_delalloc(inode, locked_page,
1662                               delalloc_start,
1663                               ((u64)(start_index + pages_locked - 1)) <<
1664                               PAGE_CACHE_SHIFT);
1665         }
1666         return ret;
1667 }
1668
1669 /*
1670  * find a contiguous range of bytes in the file marked as delalloc, not
1671  * more than 'max_bytes'.  start and end are used to return the range,
1672  *
1673  * 1 is returned if we find something, 0 if nothing was in the tree
1674  */
1675 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1676                                     struct extent_io_tree *tree,
1677                                     struct page *locked_page, u64 *start,
1678                                     u64 *end, u64 max_bytes)
1679 {
1680         u64 delalloc_start;
1681         u64 delalloc_end;
1682         u64 found;
1683         struct extent_state *cached_state = NULL;
1684         int ret;
1685         int loops = 0;
1686
1687 again:
1688         /* step one, find a bunch of delalloc bytes starting at start */
1689         delalloc_start = *start;
1690         delalloc_end = 0;
1691         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1692                                     max_bytes, &cached_state);
1693         if (!found || delalloc_end <= *start) {
1694                 *start = delalloc_start;
1695                 *end = delalloc_end;
1696                 free_extent_state(cached_state);
1697                 return 0;
1698         }
1699
1700         /*
1701          * start comes from the offset of locked_page.  We have to lock
1702          * pages in order, so we can't process delalloc bytes before
1703          * locked_page
1704          */
1705         if (delalloc_start < *start)
1706                 delalloc_start = *start;
1707
1708         /*
1709          * make sure to limit the number of pages we try to lock down
1710          */
1711         if (delalloc_end + 1 - delalloc_start > max_bytes)
1712                 delalloc_end = delalloc_start + max_bytes - 1;
1713
1714         /* step two, lock all the pages after the page that has start */
1715         ret = lock_delalloc_pages(inode, locked_page,
1716                                   delalloc_start, delalloc_end);
1717         if (ret == -EAGAIN) {
1718                 /* some of the pages are gone, lets avoid looping by
1719                  * shortening the size of the delalloc range we're searching
1720                  */
1721                 free_extent_state(cached_state);
1722                 cached_state = NULL;
1723                 if (!loops) {
1724                         max_bytes = PAGE_CACHE_SIZE;
1725                         loops = 1;
1726                         goto again;
1727                 } else {
1728                         found = 0;
1729                         goto out_failed;
1730                 }
1731         }
1732         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1733
1734         /* step three, lock the state bits for the whole range */
1735         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1736
1737         /* then test to make sure it is all still delalloc */
1738         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1739                              EXTENT_DELALLOC, 1, cached_state);
1740         if (!ret) {
1741                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1742                                      &cached_state, GFP_NOFS);
1743                 __unlock_for_delalloc(inode, locked_page,
1744                               delalloc_start, delalloc_end);
1745                 cond_resched();
1746                 goto again;
1747         }
1748         free_extent_state(cached_state);
1749         *start = delalloc_start;
1750         *end = delalloc_end;
1751 out_failed:
1752         return found;
1753 }
1754
1755 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1756                                  struct page *locked_page,
1757                                  unsigned clear_bits,
1758                                  unsigned long page_ops)
1759 {
1760         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1761         int ret;
1762         struct page *pages[16];
1763         unsigned long index = start >> PAGE_CACHE_SHIFT;
1764         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1765         unsigned long nr_pages = end_index - index + 1;
1766         int i;
1767
1768         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1769         if (page_ops == 0)
1770                 return 0;
1771
1772         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1773                 mapping_set_error(inode->i_mapping, -EIO);
1774
1775         while (nr_pages > 0) {
1776                 ret = find_get_pages_contig(inode->i_mapping, index,
1777                                      min_t(unsigned long,
1778                                      nr_pages, ARRAY_SIZE(pages)), pages);
1779                 for (i = 0; i < ret; i++) {
1780
1781                         if (page_ops & PAGE_SET_PRIVATE2)
1782                                 SetPagePrivate2(pages[i]);
1783
1784                         if (pages[i] == locked_page) {
1785                                 page_cache_release(pages[i]);
1786                                 continue;
1787                         }
1788                         if (page_ops & PAGE_CLEAR_DIRTY)
1789                                 clear_page_dirty_for_io(pages[i]);
1790                         if (page_ops & PAGE_SET_WRITEBACK)
1791                                 set_page_writeback(pages[i]);
1792                         if (page_ops & PAGE_SET_ERROR)
1793                                 SetPageError(pages[i]);
1794                         if (page_ops & PAGE_END_WRITEBACK)
1795                                 end_page_writeback(pages[i]);
1796                         if (page_ops & PAGE_UNLOCK)
1797                                 unlock_page(pages[i]);
1798                         page_cache_release(pages[i]);
1799                 }
1800                 nr_pages -= ret;
1801                 index += ret;
1802                 cond_resched();
1803         }
1804         return 0;
1805 }
1806
1807 /*
1808  * count the number of bytes in the tree that have a given bit(s)
1809  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1810  * cached.  The total number found is returned.
1811  */
1812 u64 count_range_bits(struct extent_io_tree *tree,
1813                      u64 *start, u64 search_end, u64 max_bytes,
1814                      unsigned bits, int contig)
1815 {
1816         struct rb_node *node;
1817         struct extent_state *state;
1818         u64 cur_start = *start;
1819         u64 total_bytes = 0;
1820         u64 last = 0;
1821         int found = 0;
1822
1823         if (WARN_ON(search_end <= cur_start))
1824                 return 0;
1825
1826         spin_lock(&tree->lock);
1827         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1828                 total_bytes = tree->dirty_bytes;
1829                 goto out;
1830         }
1831         /*
1832          * this search will find all the extents that end after
1833          * our range starts.
1834          */
1835         node = tree_search(tree, cur_start);
1836         if (!node)
1837                 goto out;
1838
1839         while (1) {
1840                 state = rb_entry(node, struct extent_state, rb_node);
1841                 if (state->start > search_end)
1842                         break;
1843                 if (contig && found && state->start > last + 1)
1844                         break;
1845                 if (state->end >= cur_start && (state->state & bits) == bits) {
1846                         total_bytes += min(search_end, state->end) + 1 -
1847                                        max(cur_start, state->start);
1848                         if (total_bytes >= max_bytes)
1849                                 break;
1850                         if (!found) {
1851                                 *start = max(cur_start, state->start);
1852                                 found = 1;
1853                         }
1854                         last = state->end;
1855                 } else if (contig && found) {
1856                         break;
1857                 }
1858                 node = rb_next(node);
1859                 if (!node)
1860                         break;
1861         }
1862 out:
1863         spin_unlock(&tree->lock);
1864         return total_bytes;
1865 }
1866
1867 /*
1868  * set the private field for a given byte offset in the tree.  If there isn't
1869  * an extent_state there already, this does nothing.
1870  */
1871 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1872 {
1873         struct rb_node *node;
1874         struct extent_state *state;
1875         int ret = 0;
1876
1877         spin_lock(&tree->lock);
1878         /*
1879          * this search will find all the extents that end after
1880          * our range starts.
1881          */
1882         node = tree_search(tree, start);
1883         if (!node) {
1884                 ret = -ENOENT;
1885                 goto out;
1886         }
1887         state = rb_entry(node, struct extent_state, rb_node);
1888         if (state->start != start) {
1889                 ret = -ENOENT;
1890                 goto out;
1891         }
1892         state->private = private;
1893 out:
1894         spin_unlock(&tree->lock);
1895         return ret;
1896 }
1897
1898 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1899 {
1900         struct rb_node *node;
1901         struct extent_state *state;
1902         int ret = 0;
1903
1904         spin_lock(&tree->lock);
1905         /*
1906          * this search will find all the extents that end after
1907          * our range starts.
1908          */
1909         node = tree_search(tree, start);
1910         if (!node) {
1911                 ret = -ENOENT;
1912                 goto out;
1913         }
1914         state = rb_entry(node, struct extent_state, rb_node);
1915         if (state->start != start) {
1916                 ret = -ENOENT;
1917                 goto out;
1918         }
1919         *private = state->private;
1920 out:
1921         spin_unlock(&tree->lock);
1922         return ret;
1923 }
1924
1925 /*
1926  * searches a range in the state tree for a given mask.
1927  * If 'filled' == 1, this returns 1 only if every extent in the tree
1928  * has the bits set.  Otherwise, 1 is returned if any bit in the
1929  * range is found set.
1930  */
1931 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1932                    unsigned bits, int filled, struct extent_state *cached)
1933 {
1934         struct extent_state *state = NULL;
1935         struct rb_node *node;
1936         int bitset = 0;
1937
1938         spin_lock(&tree->lock);
1939         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1940             cached->end > start)
1941                 node = &cached->rb_node;
1942         else
1943                 node = tree_search(tree, start);
1944         while (node && start <= end) {
1945                 state = rb_entry(node, struct extent_state, rb_node);
1946
1947                 if (filled && state->start > start) {
1948                         bitset = 0;
1949                         break;
1950                 }
1951
1952                 if (state->start > end)
1953                         break;
1954
1955                 if (state->state & bits) {
1956                         bitset = 1;
1957                         if (!filled)
1958                                 break;
1959                 } else if (filled) {
1960                         bitset = 0;
1961                         break;
1962                 }
1963
1964                 if (state->end == (u64)-1)
1965                         break;
1966
1967                 start = state->end + 1;
1968                 if (start > end)
1969                         break;
1970                 node = rb_next(node);
1971                 if (!node) {
1972                         if (filled)
1973                                 bitset = 0;
1974                         break;
1975                 }
1976         }
1977         spin_unlock(&tree->lock);
1978         return bitset;
1979 }
1980
1981 /*
1982  * helper function to set a given page up to date if all the
1983  * extents in the tree for that page are up to date
1984  */
1985 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1986 {
1987         u64 start = page_offset(page);
1988         u64 end = start + PAGE_CACHE_SIZE - 1;
1989         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1990                 SetPageUptodate(page);
1991 }
1992
1993 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1994 {
1995         int ret;
1996         int err = 0;
1997         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1998
1999         set_state_private(failure_tree, rec->start, 0);
2000         ret = clear_extent_bits(failure_tree, rec->start,
2001                                 rec->start + rec->len - 1,
2002                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2003         if (ret)
2004                 err = ret;
2005
2006         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
2007                                 rec->start + rec->len - 1,
2008                                 EXTENT_DAMAGED, GFP_NOFS);
2009         if (ret && !err)
2010                 err = ret;
2011
2012         kfree(rec);
2013         return err;
2014 }
2015
2016 /*
2017  * this bypasses the standard btrfs submit functions deliberately, as
2018  * the standard behavior is to write all copies in a raid setup. here we only
2019  * want to write the one bad copy. so we do the mapping for ourselves and issue
2020  * submit_bio directly.
2021  * to avoid any synchronization issues, wait for the data after writing, which
2022  * actually prevents the read that triggered the error from finishing.
2023  * currently, there can be no more than two copies of every data bit. thus,
2024  * exactly one rewrite is required.
2025  */
2026 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2027                       struct page *page, unsigned int pg_offset, int mirror_num)
2028 {
2029         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2030         struct bio *bio;
2031         struct btrfs_device *dev;
2032         u64 map_length = 0;
2033         u64 sector;
2034         struct btrfs_bio *bbio = NULL;
2035         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2036         int ret;
2037
2038         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2039         BUG_ON(!mirror_num);
2040
2041         /* we can't repair anything in raid56 yet */
2042         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2043                 return 0;
2044
2045         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2046         if (!bio)
2047                 return -EIO;
2048         bio->bi_iter.bi_size = 0;
2049         map_length = length;
2050
2051         ret = btrfs_map_block(fs_info, WRITE, logical,
2052                               &map_length, &bbio, mirror_num);
2053         if (ret) {
2054                 bio_put(bio);
2055                 return -EIO;
2056         }
2057         BUG_ON(mirror_num != bbio->mirror_num);
2058         sector = bbio->stripes[mirror_num-1].physical >> 9;
2059         bio->bi_iter.bi_sector = sector;
2060         dev = bbio->stripes[mirror_num-1].dev;
2061         btrfs_put_bbio(bbio);
2062         if (!dev || !dev->bdev || !dev->writeable) {
2063                 bio_put(bio);
2064                 return -EIO;
2065         }
2066         bio->bi_bdev = dev->bdev;
2067         bio_add_page(bio, page, length, pg_offset);
2068
2069         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2070                 /* try to remap that extent elsewhere? */
2071                 bio_put(bio);
2072                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2073                 return -EIO;
2074         }
2075
2076         printk_ratelimited_in_rcu(KERN_INFO
2077                                   "BTRFS: read error corrected: ino %llu off %llu (dev %s sector %llu)\n",
2078                                   btrfs_ino(inode), start,
2079                                   rcu_str_deref(dev->name), sector);
2080         bio_put(bio);
2081         return 0;
2082 }
2083
2084 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2085                          int mirror_num)
2086 {
2087         u64 start = eb->start;
2088         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2089         int ret = 0;
2090
2091         if (root->fs_info->sb->s_flags & MS_RDONLY)
2092                 return -EROFS;
2093
2094         for (i = 0; i < num_pages; i++) {
2095                 struct page *p = eb->pages[i];
2096
2097                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2098                                         PAGE_CACHE_SIZE, start, p,
2099                                         start - page_offset(p), mirror_num);
2100                 if (ret)
2101                         break;
2102                 start += PAGE_CACHE_SIZE;
2103         }
2104
2105         return ret;
2106 }
2107
2108 /*
2109  * each time an IO finishes, we do a fast check in the IO failure tree
2110  * to see if we need to process or clean up an io_failure_record
2111  */
2112 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2113                      unsigned int pg_offset)
2114 {
2115         u64 private;
2116         u64 private_failure;
2117         struct io_failure_record *failrec;
2118         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2119         struct extent_state *state;
2120         int num_copies;
2121         int ret;
2122
2123         private = 0;
2124         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2125                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2126         if (!ret)
2127                 return 0;
2128
2129         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2130                                 &private_failure);
2131         if (ret)
2132                 return 0;
2133
2134         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2135         BUG_ON(!failrec->this_mirror);
2136
2137         if (failrec->in_validation) {
2138                 /* there was no real error, just free the record */
2139                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2140                          failrec->start);
2141                 goto out;
2142         }
2143         if (fs_info->sb->s_flags & MS_RDONLY)
2144                 goto out;
2145
2146         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2147         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2148                                             failrec->start,
2149                                             EXTENT_LOCKED);
2150         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2151
2152         if (state && state->start <= failrec->start &&
2153             state->end >= failrec->start + failrec->len - 1) {
2154                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2155                                               failrec->len);
2156                 if (num_copies > 1)  {
2157                         repair_io_failure(inode, start, failrec->len,
2158                                           failrec->logical, page,
2159                                           pg_offset, failrec->failed_mirror);
2160                 }
2161         }
2162
2163 out:
2164         free_io_failure(inode, failrec);
2165
2166         return 0;
2167 }
2168
2169 /*
2170  * Can be called when
2171  * - hold extent lock
2172  * - under ordered extent
2173  * - the inode is freeing
2174  */
2175 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2176 {
2177         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2178         struct io_failure_record *failrec;
2179         struct extent_state *state, *next;
2180
2181         if (RB_EMPTY_ROOT(&failure_tree->state))
2182                 return;
2183
2184         spin_lock(&failure_tree->lock);
2185         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2186         while (state) {
2187                 if (state->start > end)
2188                         break;
2189
2190                 ASSERT(state->end <= end);
2191
2192                 next = next_state(state);
2193
2194                 failrec = (struct io_failure_record *)(unsigned long)state->private;
2195                 free_extent_state(state);
2196                 kfree(failrec);
2197
2198                 state = next;
2199         }
2200         spin_unlock(&failure_tree->lock);
2201 }
2202
2203 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2204                                 struct io_failure_record **failrec_ret)
2205 {
2206         struct io_failure_record *failrec;
2207         u64 private;
2208         struct extent_map *em;
2209         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2210         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2211         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2212         int ret;
2213         u64 logical;
2214
2215         ret = get_state_private(failure_tree, start, &private);
2216         if (ret) {
2217                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2218                 if (!failrec)
2219                         return -ENOMEM;
2220
2221                 failrec->start = start;
2222                 failrec->len = end - start + 1;
2223                 failrec->this_mirror = 0;
2224                 failrec->bio_flags = 0;
2225                 failrec->in_validation = 0;
2226
2227                 read_lock(&em_tree->lock);
2228                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2229                 if (!em) {
2230                         read_unlock(&em_tree->lock);
2231                         kfree(failrec);
2232                         return -EIO;
2233                 }
2234
2235                 if (em->start > start || em->start + em->len <= start) {
2236                         free_extent_map(em);
2237                         em = NULL;
2238                 }
2239                 read_unlock(&em_tree->lock);
2240                 if (!em) {
2241                         kfree(failrec);
2242                         return -EIO;
2243                 }
2244
2245                 logical = start - em->start;
2246                 logical = em->block_start + logical;
2247                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2248                         logical = em->block_start;
2249                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2250                         extent_set_compress_type(&failrec->bio_flags,
2251                                                  em->compress_type);
2252                 }
2253
2254                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2255                          logical, start, failrec->len);
2256
2257                 failrec->logical = logical;
2258                 free_extent_map(em);
2259
2260                 /* set the bits in the private failure tree */
2261                 ret = set_extent_bits(failure_tree, start, end,
2262                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2263                 if (ret >= 0)
2264                         ret = set_state_private(failure_tree, start,
2265                                                 (u64)(unsigned long)failrec);
2266                 /* set the bits in the inode's tree */
2267                 if (ret >= 0)
2268                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2269                                                 GFP_NOFS);
2270                 if (ret < 0) {
2271                         kfree(failrec);
2272                         return ret;
2273                 }
2274         } else {
2275                 failrec = (struct io_failure_record *)(unsigned long)private;
2276                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2277                          failrec->logical, failrec->start, failrec->len,
2278                          failrec->in_validation);
2279                 /*
2280                  * when data can be on disk more than twice, add to failrec here
2281                  * (e.g. with a list for failed_mirror) to make
2282                  * clean_io_failure() clean all those errors at once.
2283                  */
2284         }
2285
2286         *failrec_ret = failrec;
2287
2288         return 0;
2289 }
2290
2291 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2292                            struct io_failure_record *failrec, int failed_mirror)
2293 {
2294         int num_copies;
2295
2296         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2297                                       failrec->logical, failrec->len);
2298         if (num_copies == 1) {
2299                 /*
2300                  * we only have a single copy of the data, so don't bother with
2301                  * all the retry and error correction code that follows. no
2302                  * matter what the error is, it is very likely to persist.
2303                  */
2304                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2305                          num_copies, failrec->this_mirror, failed_mirror);
2306                 return 0;
2307         }
2308
2309         /*
2310          * there are two premises:
2311          *      a) deliver good data to the caller
2312          *      b) correct the bad sectors on disk
2313          */
2314         if (failed_bio->bi_vcnt > 1) {
2315                 /*
2316                  * to fulfill b), we need to know the exact failing sectors, as
2317                  * we don't want to rewrite any more than the failed ones. thus,
2318                  * we need separate read requests for the failed bio
2319                  *
2320                  * if the following BUG_ON triggers, our validation request got
2321                  * merged. we need separate requests for our algorithm to work.
2322                  */
2323                 BUG_ON(failrec->in_validation);
2324                 failrec->in_validation = 1;
2325                 failrec->this_mirror = failed_mirror;
2326         } else {
2327                 /*
2328                  * we're ready to fulfill a) and b) alongside. get a good copy
2329                  * of the failed sector and if we succeed, we have setup
2330                  * everything for repair_io_failure to do the rest for us.
2331                  */
2332                 if (failrec->in_validation) {
2333                         BUG_ON(failrec->this_mirror != failed_mirror);
2334                         failrec->in_validation = 0;
2335                         failrec->this_mirror = 0;
2336                 }
2337                 failrec->failed_mirror = failed_mirror;
2338                 failrec->this_mirror++;
2339                 if (failrec->this_mirror == failed_mirror)
2340                         failrec->this_mirror++;
2341         }
2342
2343         if (failrec->this_mirror > num_copies) {
2344                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2345                          num_copies, failrec->this_mirror, failed_mirror);
2346                 return 0;
2347         }
2348
2349         return 1;
2350 }
2351
2352
2353 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2354                                     struct io_failure_record *failrec,
2355                                     struct page *page, int pg_offset, int icsum,
2356                                     bio_end_io_t *endio_func, void *data)
2357 {
2358         struct bio *bio;
2359         struct btrfs_io_bio *btrfs_failed_bio;
2360         struct btrfs_io_bio *btrfs_bio;
2361
2362         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2363         if (!bio)
2364                 return NULL;
2365
2366         bio->bi_end_io = endio_func;
2367         bio->bi_iter.bi_sector = failrec->logical >> 9;
2368         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2369         bio->bi_iter.bi_size = 0;
2370         bio->bi_private = data;
2371
2372         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2373         if (btrfs_failed_bio->csum) {
2374                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2375                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2376
2377                 btrfs_bio = btrfs_io_bio(bio);
2378                 btrfs_bio->csum = btrfs_bio->csum_inline;
2379                 icsum *= csum_size;
2380                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2381                        csum_size);
2382         }
2383
2384         bio_add_page(bio, page, failrec->len, pg_offset);
2385
2386         return bio;
2387 }
2388
2389 /*
2390  * this is a generic handler for readpage errors (default
2391  * readpage_io_failed_hook). if other copies exist, read those and write back
2392  * good data to the failed position. does not investigate in remapping the
2393  * failed extent elsewhere, hoping the device will be smart enough to do this as
2394  * needed
2395  */
2396
2397 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2398                               struct page *page, u64 start, u64 end,
2399                               int failed_mirror)
2400 {
2401         struct io_failure_record *failrec;
2402         struct inode *inode = page->mapping->host;
2403         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2404         struct bio *bio;
2405         int read_mode;
2406         int ret;
2407
2408         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2409
2410         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2411         if (ret)
2412                 return ret;
2413
2414         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2415         if (!ret) {
2416                 free_io_failure(inode, failrec);
2417                 return -EIO;
2418         }
2419
2420         if (failed_bio->bi_vcnt > 1)
2421                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2422         else
2423                 read_mode = READ_SYNC;
2424
2425         phy_offset >>= inode->i_sb->s_blocksize_bits;
2426         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2427                                       start - page_offset(page),
2428                                       (int)phy_offset, failed_bio->bi_end_io,
2429                                       NULL);
2430         if (!bio) {
2431                 free_io_failure(inode, failrec);
2432                 return -EIO;
2433         }
2434
2435         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2436                  read_mode, failrec->this_mirror, failrec->in_validation);
2437
2438         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2439                                          failrec->this_mirror,
2440                                          failrec->bio_flags, 0);
2441         if (ret) {
2442                 free_io_failure(inode, failrec);
2443                 bio_put(bio);
2444         }
2445
2446         return ret;
2447 }
2448
2449 /* lots and lots of room for performance fixes in the end_bio funcs */
2450
2451 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2452 {
2453         int uptodate = (err == 0);
2454         struct extent_io_tree *tree;
2455         int ret = 0;
2456
2457         tree = &BTRFS_I(page->mapping->host)->io_tree;
2458
2459         if (tree->ops && tree->ops->writepage_end_io_hook) {
2460                 ret = tree->ops->writepage_end_io_hook(page, start,
2461                                                end, NULL, uptodate);
2462                 if (ret)
2463                         uptodate = 0;
2464         }
2465
2466         if (!uptodate) {
2467                 ClearPageUptodate(page);
2468                 SetPageError(page);
2469                 ret = ret < 0 ? ret : -EIO;
2470                 mapping_set_error(page->mapping, ret);
2471         }
2472         return 0;
2473 }
2474
2475 /*
2476  * after a writepage IO is done, we need to:
2477  * clear the uptodate bits on error
2478  * clear the writeback bits in the extent tree for this IO
2479  * end_page_writeback if the page has no more pending IO
2480  *
2481  * Scheduling is not allowed, so the extent state tree is expected
2482  * to have one and only one object corresponding to this IO.
2483  */
2484 static void end_bio_extent_writepage(struct bio *bio, int err)
2485 {
2486         struct bio_vec *bvec;
2487         u64 start;
2488         u64 end;
2489         int i;
2490
2491         bio_for_each_segment_all(bvec, bio, i) {
2492                 struct page *page = bvec->bv_page;
2493
2494                 /* We always issue full-page reads, but if some block
2495                  * in a page fails to read, blk_update_request() will
2496                  * advance bv_offset and adjust bv_len to compensate.
2497                  * Print a warning for nonzero offsets, and an error
2498                  * if they don't add up to a full page.  */
2499                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2500                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2501                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2502                                    "partial page write in btrfs with offset %u and length %u",
2503                                         bvec->bv_offset, bvec->bv_len);
2504                         else
2505                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2506                                    "incomplete page write in btrfs with offset %u and "
2507                                    "length %u",
2508                                         bvec->bv_offset, bvec->bv_len);
2509                 }
2510
2511                 start = page_offset(page);
2512                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2513
2514                 if (end_extent_writepage(page, err, start, end))
2515                         continue;
2516
2517                 end_page_writeback(page);
2518         }
2519
2520         bio_put(bio);
2521 }
2522
2523 static void
2524 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2525                               int uptodate)
2526 {
2527         struct extent_state *cached = NULL;
2528         u64 end = start + len - 1;
2529
2530         if (uptodate && tree->track_uptodate)
2531                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2532         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2533 }
2534
2535 /*
2536  * after a readpage IO is done, we need to:
2537  * clear the uptodate bits on error
2538  * set the uptodate bits if things worked
2539  * set the page up to date if all extents in the tree are uptodate
2540  * clear the lock bit in the extent tree
2541  * unlock the page if there are no other extents locked for it
2542  *
2543  * Scheduling is not allowed, so the extent state tree is expected
2544  * to have one and only one object corresponding to this IO.
2545  */
2546 static void end_bio_extent_readpage(struct bio *bio, int err)
2547 {
2548         struct bio_vec *bvec;
2549         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2550         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2551         struct extent_io_tree *tree;
2552         u64 offset = 0;
2553         u64 start;
2554         u64 end;
2555         u64 len;
2556         u64 extent_start = 0;
2557         u64 extent_len = 0;
2558         int mirror;
2559         int ret;
2560         int i;
2561
2562         if (err)
2563                 uptodate = 0;
2564
2565         bio_for_each_segment_all(bvec, bio, i) {
2566                 struct page *page = bvec->bv_page;
2567                 struct inode *inode = page->mapping->host;
2568
2569                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2570                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector, err,
2571                          io_bio->mirror_num);
2572                 tree = &BTRFS_I(inode)->io_tree;
2573
2574                 /* We always issue full-page reads, but if some block
2575                  * in a page fails to read, blk_update_request() will
2576                  * advance bv_offset and adjust bv_len to compensate.
2577                  * Print a warning for nonzero offsets, and an error
2578                  * if they don't add up to a full page.  */
2579                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2580                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2581                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2582                                    "partial page read in btrfs with offset %u and length %u",
2583                                         bvec->bv_offset, bvec->bv_len);
2584                         else
2585                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2586                                    "incomplete page read in btrfs with offset %u and "
2587                                    "length %u",
2588                                         bvec->bv_offset, bvec->bv_len);
2589                 }
2590
2591                 start = page_offset(page);
2592                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2593                 len = bvec->bv_len;
2594
2595                 mirror = io_bio->mirror_num;
2596                 if (likely(uptodate && tree->ops &&
2597                            tree->ops->readpage_end_io_hook)) {
2598                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2599                                                               page, start, end,
2600                                                               mirror);
2601                         if (ret)
2602                                 uptodate = 0;
2603                         else
2604                                 clean_io_failure(inode, start, page, 0);
2605                 }
2606
2607                 if (likely(uptodate))
2608                         goto readpage_ok;
2609
2610                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2611                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2612                         if (!ret && !err &&
2613                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2614                                 uptodate = 1;
2615                 } else {
2616                         /*
2617                          * The generic bio_readpage_error handles errors the
2618                          * following way: If possible, new read requests are
2619                          * created and submitted and will end up in
2620                          * end_bio_extent_readpage as well (if we're lucky, not
2621                          * in the !uptodate case). In that case it returns 0 and
2622                          * we just go on with the next page in our bio. If it
2623                          * can't handle the error it will return -EIO and we
2624                          * remain responsible for that page.
2625                          */
2626                         ret = bio_readpage_error(bio, offset, page, start, end,
2627                                                  mirror);
2628                         if (ret == 0) {
2629                                 uptodate =
2630                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2631                                 if (err)
2632                                         uptodate = 0;
2633                                 offset += len;
2634                                 continue;
2635                         }
2636                 }
2637 readpage_ok:
2638                 if (likely(uptodate)) {
2639                         loff_t i_size = i_size_read(inode);
2640                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2641                         unsigned off;
2642
2643                         /* Zero out the end if this page straddles i_size */
2644                         off = i_size & (PAGE_CACHE_SIZE-1);
2645                         if (page->index == end_index && off)
2646                                 zero_user_segment(page, off, PAGE_CACHE_SIZE);
2647                         SetPageUptodate(page);
2648                 } else {
2649                         ClearPageUptodate(page);
2650                         SetPageError(page);
2651                 }
2652                 unlock_page(page);
2653                 offset += len;
2654
2655                 if (unlikely(!uptodate)) {
2656                         if (extent_len) {
2657                                 endio_readpage_release_extent(tree,
2658                                                               extent_start,
2659                                                               extent_len, 1);
2660                                 extent_start = 0;
2661                                 extent_len = 0;
2662                         }
2663                         endio_readpage_release_extent(tree, start,
2664                                                       end - start + 1, 0);
2665                 } else if (!extent_len) {
2666                         extent_start = start;
2667                         extent_len = end + 1 - start;
2668                 } else if (extent_start + extent_len == start) {
2669                         extent_len += end + 1 - start;
2670                 } else {
2671                         endio_readpage_release_extent(tree, extent_start,
2672                                                       extent_len, uptodate);
2673                         extent_start = start;
2674                         extent_len = end + 1 - start;
2675                 }
2676         }
2677
2678         if (extent_len)
2679                 endio_readpage_release_extent(tree, extent_start, extent_len,
2680                                               uptodate);
2681         if (io_bio->end_io)
2682                 io_bio->end_io(io_bio, err);
2683         bio_put(bio);
2684 }
2685
2686 /*
2687  * this allocates from the btrfs_bioset.  We're returning a bio right now
2688  * but you can call btrfs_io_bio for the appropriate container_of magic
2689  */
2690 struct bio *
2691 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2692                 gfp_t gfp_flags)
2693 {
2694         struct btrfs_io_bio *btrfs_bio;
2695         struct bio *bio;
2696
2697         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2698
2699         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2700                 while (!bio && (nr_vecs /= 2)) {
2701                         bio = bio_alloc_bioset(gfp_flags,
2702                                                nr_vecs, btrfs_bioset);
2703                 }
2704         }
2705
2706         if (bio) {
2707                 bio->bi_bdev = bdev;
2708                 bio->bi_iter.bi_sector = first_sector;
2709                 btrfs_bio = btrfs_io_bio(bio);
2710                 btrfs_bio->csum = NULL;
2711                 btrfs_bio->csum_allocated = NULL;
2712                 btrfs_bio->end_io = NULL;
2713         }
2714         return bio;
2715 }
2716
2717 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2718 {
2719         struct btrfs_io_bio *btrfs_bio;
2720         struct bio *new;
2721
2722         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2723         if (new) {
2724                 btrfs_bio = btrfs_io_bio(new);
2725                 btrfs_bio->csum = NULL;
2726                 btrfs_bio->csum_allocated = NULL;
2727                 btrfs_bio->end_io = NULL;
2728         }
2729         return new;
2730 }
2731
2732 /* this also allocates from the btrfs_bioset */
2733 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2734 {
2735         struct btrfs_io_bio *btrfs_bio;
2736         struct bio *bio;
2737
2738         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2739         if (bio) {
2740                 btrfs_bio = btrfs_io_bio(bio);
2741                 btrfs_bio->csum = NULL;
2742                 btrfs_bio->csum_allocated = NULL;
2743                 btrfs_bio->end_io = NULL;
2744         }
2745         return bio;
2746 }
2747
2748
2749 static int __must_check submit_one_bio(int rw, struct bio *bio,
2750                                        int mirror_num, unsigned long bio_flags)
2751 {
2752         int ret = 0;
2753         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2754         struct page *page = bvec->bv_page;
2755         struct extent_io_tree *tree = bio->bi_private;
2756         u64 start;
2757
2758         start = page_offset(page) + bvec->bv_offset;
2759
2760         bio->bi_private = NULL;
2761
2762         bio_get(bio);
2763
2764         if (tree->ops && tree->ops->submit_bio_hook)
2765                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2766                                            mirror_num, bio_flags, start);
2767         else
2768                 btrfsic_submit_bio(rw, bio);
2769
2770         bio_put(bio);
2771         return ret;
2772 }
2773
2774 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2775                      unsigned long offset, size_t size, struct bio *bio,
2776                      unsigned long bio_flags)
2777 {
2778         int ret = 0;
2779         if (tree->ops && tree->ops->merge_bio_hook)
2780                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2781                                                 bio_flags);
2782         BUG_ON(ret < 0);
2783         return ret;
2784
2785 }
2786
2787 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2788                               struct page *page, sector_t sector,
2789                               size_t size, unsigned long offset,
2790                               struct block_device *bdev,
2791                               struct bio **bio_ret,
2792                               unsigned long max_pages,
2793                               bio_end_io_t end_io_func,
2794                               int mirror_num,
2795                               unsigned long prev_bio_flags,
2796                               unsigned long bio_flags)
2797 {
2798         int ret = 0;
2799         struct bio *bio;
2800         int nr;
2801         int contig = 0;
2802         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2803         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2804         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2805
2806         if (bio_ret && *bio_ret) {
2807                 bio = *bio_ret;
2808                 if (old_compressed)
2809                         contig = bio->bi_iter.bi_sector == sector;
2810                 else
2811                         contig = bio_end_sector(bio) == sector;
2812
2813                 if (prev_bio_flags != bio_flags || !contig ||
2814                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2815                     bio_add_page(bio, page, page_size, offset) < page_size) {
2816                         ret = submit_one_bio(rw, bio, mirror_num,
2817                                              prev_bio_flags);
2818                         if (ret < 0) {
2819                                 *bio_ret = NULL;
2820                                 return ret;
2821                         }
2822                         bio = NULL;
2823                 } else {
2824                         return 0;
2825                 }
2826         }
2827         if (this_compressed)
2828                 nr = BIO_MAX_PAGES;
2829         else
2830                 nr = bio_get_nr_vecs(bdev);
2831
2832         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2833         if (!bio)
2834                 return -ENOMEM;
2835
2836         bio_add_page(bio, page, page_size, offset);
2837         bio->bi_end_io = end_io_func;
2838         bio->bi_private = tree;
2839
2840         if (bio_ret)
2841                 *bio_ret = bio;
2842         else
2843                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2844
2845         return ret;
2846 }
2847
2848 static void attach_extent_buffer_page(struct extent_buffer *eb,
2849                                       struct page *page)
2850 {
2851         if (!PagePrivate(page)) {
2852                 SetPagePrivate(page);
2853                 page_cache_get(page);
2854                 set_page_private(page, (unsigned long)eb);
2855         } else {
2856                 WARN_ON(page->private != (unsigned long)eb);
2857         }
2858 }
2859
2860 void set_page_extent_mapped(struct page *page)
2861 {
2862         if (!PagePrivate(page)) {
2863                 SetPagePrivate(page);
2864                 page_cache_get(page);
2865                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2866         }
2867 }
2868
2869 static struct extent_map *
2870 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2871                  u64 start, u64 len, get_extent_t *get_extent,
2872                  struct extent_map **em_cached)
2873 {
2874         struct extent_map *em;
2875
2876         if (em_cached && *em_cached) {
2877                 em = *em_cached;
2878                 if (extent_map_in_tree(em) && start >= em->start &&
2879                     start < extent_map_end(em)) {
2880                         atomic_inc(&em->refs);
2881                         return em;
2882                 }
2883
2884                 free_extent_map(em);
2885                 *em_cached = NULL;
2886         }
2887
2888         em = get_extent(inode, page, pg_offset, start, len, 0);
2889         if (em_cached && !IS_ERR_OR_NULL(em)) {
2890                 BUG_ON(*em_cached);
2891                 atomic_inc(&em->refs);
2892                 *em_cached = em;
2893         }
2894         return em;
2895 }
2896 /*
2897  * basic readpage implementation.  Locked extent state structs are inserted
2898  * into the tree that are removed when the IO is done (by the end_io
2899  * handlers)
2900  * XXX JDM: This needs looking at to ensure proper page locking
2901  */
2902 static int __do_readpage(struct extent_io_tree *tree,
2903                          struct page *page,
2904                          get_extent_t *get_extent,
2905                          struct extent_map **em_cached,
2906                          struct bio **bio, int mirror_num,
2907                          unsigned long *bio_flags, int rw)
2908 {
2909         struct inode *inode = page->mapping->host;
2910         u64 start = page_offset(page);
2911         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2912         u64 end;
2913         u64 cur = start;
2914         u64 extent_offset;
2915         u64 last_byte = i_size_read(inode);
2916         u64 block_start;
2917         u64 cur_end;
2918         sector_t sector;
2919         struct extent_map *em;
2920         struct block_device *bdev;
2921         int ret;
2922         int nr = 0;
2923         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2924         size_t pg_offset = 0;
2925         size_t iosize;
2926         size_t disk_io_size;
2927         size_t blocksize = inode->i_sb->s_blocksize;
2928         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2929
2930         set_page_extent_mapped(page);
2931
2932         end = page_end;
2933         if (!PageUptodate(page)) {
2934                 if (cleancache_get_page(page) == 0) {
2935                         BUG_ON(blocksize != PAGE_SIZE);
2936                         unlock_extent(tree, start, end);
2937                         goto out;
2938                 }
2939         }
2940
2941         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2942                 char *userpage;
2943                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2944
2945                 if (zero_offset) {
2946                         iosize = PAGE_CACHE_SIZE - zero_offset;
2947                         userpage = kmap_atomic(page);
2948                         memset(userpage + zero_offset, 0, iosize);
2949                         flush_dcache_page(page);
2950                         kunmap_atomic(userpage);
2951                 }
2952         }
2953         while (cur <= end) {
2954                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2955
2956                 if (cur >= last_byte) {
2957                         char *userpage;
2958                         struct extent_state *cached = NULL;
2959
2960                         iosize = PAGE_CACHE_SIZE - pg_offset;
2961                         userpage = kmap_atomic(page);
2962                         memset(userpage + pg_offset, 0, iosize);
2963                         flush_dcache_page(page);
2964                         kunmap_atomic(userpage);
2965                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2966                                             &cached, GFP_NOFS);
2967                         if (!parent_locked)
2968                                 unlock_extent_cached(tree, cur,
2969                                                      cur + iosize - 1,
2970                                                      &cached, GFP_NOFS);
2971                         break;
2972                 }
2973                 em = __get_extent_map(inode, page, pg_offset, cur,
2974                                       end - cur + 1, get_extent, em_cached);
2975                 if (IS_ERR_OR_NULL(em)) {
2976                         SetPageError(page);
2977                         if (!parent_locked)
2978                                 unlock_extent(tree, cur, end);
2979                         break;
2980                 }
2981                 extent_offset = cur - em->start;
2982                 BUG_ON(extent_map_end(em) <= cur);
2983                 BUG_ON(end < cur);
2984
2985                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2986                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2987                         extent_set_compress_type(&this_bio_flag,
2988                                                  em->compress_type);
2989                 }
2990
2991                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2992                 cur_end = min(extent_map_end(em) - 1, end);
2993                 iosize = ALIGN(iosize, blocksize);
2994                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2995                         disk_io_size = em->block_len;
2996                         sector = em->block_start >> 9;
2997                 } else {
2998                         sector = (em->block_start + extent_offset) >> 9;
2999                         disk_io_size = iosize;
3000                 }
3001                 bdev = em->bdev;
3002                 block_start = em->block_start;
3003                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3004                         block_start = EXTENT_MAP_HOLE;
3005                 free_extent_map(em);
3006                 em = NULL;
3007
3008                 /* we've found a hole, just zero and go on */
3009                 if (block_start == EXTENT_MAP_HOLE) {
3010                         char *userpage;
3011                         struct extent_state *cached = NULL;
3012
3013                         userpage = kmap_atomic(page);
3014                         memset(userpage + pg_offset, 0, iosize);
3015                         flush_dcache_page(page);
3016                         kunmap_atomic(userpage);
3017
3018                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3019                                             &cached, GFP_NOFS);
3020                         unlock_extent_cached(tree, cur, cur + iosize - 1,
3021                                              &cached, GFP_NOFS);
3022                         cur = cur + iosize;
3023                         pg_offset += iosize;
3024                         continue;
3025                 }
3026                 /* the get_extent function already copied into the page */
3027                 if (test_range_bit(tree, cur, cur_end,
3028                                    EXTENT_UPTODATE, 1, NULL)) {
3029                         check_page_uptodate(tree, page);
3030                         if (!parent_locked)
3031                                 unlock_extent(tree, cur, cur + iosize - 1);
3032                         cur = cur + iosize;
3033                         pg_offset += iosize;
3034                         continue;
3035                 }
3036                 /* we have an inline extent but it didn't get marked up
3037                  * to date.  Error out
3038                  */
3039                 if (block_start == EXTENT_MAP_INLINE) {
3040                         SetPageError(page);
3041                         if (!parent_locked)
3042                                 unlock_extent(tree, cur, cur + iosize - 1);
3043                         cur = cur + iosize;
3044                         pg_offset += iosize;
3045                         continue;
3046                 }
3047
3048                 pnr -= page->index;
3049                 ret = submit_extent_page(rw, tree, page,
3050                                          sector, disk_io_size, pg_offset,
3051                                          bdev, bio, pnr,
3052                                          end_bio_extent_readpage, mirror_num,
3053                                          *bio_flags,
3054                                          this_bio_flag);
3055                 if (!ret) {
3056                         nr++;
3057                         *bio_flags = this_bio_flag;
3058                 } else {
3059                         SetPageError(page);
3060                         if (!parent_locked)
3061                                 unlock_extent(tree, cur, cur + iosize - 1);
3062                 }
3063                 cur = cur + iosize;
3064                 pg_offset += iosize;
3065         }
3066 out:
3067         if (!nr) {
3068                 if (!PageError(page))
3069                         SetPageUptodate(page);
3070                 unlock_page(page);
3071         }
3072         return 0;
3073 }
3074
3075 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3076                                              struct page *pages[], int nr_pages,
3077                                              u64 start, u64 end,
3078                                              get_extent_t *get_extent,
3079                                              struct extent_map **em_cached,
3080                                              struct bio **bio, int mirror_num,
3081                                              unsigned long *bio_flags, int rw)
3082 {
3083         struct inode *inode;
3084         struct btrfs_ordered_extent *ordered;
3085         int index;
3086
3087         inode = pages[0]->mapping->host;
3088         while (1) {
3089                 lock_extent(tree, start, end);
3090                 ordered = btrfs_lookup_ordered_range(inode, start,
3091                                                      end - start + 1);
3092                 if (!ordered)
3093                         break;
3094                 unlock_extent(tree, start, end);
3095                 btrfs_start_ordered_extent(inode, ordered, 1);
3096                 btrfs_put_ordered_extent(ordered);
3097         }
3098
3099         for (index = 0; index < nr_pages; index++) {
3100                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3101                               mirror_num, bio_flags, rw);
3102                 page_cache_release(pages[index]);
3103         }
3104 }
3105
3106 static void __extent_readpages(struct extent_io_tree *tree,
3107                                struct page *pages[],
3108                                int nr_pages, get_extent_t *get_extent,
3109                                struct extent_map **em_cached,
3110                                struct bio **bio, int mirror_num,
3111                                unsigned long *bio_flags, int rw)
3112 {
3113         u64 start = 0;
3114         u64 end = 0;
3115         u64 page_start;
3116         int index;
3117         int first_index = 0;
3118
3119         for (index = 0; index < nr_pages; index++) {
3120                 page_start = page_offset(pages[index]);
3121                 if (!end) {
3122                         start = page_start;
3123                         end = start + PAGE_CACHE_SIZE - 1;
3124                         first_index = index;
3125                 } else if (end + 1 == page_start) {
3126                         end += PAGE_CACHE_SIZE;
3127                 } else {
3128                         __do_contiguous_readpages(tree, &pages[first_index],
3129                                                   index - first_index, start,
3130                                                   end, get_extent, em_cached,
3131                                                   bio, mirror_num, bio_flags,
3132                                                   rw);
3133                         start = page_start;
3134                         end = start + PAGE_CACHE_SIZE - 1;
3135                         first_index = index;
3136                 }
3137         }
3138
3139         if (end)
3140                 __do_contiguous_readpages(tree, &pages[first_index],
3141                                           index - first_index, start,
3142                                           end, get_extent, em_cached, bio,
3143                                           mirror_num, bio_flags, rw);
3144 }
3145
3146 static int __extent_read_full_page(struct extent_io_tree *tree,
3147                                    struct page *page,
3148                                    get_extent_t *get_extent,
3149                                    struct bio **bio, int mirror_num,
3150                                    unsigned long *bio_flags, int rw)
3151 {
3152         struct inode *inode = page->mapping->host;
3153         struct btrfs_ordered_extent *ordered;
3154         u64 start = page_offset(page);
3155         u64 end = start + PAGE_CACHE_SIZE - 1;
3156         int ret;
3157
3158         while (1) {
3159                 lock_extent(tree, start, end);
3160                 ordered = btrfs_lookup_ordered_extent(inode, start);
3161                 if (!ordered)
3162                         break;
3163                 unlock_extent(tree, start, end);
3164                 btrfs_start_ordered_extent(inode, ordered, 1);
3165                 btrfs_put_ordered_extent(ordered);
3166         }
3167
3168         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3169                             bio_flags, rw);
3170         return ret;
3171 }
3172
3173 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3174                             get_extent_t *get_extent, int mirror_num)
3175 {
3176         struct bio *bio = NULL;
3177         unsigned long bio_flags = 0;
3178         int ret;
3179
3180         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3181                                       &bio_flags, READ);
3182         if (bio)
3183                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3184         return ret;
3185 }
3186
3187 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3188                                  get_extent_t *get_extent, int mirror_num)
3189 {
3190         struct bio *bio = NULL;
3191         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3192         int ret;
3193
3194         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3195                                       &bio_flags, READ);
3196         if (bio)
3197                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3198         return ret;
3199 }
3200
3201 static noinline void update_nr_written(struct page *page,
3202                                       struct writeback_control *wbc,
3203                                       unsigned long nr_written)
3204 {
3205         wbc->nr_to_write -= nr_written;
3206         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3207             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3208                 page->mapping->writeback_index = page->index + nr_written;
3209 }
3210
3211 /*
3212  * helper for __extent_writepage, doing all of the delayed allocation setup.
3213  *
3214  * This returns 1 if our fill_delalloc function did all the work required
3215  * to write the page (copy into inline extent).  In this case the IO has
3216  * been started and the page is already unlocked.
3217  *
3218  * This returns 0 if all went well (page still locked)
3219  * This returns < 0 if there were errors (page still locked)
3220  */
3221 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3222                               struct page *page, struct writeback_control *wbc,
3223                               struct extent_page_data *epd,
3224                               u64 delalloc_start,
3225                               unsigned long *nr_written)
3226 {
3227         struct extent_io_tree *tree = epd->tree;
3228         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3229         u64 nr_delalloc;
3230         u64 delalloc_to_write = 0;
3231         u64 delalloc_end = 0;
3232         int ret;
3233         int page_started = 0;
3234
3235         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3236                 return 0;
3237
3238         while (delalloc_end < page_end) {
3239                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3240                                                page,
3241                                                &delalloc_start,
3242                                                &delalloc_end,
3243                                                BTRFS_MAX_EXTENT_SIZE);
3244                 if (nr_delalloc == 0) {
3245                         delalloc_start = delalloc_end + 1;
3246                         continue;
3247                 }
3248                 ret = tree->ops->fill_delalloc(inode, page,
3249                                                delalloc_start,
3250                                                delalloc_end,
3251                                                &page_started,
3252                                                nr_written);
3253                 /* File system has been set read-only */
3254                 if (ret) {
3255                         SetPageError(page);
3256                         /* fill_delalloc should be return < 0 for error
3257                          * but just in case, we use > 0 here meaning the
3258                          * IO is started, so we don't want to return > 0
3259                          * unless things are going well.
3260                          */
3261                         ret = ret < 0 ? ret : -EIO;
3262                         goto done;
3263                 }
3264                 /*
3265                  * delalloc_end is already one less than the total
3266                  * length, so we don't subtract one from
3267                  * PAGE_CACHE_SIZE
3268                  */
3269                 delalloc_to_write += (delalloc_end - delalloc_start +
3270                                       PAGE_CACHE_SIZE) >>
3271                                       PAGE_CACHE_SHIFT;
3272                 delalloc_start = delalloc_end + 1;
3273         }
3274         if (wbc->nr_to_write < delalloc_to_write) {
3275                 int thresh = 8192;
3276
3277                 if (delalloc_to_write < thresh * 2)
3278                         thresh = delalloc_to_write;
3279                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3280                                          thresh);
3281         }
3282
3283         /* did the fill delalloc function already unlock and start
3284          * the IO?
3285          */
3286         if (page_started) {
3287                 /*
3288                  * we've unlocked the page, so we can't update
3289                  * the mapping's writeback index, just update
3290                  * nr_to_write.
3291                  */
3292                 wbc->nr_to_write -= *nr_written;
3293                 return 1;
3294         }
3295
3296         ret = 0;
3297
3298 done:
3299         return ret;
3300 }
3301
3302 /*
3303  * helper for __extent_writepage.  This calls the writepage start hooks,
3304  * and does the loop to map the page into extents and bios.
3305  *
3306  * We return 1 if the IO is started and the page is unlocked,
3307  * 0 if all went well (page still locked)
3308  * < 0 if there were errors (page still locked)
3309  */
3310 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3311                                  struct page *page,
3312                                  struct writeback_control *wbc,
3313                                  struct extent_page_data *epd,
3314                                  loff_t i_size,
3315                                  unsigned long nr_written,
3316                                  int write_flags, int *nr_ret)
3317 {
3318         struct extent_io_tree *tree = epd->tree;
3319         u64 start = page_offset(page);
3320         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3321         u64 end;
3322         u64 cur = start;
3323         u64 extent_offset;
3324         u64 block_start;
3325         u64 iosize;
3326         sector_t sector;
3327         struct extent_state *cached_state = NULL;
3328         struct extent_map *em;
3329         struct block_device *bdev;
3330         size_t pg_offset = 0;
3331         size_t blocksize;
3332         int ret = 0;
3333         int nr = 0;
3334         bool compressed;
3335
3336         if (tree->ops && tree->ops->writepage_start_hook) {
3337                 ret = tree->ops->writepage_start_hook(page, start,
3338                                                       page_end);
3339                 if (ret) {
3340                         /* Fixup worker will requeue */
3341                         if (ret == -EBUSY)
3342                                 wbc->pages_skipped++;
3343                         else
3344                                 redirty_page_for_writepage(wbc, page);
3345
3346                         update_nr_written(page, wbc, nr_written);
3347                         unlock_page(page);
3348                         ret = 1;
3349                         goto done_unlocked;
3350                 }
3351         }
3352
3353         /*
3354          * we don't want to touch the inode after unlocking the page,
3355          * so we update the mapping writeback index now
3356          */
3357         update_nr_written(page, wbc, nr_written + 1);
3358
3359         end = page_end;
3360         if (i_size <= start) {
3361                 if (tree->ops && tree->ops->writepage_end_io_hook)
3362                         tree->ops->writepage_end_io_hook(page, start,
3363                                                          page_end, NULL, 1);
3364                 goto done;
3365         }
3366
3367         blocksize = inode->i_sb->s_blocksize;
3368
3369         while (cur <= end) {
3370                 u64 em_end;
3371                 if (cur >= i_size) {
3372                         if (tree->ops && tree->ops->writepage_end_io_hook)
3373                                 tree->ops->writepage_end_io_hook(page, cur,
3374                                                          page_end, NULL, 1);
3375                         break;
3376                 }
3377                 em = epd->get_extent(inode, page, pg_offset, cur,
3378                                      end - cur + 1, 1);
3379                 if (IS_ERR_OR_NULL(em)) {
3380                         SetPageError(page);
3381                         ret = PTR_ERR_OR_ZERO(em);
3382                         break;
3383                 }
3384
3385                 extent_offset = cur - em->start;
3386                 em_end = extent_map_end(em);
3387                 BUG_ON(em_end <= cur);
3388                 BUG_ON(end < cur);
3389                 iosize = min(em_end - cur, end - cur + 1);
3390                 iosize = ALIGN(iosize, blocksize);
3391                 sector = (em->block_start + extent_offset) >> 9;
3392                 bdev = em->bdev;
3393                 block_start = em->block_start;
3394                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3395                 free_extent_map(em);
3396                 em = NULL;
3397
3398                 /*
3399                  * compressed and inline extents are written through other
3400                  * paths in the FS
3401                  */
3402                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3403                     block_start == EXTENT_MAP_INLINE) {
3404                         /*
3405                          * end_io notification does not happen here for
3406                          * compressed extents
3407                          */
3408                         if (!compressed && tree->ops &&
3409                             tree->ops->writepage_end_io_hook)
3410                                 tree->ops->writepage_end_io_hook(page, cur,
3411                                                          cur + iosize - 1,
3412                                                          NULL, 1);
3413                         else if (compressed) {
3414                                 /* we don't want to end_page_writeback on
3415                                  * a compressed extent.  this happens
3416                                  * elsewhere
3417                                  */
3418                                 nr++;
3419                         }
3420
3421                         cur += iosize;
3422                         pg_offset += iosize;
3423                         continue;
3424                 }
3425
3426                 if (tree->ops && tree->ops->writepage_io_hook) {
3427                         ret = tree->ops->writepage_io_hook(page, cur,
3428                                                 cur + iosize - 1);
3429                 } else {
3430                         ret = 0;
3431                 }
3432                 if (ret) {
3433                         SetPageError(page);
3434                 } else {
3435                         unsigned long max_nr = (i_size >> PAGE_CACHE_SHIFT) + 1;
3436
3437                         set_range_writeback(tree, cur, cur + iosize - 1);
3438                         if (!PageWriteback(page)) {
3439                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3440                                            "page %lu not writeback, cur %llu end %llu",
3441                                        page->index, cur, end);
3442                         }
3443
3444                         ret = submit_extent_page(write_flags, tree, page,
3445                                                  sector, iosize, pg_offset,
3446                                                  bdev, &epd->bio, max_nr,
3447                                                  end_bio_extent_writepage,
3448                                                  0, 0, 0);
3449                         if (ret)
3450                                 SetPageError(page);
3451                 }
3452                 cur = cur + iosize;
3453                 pg_offset += iosize;
3454                 nr++;
3455         }
3456 done:
3457         *nr_ret = nr;
3458
3459 done_unlocked:
3460
3461         /* drop our reference on any cached states */
3462         free_extent_state(cached_state);
3463         return ret;
3464 }
3465
3466 /*
3467  * the writepage semantics are similar to regular writepage.  extent
3468  * records are inserted to lock ranges in the tree, and as dirty areas
3469  * are found, they are marked writeback.  Then the lock bits are removed
3470  * and the end_io handler clears the writeback ranges
3471  */
3472 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3473                               void *data)
3474 {
3475         struct inode *inode = page->mapping->host;
3476         struct extent_page_data *epd = data;
3477         u64 start = page_offset(page);
3478         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3479         int ret;
3480         int nr = 0;
3481         size_t pg_offset = 0;
3482         loff_t i_size = i_size_read(inode);
3483         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3484         int write_flags;
3485         unsigned long nr_written = 0;
3486
3487         if (wbc->sync_mode == WB_SYNC_ALL)
3488                 write_flags = WRITE_SYNC;
3489         else
3490                 write_flags = WRITE;
3491
3492         trace___extent_writepage(page, inode, wbc);
3493
3494         WARN_ON(!PageLocked(page));
3495
3496         ClearPageError(page);
3497
3498         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3499         if (page->index > end_index ||
3500            (page->index == end_index && !pg_offset)) {
3501                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3502                 unlock_page(page);
3503                 return 0;
3504         }
3505
3506         if (page->index == end_index) {
3507                 char *userpage;
3508
3509                 userpage = kmap_atomic(page);
3510                 memset(userpage + pg_offset, 0,
3511                        PAGE_CACHE_SIZE - pg_offset);
3512                 kunmap_atomic(userpage);
3513                 flush_dcache_page(page);
3514         }
3515
3516         pg_offset = 0;
3517
3518         set_page_extent_mapped(page);
3519
3520         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3521         if (ret == 1)
3522                 goto done_unlocked;
3523         if (ret)
3524                 goto done;
3525
3526         ret = __extent_writepage_io(inode, page, wbc, epd,
3527                                     i_size, nr_written, write_flags, &nr);
3528         if (ret == 1)
3529                 goto done_unlocked;
3530
3531 done:
3532         if (nr == 0) {
3533                 /* make sure the mapping tag for page dirty gets cleared */
3534                 set_page_writeback(page);
3535                 end_page_writeback(page);
3536         }
3537         if (PageError(page)) {
3538                 ret = ret < 0 ? ret : -EIO;
3539                 end_extent_writepage(page, ret, start, page_end);
3540         }
3541         unlock_page(page);
3542         return ret;
3543
3544 done_unlocked:
3545         return 0;
3546 }
3547
3548 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3549 {
3550         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3551                        TASK_UNINTERRUPTIBLE);
3552 }
3553
3554 static noinline_for_stack int
3555 lock_extent_buffer_for_io(struct extent_buffer *eb,
3556                           struct btrfs_fs_info *fs_info,
3557                           struct extent_page_data *epd)
3558 {
3559         unsigned long i, num_pages;
3560         int flush = 0;
3561         int ret = 0;
3562
3563         if (!btrfs_try_tree_write_lock(eb)) {
3564                 flush = 1;
3565                 flush_write_bio(epd);
3566                 btrfs_tree_lock(eb);
3567         }
3568
3569         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3570                 btrfs_tree_unlock(eb);
3571                 if (!epd->sync_io)
3572                         return 0;
3573                 if (!flush) {
3574                         flush_write_bio(epd);
3575                         flush = 1;
3576                 }
3577                 while (1) {
3578                         wait_on_extent_buffer_writeback(eb);
3579                         btrfs_tree_lock(eb);
3580                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3581                                 break;
3582                         btrfs_tree_unlock(eb);
3583                 }
3584         }
3585
3586         /*
3587          * We need to do this to prevent races in people who check if the eb is
3588          * under IO since we can end up having no IO bits set for a short period
3589          * of time.
3590          */
3591         spin_lock(&eb->refs_lock);
3592         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3593                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3594                 spin_unlock(&eb->refs_lock);
3595                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3596                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3597                                      -eb->len,
3598                                      fs_info->dirty_metadata_batch);
3599                 ret = 1;
3600         } else {
3601                 spin_unlock(&eb->refs_lock);
3602         }
3603
3604         btrfs_tree_unlock(eb);
3605
3606         if (!ret)
3607                 return ret;
3608
3609         num_pages = num_extent_pages(eb->start, eb->len);
3610         for (i = 0; i < num_pages; i++) {
3611                 struct page *p = eb->pages[i];
3612
3613                 if (!trylock_page(p)) {
3614                         if (!flush) {
3615                                 flush_write_bio(epd);
3616                                 flush = 1;
3617                         }
3618                         lock_page(p);
3619                 }
3620         }
3621
3622         return ret;
3623 }
3624
3625 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3626 {
3627         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3628         smp_mb__after_atomic();
3629         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3630 }
3631
3632 static void set_btree_ioerr(struct page *page)
3633 {
3634         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3635         struct btrfs_inode *btree_ino = BTRFS_I(eb->fs_info->btree_inode);
3636
3637         SetPageError(page);
3638         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3639                 return;
3640
3641         /*
3642          * If writeback for a btree extent that doesn't belong to a log tree
3643          * failed, increment the counter transaction->eb_write_errors.
3644          * We do this because while the transaction is running and before it's
3645          * committing (when we call filemap_fdata[write|wait]_range against
3646          * the btree inode), we might have
3647          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3648          * returns an error or an error happens during writeback, when we're
3649          * committing the transaction we wouldn't know about it, since the pages
3650          * can be no longer dirty nor marked anymore for writeback (if a
3651          * subsequent modification to the extent buffer didn't happen before the
3652          * transaction commit), which makes filemap_fdata[write|wait]_range not
3653          * able to find the pages tagged with SetPageError at transaction
3654          * commit time. So if this happens we must abort the transaction,
3655          * otherwise we commit a super block with btree roots that point to
3656          * btree nodes/leafs whose content on disk is invalid - either garbage
3657          * or the content of some node/leaf from a past generation that got
3658          * cowed or deleted and is no longer valid.
3659          *
3660          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3661          * not be enough - we need to distinguish between log tree extents vs
3662          * non-log tree extents, and the next filemap_fdatawait_range() call
3663          * will catch and clear such errors in the mapping - and that call might
3664          * be from a log sync and not from a transaction commit. Also, checking
3665          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3666          * not done and would not be reliable - the eb might have been released
3667          * from memory and reading it back again means that flag would not be
3668          * set (since it's a runtime flag, not persisted on disk).
3669          *
3670          * Using the flags below in the btree inode also makes us achieve the
3671          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3672          * writeback for all dirty pages and before filemap_fdatawait_range()
3673          * is called, the writeback for all dirty pages had already finished
3674          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3675          * filemap_fdatawait_range() would return success, as it could not know
3676          * that writeback errors happened (the pages were no longer tagged for
3677          * writeback).
3678          */
3679         switch (eb->log_index) {
3680         case -1:
3681                 set_bit(BTRFS_INODE_BTREE_ERR, &btree_ino->runtime_flags);
3682                 break;
3683         case 0:
3684                 set_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
3685                 break;
3686         case 1:
3687                 set_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
3688                 break;
3689         default:
3690                 BUG(); /* unexpected, logic error */
3691         }
3692 }
3693
3694 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3695 {
3696         struct bio_vec *bvec;
3697         struct extent_buffer *eb;
3698         int i, done;
3699
3700         bio_for_each_segment_all(bvec, bio, i) {
3701                 struct page *page = bvec->bv_page;
3702
3703                 eb = (struct extent_buffer *)page->private;
3704                 BUG_ON(!eb);
3705                 done = atomic_dec_and_test(&eb->io_pages);
3706
3707                 if (err || test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3708                         ClearPageUptodate(page);
3709                         set_btree_ioerr(page);
3710                 }
3711
3712                 end_page_writeback(page);
3713
3714                 if (!done)
3715                         continue;
3716
3717                 end_extent_buffer_writeback(eb);
3718         }
3719
3720         bio_put(bio);
3721 }
3722
3723 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3724                         struct btrfs_fs_info *fs_info,
3725                         struct writeback_control *wbc,
3726                         struct extent_page_data *epd)
3727 {
3728         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3729         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3730         u64 offset = eb->start;
3731         unsigned long i, num_pages;
3732         unsigned long bio_flags = 0;
3733         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3734         int ret = 0;
3735
3736         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3737         num_pages = num_extent_pages(eb->start, eb->len);
3738         atomic_set(&eb->io_pages, num_pages);
3739         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3740                 bio_flags = EXTENT_BIO_TREE_LOG;
3741
3742         for (i = 0; i < num_pages; i++) {
3743                 struct page *p = eb->pages[i];
3744
3745                 clear_page_dirty_for_io(p);
3746                 set_page_writeback(p);
3747                 ret = submit_extent_page(rw, tree, p, offset >> 9,
3748                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3749                                          -1, end_bio_extent_buffer_writepage,
3750                                          0, epd->bio_flags, bio_flags);
3751                 epd->bio_flags = bio_flags;
3752                 if (ret) {
3753                         set_btree_ioerr(p);
3754                         end_page_writeback(p);
3755                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3756                                 end_extent_buffer_writeback(eb);
3757                         ret = -EIO;
3758                         break;
3759                 }
3760                 offset += PAGE_CACHE_SIZE;
3761                 update_nr_written(p, wbc, 1);
3762                 unlock_page(p);
3763         }
3764
3765         if (unlikely(ret)) {
3766                 for (; i < num_pages; i++) {
3767                         struct page *p = eb->pages[i];
3768                         clear_page_dirty_for_io(p);
3769                         unlock_page(p);
3770                 }
3771         }
3772
3773         return ret;
3774 }
3775
3776 int btree_write_cache_pages(struct address_space *mapping,
3777                                    struct writeback_control *wbc)
3778 {
3779         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3780         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3781         struct extent_buffer *eb, *prev_eb = NULL;
3782         struct extent_page_data epd = {
3783                 .bio = NULL,
3784                 .tree = tree,
3785                 .extent_locked = 0,
3786                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3787                 .bio_flags = 0,
3788         };
3789         int ret = 0;
3790         int done = 0;
3791         int nr_to_write_done = 0;
3792         struct pagevec pvec;
3793         int nr_pages;
3794         pgoff_t index;
3795         pgoff_t end;            /* Inclusive */
3796         int scanned = 0;
3797         int tag;
3798
3799         pagevec_init(&pvec, 0);
3800         if (wbc->range_cyclic) {
3801                 index = mapping->writeback_index; /* Start from prev offset */
3802                 end = -1;
3803         } else {
3804                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3805                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3806                 scanned = 1;
3807         }
3808         if (wbc->sync_mode == WB_SYNC_ALL)
3809                 tag = PAGECACHE_TAG_TOWRITE;
3810         else
3811                 tag = PAGECACHE_TAG_DIRTY;
3812 retry:
3813         if (wbc->sync_mode == WB_SYNC_ALL)
3814                 tag_pages_for_writeback(mapping, index, end);
3815         while (!done && !nr_to_write_done && (index <= end) &&
3816                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3817                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3818                 unsigned i;
3819
3820                 scanned = 1;
3821                 for (i = 0; i < nr_pages; i++) {
3822                         struct page *page = pvec.pages[i];
3823
3824                         if (!PagePrivate(page))
3825                                 continue;
3826
3827                         if (!wbc->range_cyclic && page->index > end) {
3828                                 done = 1;
3829                                 break;
3830                         }
3831
3832                         spin_lock(&mapping->private_lock);
3833                         if (!PagePrivate(page)) {
3834                                 spin_unlock(&mapping->private_lock);
3835                                 continue;
3836                         }
3837
3838                         eb = (struct extent_buffer *)page->private;
3839
3840                         /*
3841                          * Shouldn't happen and normally this would be a BUG_ON
3842                          * but no sense in crashing the users box for something
3843                          * we can survive anyway.
3844                          */
3845                         if (WARN_ON(!eb)) {
3846                                 spin_unlock(&mapping->private_lock);
3847                                 continue;
3848                         }
3849
3850                         if (eb == prev_eb) {
3851                                 spin_unlock(&mapping->private_lock);
3852                                 continue;
3853                         }
3854
3855                         ret = atomic_inc_not_zero(&eb->refs);
3856                         spin_unlock(&mapping->private_lock);
3857                         if (!ret)
3858                                 continue;
3859
3860                         prev_eb = eb;
3861                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3862                         if (!ret) {
3863                                 free_extent_buffer(eb);
3864                                 continue;
3865                         }
3866
3867                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3868                         if (ret) {
3869                                 done = 1;
3870                                 free_extent_buffer(eb);
3871                                 break;
3872                         }
3873                         free_extent_buffer(eb);
3874
3875                         /*
3876                          * the filesystem may choose to bump up nr_to_write.
3877                          * We have to make sure to honor the new nr_to_write
3878                          * at any time
3879                          */
3880                         nr_to_write_done = wbc->nr_to_write <= 0;
3881                 }
3882                 pagevec_release(&pvec);
3883                 cond_resched();
3884         }
3885         if (!scanned && !done) {
3886                 /*
3887                  * We hit the last page and there is more work to be done: wrap
3888                  * back to the start of the file
3889                  */
3890                 scanned = 1;
3891                 index = 0;
3892                 goto retry;
3893         }
3894         flush_write_bio(&epd);
3895         return ret;
3896 }
3897
3898 /**
3899  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3900  * @mapping: address space structure to write
3901  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3902  * @writepage: function called for each page
3903  * @data: data passed to writepage function
3904  *
3905  * If a page is already under I/O, write_cache_pages() skips it, even
3906  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3907  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3908  * and msync() need to guarantee that all the data which was dirty at the time
3909  * the call was made get new I/O started against them.  If wbc->sync_mode is
3910  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3911  * existing IO to complete.
3912  */
3913 static int extent_write_cache_pages(struct extent_io_tree *tree,
3914                              struct address_space *mapping,
3915                              struct writeback_control *wbc,
3916                              writepage_t writepage, void *data,
3917                              void (*flush_fn)(void *))
3918 {
3919         struct inode *inode = mapping->host;
3920         int ret = 0;
3921         int done = 0;
3922         int err = 0;
3923         int nr_to_write_done = 0;
3924         struct pagevec pvec;
3925         int nr_pages;
3926         pgoff_t index;
3927         pgoff_t end;            /* Inclusive */
3928         int scanned = 0;
3929         int tag;
3930
3931         /*
3932          * We have to hold onto the inode so that ordered extents can do their
3933          * work when the IO finishes.  The alternative to this is failing to add
3934          * an ordered extent if the igrab() fails there and that is a huge pain
3935          * to deal with, so instead just hold onto the inode throughout the
3936          * writepages operation.  If it fails here we are freeing up the inode
3937          * anyway and we'd rather not waste our time writing out stuff that is
3938          * going to be truncated anyway.
3939          */
3940         if (!igrab(inode))
3941                 return 0;
3942
3943         pagevec_init(&pvec, 0);
3944         if (wbc->range_cyclic) {
3945                 index = mapping->writeback_index; /* Start from prev offset */
3946                 end = -1;
3947         } else {
3948                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3949                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3950                 scanned = 1;
3951         }
3952         if (wbc->sync_mode == WB_SYNC_ALL)
3953                 tag = PAGECACHE_TAG_TOWRITE;
3954         else
3955                 tag = PAGECACHE_TAG_DIRTY;
3956 retry:
3957         if (wbc->sync_mode == WB_SYNC_ALL)
3958                 tag_pages_for_writeback(mapping, index, end);
3959         while (!done && !nr_to_write_done && (index <= end) &&
3960                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3961                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3962                 unsigned i;
3963
3964                 scanned = 1;
3965                 for (i = 0; i < nr_pages; i++) {
3966                         struct page *page = pvec.pages[i];
3967
3968                         /*
3969                          * At this point we hold neither mapping->tree_lock nor
3970                          * lock on the page itself: the page may be truncated or
3971                          * invalidated (changing page->mapping to NULL), or even
3972                          * swizzled back from swapper_space to tmpfs file
3973                          * mapping
3974                          */
3975                         if (!trylock_page(page)) {
3976                                 flush_fn(data);
3977                                 lock_page(page);
3978                         }
3979
3980                         if (unlikely(page->mapping != mapping)) {
3981                                 unlock_page(page);
3982                                 continue;
3983                         }
3984
3985                         if (!wbc->range_cyclic && page->index > end) {
3986                                 done = 1;
3987                                 unlock_page(page);
3988                                 continue;
3989                         }
3990
3991                         if (wbc->sync_mode != WB_SYNC_NONE) {
3992                                 if (PageWriteback(page))
3993                                         flush_fn(data);
3994                                 wait_on_page_writeback(page);
3995                         }
3996
3997                         if (PageWriteback(page) ||
3998                             !clear_page_dirty_for_io(page)) {
3999                                 unlock_page(page);
4000                                 continue;
4001                         }
4002
4003                         ret = (*writepage)(page, wbc, data);
4004
4005                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4006                                 unlock_page(page);
4007                                 ret = 0;
4008                         }
4009                         if (!err && ret < 0)
4010                                 err = ret;
4011
4012                         /*
4013                          * the filesystem may choose to bump up nr_to_write.
4014                          * We have to make sure to honor the new nr_to_write
4015                          * at any time
4016                          */
4017                         nr_to_write_done = wbc->nr_to_write <= 0;
4018                 }
4019                 pagevec_release(&pvec);
4020                 cond_resched();
4021         }
4022         if (!scanned && !done && !err) {
4023                 /*
4024                  * We hit the last page and there is more work to be done: wrap
4025                  * back to the start of the file
4026                  */
4027                 scanned = 1;
4028                 index = 0;
4029                 goto retry;
4030         }
4031         btrfs_add_delayed_iput(inode);
4032         return err;
4033 }
4034
4035 static void flush_epd_write_bio(struct extent_page_data *epd)
4036 {
4037         if (epd->bio) {
4038                 int rw = WRITE;
4039                 int ret;
4040
4041                 if (epd->sync_io)
4042                         rw = WRITE_SYNC;
4043
4044                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
4045                 BUG_ON(ret < 0); /* -ENOMEM */
4046                 epd->bio = NULL;
4047         }
4048 }
4049
4050 static noinline void flush_write_bio(void *data)
4051 {
4052         struct extent_page_data *epd = data;
4053         flush_epd_write_bio(epd);
4054 }
4055
4056 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4057                           get_extent_t *get_extent,
4058                           struct writeback_control *wbc)
4059 {
4060         int ret;
4061         struct extent_page_data epd = {
4062                 .bio = NULL,
4063                 .tree = tree,
4064                 .get_extent = get_extent,
4065                 .extent_locked = 0,
4066                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4067                 .bio_flags = 0,
4068         };
4069
4070         ret = __extent_writepage(page, wbc, &epd);
4071
4072         flush_epd_write_bio(&epd);
4073         return ret;
4074 }
4075
4076 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4077                               u64 start, u64 end, get_extent_t *get_extent,
4078                               int mode)
4079 {
4080         int ret = 0;
4081         struct address_space *mapping = inode->i_mapping;
4082         struct page *page;
4083         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
4084                 PAGE_CACHE_SHIFT;
4085
4086         struct extent_page_data epd = {
4087                 .bio = NULL,
4088                 .tree = tree,
4089                 .get_extent = get_extent,
4090                 .extent_locked = 1,
4091                 .sync_io = mode == WB_SYNC_ALL,
4092                 .bio_flags = 0,
4093         };
4094         struct writeback_control wbc_writepages = {
4095                 .sync_mode      = mode,
4096                 .nr_to_write    = nr_pages * 2,
4097                 .range_start    = start,
4098                 .range_end      = end + 1,
4099         };
4100
4101         while (start <= end) {
4102                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
4103                 if (clear_page_dirty_for_io(page))
4104                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4105                 else {
4106                         if (tree->ops && tree->ops->writepage_end_io_hook)
4107                                 tree->ops->writepage_end_io_hook(page, start,
4108                                                  start + PAGE_CACHE_SIZE - 1,
4109                                                  NULL, 1);
4110                         unlock_page(page);
4111                 }
4112                 page_cache_release(page);
4113                 start += PAGE_CACHE_SIZE;
4114         }
4115
4116         flush_epd_write_bio(&epd);
4117         return ret;
4118 }
4119
4120 int extent_writepages(struct extent_io_tree *tree,
4121                       struct address_space *mapping,
4122                       get_extent_t *get_extent,
4123                       struct writeback_control *wbc)
4124 {
4125         int ret = 0;
4126         struct extent_page_data epd = {
4127                 .bio = NULL,
4128                 .tree = tree,
4129                 .get_extent = get_extent,
4130                 .extent_locked = 0,
4131                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4132                 .bio_flags = 0,
4133         };
4134
4135         ret = extent_write_cache_pages(tree, mapping, wbc,
4136                                        __extent_writepage, &epd,
4137                                        flush_write_bio);
4138         flush_epd_write_bio(&epd);
4139         return ret;
4140 }
4141
4142 int extent_readpages(struct extent_io_tree *tree,
4143                      struct address_space *mapping,
4144                      struct list_head *pages, unsigned nr_pages,
4145                      get_extent_t get_extent)
4146 {
4147         struct bio *bio = NULL;
4148         unsigned page_idx;
4149         unsigned long bio_flags = 0;
4150         struct page *pagepool[16];
4151         struct page *page;
4152         struct extent_map *em_cached = NULL;
4153         int nr = 0;
4154
4155         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4156                 page = list_entry(pages->prev, struct page, lru);
4157
4158                 prefetchw(&page->flags);
4159                 list_del(&page->lru);
4160                 if (add_to_page_cache_lru(page, mapping,
4161                                         page->index, GFP_NOFS)) {
4162                         page_cache_release(page);
4163                         continue;
4164                 }
4165
4166                 pagepool[nr++] = page;
4167                 if (nr < ARRAY_SIZE(pagepool))
4168                         continue;
4169                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4170                                    &bio, 0, &bio_flags, READ);
4171                 nr = 0;
4172         }
4173         if (nr)
4174                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4175                                    &bio, 0, &bio_flags, READ);
4176
4177         if (em_cached)
4178                 free_extent_map(em_cached);
4179
4180         BUG_ON(!list_empty(pages));
4181         if (bio)
4182                 return submit_one_bio(READ, bio, 0, bio_flags);
4183         return 0;
4184 }
4185
4186 /*
4187  * basic invalidatepage code, this waits on any locked or writeback
4188  * ranges corresponding to the page, and then deletes any extent state
4189  * records from the tree
4190  */
4191 int extent_invalidatepage(struct extent_io_tree *tree,
4192                           struct page *page, unsigned long offset)
4193 {
4194         struct extent_state *cached_state = NULL;
4195         u64 start = page_offset(page);
4196         u64 end = start + PAGE_CACHE_SIZE - 1;
4197         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4198
4199         start += ALIGN(offset, blocksize);
4200         if (start > end)
4201                 return 0;
4202
4203         lock_extent_bits(tree, start, end, 0, &cached_state);
4204         wait_on_page_writeback(page);
4205         clear_extent_bit(tree, start, end,
4206                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4207                          EXTENT_DO_ACCOUNTING,
4208                          1, 1, &cached_state, GFP_NOFS);
4209         return 0;
4210 }
4211
4212 /*
4213  * a helper for releasepage, this tests for areas of the page that
4214  * are locked or under IO and drops the related state bits if it is safe
4215  * to drop the page.
4216  */
4217 static int try_release_extent_state(struct extent_map_tree *map,
4218                                     struct extent_io_tree *tree,
4219                                     struct page *page, gfp_t mask)
4220 {
4221         u64 start = page_offset(page);
4222         u64 end = start + PAGE_CACHE_SIZE - 1;
4223         int ret = 1;
4224
4225         if (test_range_bit(tree, start, end,
4226                            EXTENT_IOBITS, 0, NULL))
4227                 ret = 0;
4228         else {
4229                 if ((mask & GFP_NOFS) == GFP_NOFS)
4230                         mask = GFP_NOFS;
4231                 /*
4232                  * at this point we can safely clear everything except the
4233                  * locked bit and the nodatasum bit
4234                  */
4235                 ret = clear_extent_bit(tree, start, end,
4236                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4237                                  0, 0, NULL, mask);
4238
4239                 /* if clear_extent_bit failed for enomem reasons,
4240                  * we can't allow the release to continue.
4241                  */
4242                 if (ret < 0)
4243                         ret = 0;
4244                 else
4245                         ret = 1;
4246         }
4247         return ret;
4248 }
4249
4250 /*
4251  * a helper for releasepage.  As long as there are no locked extents
4252  * in the range corresponding to the page, both state records and extent
4253  * map records are removed
4254  */
4255 int try_release_extent_mapping(struct extent_map_tree *map,
4256                                struct extent_io_tree *tree, struct page *page,
4257                                gfp_t mask)
4258 {
4259         struct extent_map *em;
4260         u64 start = page_offset(page);
4261         u64 end = start + PAGE_CACHE_SIZE - 1;
4262
4263         if ((mask & __GFP_WAIT) &&
4264             page->mapping->host->i_size > 16 * 1024 * 1024) {
4265                 u64 len;
4266                 while (start <= end) {
4267                         len = end - start + 1;
4268                         write_lock(&map->lock);
4269                         em = lookup_extent_mapping(map, start, len);
4270                         if (!em) {
4271                                 write_unlock(&map->lock);
4272                                 break;
4273                         }
4274                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4275                             em->start != start) {
4276                                 write_unlock(&map->lock);
4277                                 free_extent_map(em);
4278                                 break;
4279                         }
4280                         if (!test_range_bit(tree, em->start,
4281                                             extent_map_end(em) - 1,
4282                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4283                                             0, NULL)) {
4284                                 remove_extent_mapping(map, em);
4285                                 /* once for the rb tree */
4286                                 free_extent_map(em);
4287                         }
4288                         start = extent_map_end(em);
4289                         write_unlock(&map->lock);
4290
4291                         /* once for us */
4292                         free_extent_map(em);
4293                 }
4294         }
4295         return try_release_extent_state(map, tree, page, mask);
4296 }
4297
4298 /*
4299  * helper function for fiemap, which doesn't want to see any holes.
4300  * This maps until we find something past 'last'
4301  */
4302 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4303                                                 u64 offset,
4304                                                 u64 last,
4305                                                 get_extent_t *get_extent)
4306 {
4307         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4308         struct extent_map *em;
4309         u64 len;
4310
4311         if (offset >= last)
4312                 return NULL;
4313
4314         while (1) {
4315                 len = last - offset;
4316                 if (len == 0)
4317                         break;
4318                 len = ALIGN(len, sectorsize);
4319                 em = get_extent(inode, NULL, 0, offset, len, 0);
4320                 if (IS_ERR_OR_NULL(em))
4321                         return em;
4322
4323                 /* if this isn't a hole return it */
4324                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4325                     em->block_start != EXTENT_MAP_HOLE) {
4326                         return em;
4327                 }
4328
4329                 /* this is a hole, advance to the next extent */
4330                 offset = extent_map_end(em);
4331                 free_extent_map(em);
4332                 if (offset >= last)
4333                         break;
4334         }
4335         return NULL;
4336 }
4337
4338 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4339                 __u64 start, __u64 len, get_extent_t *get_extent)
4340 {
4341         int ret = 0;
4342         u64 off = start;
4343         u64 max = start + len;
4344         u32 flags = 0;
4345         u32 found_type;
4346         u64 last;
4347         u64 last_for_get_extent = 0;
4348         u64 disko = 0;
4349         u64 isize = i_size_read(inode);
4350         struct btrfs_key found_key;
4351         struct extent_map *em = NULL;
4352         struct extent_state *cached_state = NULL;
4353         struct btrfs_path *path;
4354         struct btrfs_root *root = BTRFS_I(inode)->root;
4355         int end = 0;
4356         u64 em_start = 0;
4357         u64 em_len = 0;
4358         u64 em_end = 0;
4359
4360         if (len == 0)
4361                 return -EINVAL;
4362
4363         path = btrfs_alloc_path();
4364         if (!path)
4365                 return -ENOMEM;
4366         path->leave_spinning = 1;
4367
4368         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4369         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4370
4371         /*
4372          * lookup the last file extent.  We're not using i_size here
4373          * because there might be preallocation past i_size
4374          */
4375         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
4376                                        0);
4377         if (ret < 0) {
4378                 btrfs_free_path(path);
4379                 return ret;
4380         }
4381         WARN_ON(!ret);
4382         path->slots[0]--;
4383         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4384         found_type = found_key.type;
4385
4386         /* No extents, but there might be delalloc bits */
4387         if (found_key.objectid != btrfs_ino(inode) ||
4388             found_type != BTRFS_EXTENT_DATA_KEY) {
4389                 /* have to trust i_size as the end */
4390                 last = (u64)-1;
4391                 last_for_get_extent = isize;
4392         } else {
4393                 /*
4394                  * remember the start of the last extent.  There are a
4395                  * bunch of different factors that go into the length of the
4396                  * extent, so its much less complex to remember where it started
4397                  */
4398                 last = found_key.offset;
4399                 last_for_get_extent = last + 1;
4400         }
4401         btrfs_release_path(path);
4402
4403         /*
4404          * we might have some extents allocated but more delalloc past those
4405          * extents.  so, we trust isize unless the start of the last extent is
4406          * beyond isize
4407          */
4408         if (last < isize) {
4409                 last = (u64)-1;
4410                 last_for_get_extent = isize;
4411         }
4412
4413         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4414                          &cached_state);
4415
4416         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4417                                    get_extent);
4418         if (!em)
4419                 goto out;
4420         if (IS_ERR(em)) {
4421                 ret = PTR_ERR(em);
4422                 goto out;
4423         }
4424
4425         while (!end) {
4426                 u64 offset_in_extent = 0;
4427
4428                 /* break if the extent we found is outside the range */
4429                 if (em->start >= max || extent_map_end(em) < off)
4430                         break;
4431
4432                 /*
4433                  * get_extent may return an extent that starts before our
4434                  * requested range.  We have to make sure the ranges
4435                  * we return to fiemap always move forward and don't
4436                  * overlap, so adjust the offsets here
4437                  */
4438                 em_start = max(em->start, off);
4439
4440                 /*
4441                  * record the offset from the start of the extent
4442                  * for adjusting the disk offset below.  Only do this if the
4443                  * extent isn't compressed since our in ram offset may be past
4444                  * what we have actually allocated on disk.
4445                  */
4446                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4447                         offset_in_extent = em_start - em->start;
4448                 em_end = extent_map_end(em);
4449                 em_len = em_end - em_start;
4450                 disko = 0;
4451                 flags = 0;
4452
4453                 /*
4454                  * bump off for our next call to get_extent
4455                  */
4456                 off = extent_map_end(em);
4457                 if (off >= max)
4458                         end = 1;
4459
4460                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4461                         end = 1;
4462                         flags |= FIEMAP_EXTENT_LAST;
4463                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4464                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4465                                   FIEMAP_EXTENT_NOT_ALIGNED);
4466                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4467                         flags |= (FIEMAP_EXTENT_DELALLOC |
4468                                   FIEMAP_EXTENT_UNKNOWN);
4469                 } else if (fieinfo->fi_extents_max) {
4470                         u64 bytenr = em->block_start -
4471                                 (em->start - em->orig_start);
4472
4473                         disko = em->block_start + offset_in_extent;
4474
4475                         /*
4476                          * As btrfs supports shared space, this information
4477                          * can be exported to userspace tools via
4478                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4479                          * then we're just getting a count and we can skip the
4480                          * lookup stuff.
4481                          */
4482                         ret = btrfs_check_shared(NULL, root->fs_info,
4483                                                  root->objectid,
4484                                                  btrfs_ino(inode), bytenr);
4485                         if (ret < 0)
4486                                 goto out_free;
4487                         if (ret)
4488                                 flags |= FIEMAP_EXTENT_SHARED;
4489                         ret = 0;
4490                 }
4491                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4492                         flags |= FIEMAP_EXTENT_ENCODED;
4493
4494                 free_extent_map(em);
4495                 em = NULL;
4496                 if ((em_start >= last) || em_len == (u64)-1 ||
4497                    (last == (u64)-1 && isize <= em_end)) {
4498                         flags |= FIEMAP_EXTENT_LAST;
4499                         end = 1;
4500                 }
4501
4502                 /* now scan forward to see if this is really the last extent. */
4503                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4504                                            get_extent);
4505                 if (IS_ERR(em)) {
4506                         ret = PTR_ERR(em);
4507                         goto out;
4508                 }
4509                 if (!em) {
4510                         flags |= FIEMAP_EXTENT_LAST;
4511                         end = 1;
4512                 }
4513                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4514                                               em_len, flags);
4515                 if (ret) {
4516                         if (ret == 1)
4517                                 ret = 0;
4518                         goto out_free;
4519                 }
4520         }
4521 out_free:
4522         free_extent_map(em);
4523 out:
4524         btrfs_free_path(path);
4525         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4526                              &cached_state, GFP_NOFS);
4527         return ret;
4528 }
4529
4530 static void __free_extent_buffer(struct extent_buffer *eb)
4531 {
4532         btrfs_leak_debug_del(&eb->leak_list);
4533         kmem_cache_free(extent_buffer_cache, eb);
4534 }
4535
4536 int extent_buffer_under_io(struct extent_buffer *eb)
4537 {
4538         return (atomic_read(&eb->io_pages) ||
4539                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4540                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4541 }
4542
4543 /*
4544  * Helper for releasing extent buffer page.
4545  */
4546 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4547 {
4548         unsigned long index;
4549         struct page *page;
4550         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4551
4552         BUG_ON(extent_buffer_under_io(eb));
4553
4554         index = num_extent_pages(eb->start, eb->len);
4555         if (index == 0)
4556                 return;
4557
4558         do {
4559                 index--;
4560                 page = eb->pages[index];
4561                 if (!page)
4562                         continue;
4563                 if (mapped)
4564                         spin_lock(&page->mapping->private_lock);
4565                 /*
4566                  * We do this since we'll remove the pages after we've
4567                  * removed the eb from the radix tree, so we could race
4568                  * and have this page now attached to the new eb.  So
4569                  * only clear page_private if it's still connected to
4570                  * this eb.
4571                  */
4572                 if (PagePrivate(page) &&
4573                     page->private == (unsigned long)eb) {
4574                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4575                         BUG_ON(PageDirty(page));
4576                         BUG_ON(PageWriteback(page));
4577                         /*
4578                          * We need to make sure we haven't be attached
4579                          * to a new eb.
4580                          */
4581                         ClearPagePrivate(page);
4582                         set_page_private(page, 0);
4583                         /* One for the page private */
4584                         page_cache_release(page);
4585                 }
4586
4587                 if (mapped)
4588                         spin_unlock(&page->mapping->private_lock);
4589
4590                 /* One for when we alloced the page */
4591                 page_cache_release(page);
4592         } while (index != 0);
4593 }
4594
4595 /*
4596  * Helper for releasing the extent buffer.
4597  */
4598 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4599 {
4600         btrfs_release_extent_buffer_page(eb);
4601         __free_extent_buffer(eb);
4602 }
4603
4604 static struct extent_buffer *
4605 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4606                       unsigned long len)
4607 {
4608         struct extent_buffer *eb = NULL;
4609
4610         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS);
4611         if (eb == NULL)
4612                 return NULL;
4613         eb->start = start;
4614         eb->len = len;
4615         eb->fs_info = fs_info;
4616         eb->bflags = 0;
4617         rwlock_init(&eb->lock);
4618         atomic_set(&eb->write_locks, 0);
4619         atomic_set(&eb->read_locks, 0);
4620         atomic_set(&eb->blocking_readers, 0);
4621         atomic_set(&eb->blocking_writers, 0);
4622         atomic_set(&eb->spinning_readers, 0);
4623         atomic_set(&eb->spinning_writers, 0);
4624         eb->lock_nested = 0;
4625         init_waitqueue_head(&eb->write_lock_wq);
4626         init_waitqueue_head(&eb->read_lock_wq);
4627
4628         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4629
4630         spin_lock_init(&eb->refs_lock);
4631         atomic_set(&eb->refs, 1);
4632         atomic_set(&eb->io_pages, 0);
4633
4634         /*
4635          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4636          */
4637         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4638                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4639         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4640
4641         return eb;
4642 }
4643
4644 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4645 {
4646         unsigned long i;
4647         struct page *p;
4648         struct extent_buffer *new;
4649         unsigned long num_pages = num_extent_pages(src->start, src->len);
4650
4651         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4652         if (new == NULL)
4653                 return NULL;
4654
4655         for (i = 0; i < num_pages; i++) {
4656                 p = alloc_page(GFP_NOFS);
4657                 if (!p) {
4658                         btrfs_release_extent_buffer(new);
4659                         return NULL;
4660                 }
4661                 attach_extent_buffer_page(new, p);
4662                 WARN_ON(PageDirty(p));
4663                 SetPageUptodate(p);
4664                 new->pages[i] = p;
4665         }
4666
4667         copy_extent_buffer(new, src, 0, 0, src->len);
4668         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4669         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4670
4671         return new;
4672 }
4673
4674 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4675                                                 u64 start)
4676 {
4677         struct extent_buffer *eb;
4678         unsigned long len;
4679         unsigned long num_pages;
4680         unsigned long i;
4681
4682         if (!fs_info) {
4683                 /*
4684                  * Called only from tests that don't always have a fs_info
4685                  * available, but we know that nodesize is 4096
4686                  */
4687                 len = 4096;
4688         } else {
4689                 len = fs_info->tree_root->nodesize;
4690         }
4691         num_pages = num_extent_pages(0, len);
4692
4693         eb = __alloc_extent_buffer(fs_info, start, len);
4694         if (!eb)
4695                 return NULL;
4696
4697         for (i = 0; i < num_pages; i++) {
4698                 eb->pages[i] = alloc_page(GFP_NOFS);
4699                 if (!eb->pages[i])
4700                         goto err;
4701         }
4702         set_extent_buffer_uptodate(eb);
4703         btrfs_set_header_nritems(eb, 0);
4704         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4705
4706         return eb;
4707 err:
4708         for (; i > 0; i--)
4709                 __free_page(eb->pages[i - 1]);
4710         __free_extent_buffer(eb);
4711         return NULL;
4712 }
4713
4714 static void check_buffer_tree_ref(struct extent_buffer *eb)
4715 {
4716         int refs;
4717         /* the ref bit is tricky.  We have to make sure it is set
4718          * if we have the buffer dirty.   Otherwise the
4719          * code to free a buffer can end up dropping a dirty
4720          * page
4721          *
4722          * Once the ref bit is set, it won't go away while the
4723          * buffer is dirty or in writeback, and it also won't
4724          * go away while we have the reference count on the
4725          * eb bumped.
4726          *
4727          * We can't just set the ref bit without bumping the
4728          * ref on the eb because free_extent_buffer might
4729          * see the ref bit and try to clear it.  If this happens
4730          * free_extent_buffer might end up dropping our original
4731          * ref by mistake and freeing the page before we are able
4732          * to add one more ref.
4733          *
4734          * So bump the ref count first, then set the bit.  If someone
4735          * beat us to it, drop the ref we added.
4736          */
4737         refs = atomic_read(&eb->refs);
4738         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4739                 return;
4740
4741         spin_lock(&eb->refs_lock);
4742         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4743                 atomic_inc(&eb->refs);
4744         spin_unlock(&eb->refs_lock);
4745 }
4746
4747 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4748                 struct page *accessed)
4749 {
4750         unsigned long num_pages, i;
4751
4752         check_buffer_tree_ref(eb);
4753
4754         num_pages = num_extent_pages(eb->start, eb->len);
4755         for (i = 0; i < num_pages; i++) {
4756                 struct page *p = eb->pages[i];
4757
4758                 if (p != accessed)
4759                         mark_page_accessed(p);
4760         }
4761 }
4762
4763 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4764                                          u64 start)
4765 {
4766         struct extent_buffer *eb;
4767
4768         rcu_read_lock();
4769         eb = radix_tree_lookup(&fs_info->buffer_radix,
4770                                start >> PAGE_CACHE_SHIFT);
4771         if (eb && atomic_inc_not_zero(&eb->refs)) {
4772                 rcu_read_unlock();
4773                 /*
4774                  * Lock our eb's refs_lock to avoid races with
4775                  * free_extent_buffer. When we get our eb it might be flagged
4776                  * with EXTENT_BUFFER_STALE and another task running
4777                  * free_extent_buffer might have seen that flag set,
4778                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4779                  * writeback flags not set) and it's still in the tree (flag
4780                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4781                  * of decrementing the extent buffer's reference count twice.
4782                  * So here we could race and increment the eb's reference count,
4783                  * clear its stale flag, mark it as dirty and drop our reference
4784                  * before the other task finishes executing free_extent_buffer,
4785                  * which would later result in an attempt to free an extent
4786                  * buffer that is dirty.
4787                  */
4788                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4789                         spin_lock(&eb->refs_lock);
4790                         spin_unlock(&eb->refs_lock);
4791                 }
4792                 mark_extent_buffer_accessed(eb, NULL);
4793                 return eb;
4794         }
4795         rcu_read_unlock();
4796
4797         return NULL;
4798 }
4799
4800 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4801 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4802                                                u64 start)
4803 {
4804         struct extent_buffer *eb, *exists = NULL;
4805         int ret;
4806
4807         eb = find_extent_buffer(fs_info, start);
4808         if (eb)
4809                 return eb;
4810         eb = alloc_dummy_extent_buffer(fs_info, start);
4811         if (!eb)
4812                 return NULL;
4813         eb->fs_info = fs_info;
4814 again:
4815         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4816         if (ret)
4817                 goto free_eb;
4818         spin_lock(&fs_info->buffer_lock);
4819         ret = radix_tree_insert(&fs_info->buffer_radix,
4820                                 start >> PAGE_CACHE_SHIFT, eb);
4821         spin_unlock(&fs_info->buffer_lock);
4822         radix_tree_preload_end();
4823         if (ret == -EEXIST) {
4824                 exists = find_extent_buffer(fs_info, start);
4825                 if (exists)
4826                         goto free_eb;
4827                 else
4828                         goto again;
4829         }
4830         check_buffer_tree_ref(eb);
4831         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4832
4833         /*
4834          * We will free dummy extent buffer's if they come into
4835          * free_extent_buffer with a ref count of 2, but if we are using this we
4836          * want the buffers to stay in memory until we're done with them, so
4837          * bump the ref count again.
4838          */
4839         atomic_inc(&eb->refs);
4840         return eb;
4841 free_eb:
4842         btrfs_release_extent_buffer(eb);
4843         return exists;
4844 }
4845 #endif
4846
4847 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4848                                           u64 start)
4849 {
4850         unsigned long len = fs_info->tree_root->nodesize;
4851         unsigned long num_pages = num_extent_pages(start, len);
4852         unsigned long i;
4853         unsigned long index = start >> PAGE_CACHE_SHIFT;
4854         struct extent_buffer *eb;
4855         struct extent_buffer *exists = NULL;
4856         struct page *p;
4857         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4858         int uptodate = 1;
4859         int ret;
4860
4861         eb = find_extent_buffer(fs_info, start);
4862         if (eb)
4863                 return eb;
4864
4865         eb = __alloc_extent_buffer(fs_info, start, len);
4866         if (!eb)
4867                 return NULL;
4868
4869         for (i = 0; i < num_pages; i++, index++) {
4870                 p = find_or_create_page(mapping, index, GFP_NOFS);
4871                 if (!p)
4872                         goto free_eb;
4873
4874                 spin_lock(&mapping->private_lock);
4875                 if (PagePrivate(p)) {
4876                         /*
4877                          * We could have already allocated an eb for this page
4878                          * and attached one so lets see if we can get a ref on
4879                          * the existing eb, and if we can we know it's good and
4880                          * we can just return that one, else we know we can just
4881                          * overwrite page->private.
4882                          */
4883                         exists = (struct extent_buffer *)p->private;
4884                         if (atomic_inc_not_zero(&exists->refs)) {
4885                                 spin_unlock(&mapping->private_lock);
4886                                 unlock_page(p);
4887                                 page_cache_release(p);
4888                                 mark_extent_buffer_accessed(exists, p);
4889                                 goto free_eb;
4890                         }
4891                         exists = NULL;
4892
4893                         /*
4894                          * Do this so attach doesn't complain and we need to
4895                          * drop the ref the old guy had.
4896                          */
4897                         ClearPagePrivate(p);
4898                         WARN_ON(PageDirty(p));
4899                         page_cache_release(p);
4900                 }
4901                 attach_extent_buffer_page(eb, p);
4902                 spin_unlock(&mapping->private_lock);
4903                 WARN_ON(PageDirty(p));
4904                 eb->pages[i] = p;
4905                 if (!PageUptodate(p))
4906                         uptodate = 0;
4907
4908                 /*
4909                  * see below about how we avoid a nasty race with release page
4910                  * and why we unlock later
4911                  */
4912         }
4913         if (uptodate)
4914                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4915 again:
4916         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4917         if (ret)
4918                 goto free_eb;
4919
4920         spin_lock(&fs_info->buffer_lock);
4921         ret = radix_tree_insert(&fs_info->buffer_radix,
4922                                 start >> PAGE_CACHE_SHIFT, eb);
4923         spin_unlock(&fs_info->buffer_lock);
4924         radix_tree_preload_end();
4925         if (ret == -EEXIST) {
4926                 exists = find_extent_buffer(fs_info, start);
4927                 if (exists)
4928                         goto free_eb;
4929                 else
4930                         goto again;
4931         }
4932         /* add one reference for the tree */
4933         check_buffer_tree_ref(eb);
4934         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4935
4936         /*
4937          * there is a race where release page may have
4938          * tried to find this extent buffer in the radix
4939          * but failed.  It will tell the VM it is safe to
4940          * reclaim the, and it will clear the page private bit.
4941          * We must make sure to set the page private bit properly
4942          * after the extent buffer is in the radix tree so
4943          * it doesn't get lost
4944          */
4945         SetPageChecked(eb->pages[0]);
4946         for (i = 1; i < num_pages; i++) {
4947                 p = eb->pages[i];
4948                 ClearPageChecked(p);
4949                 unlock_page(p);
4950         }
4951         unlock_page(eb->pages[0]);
4952         return eb;
4953
4954 free_eb:
4955         WARN_ON(!atomic_dec_and_test(&eb->refs));
4956         for (i = 0; i < num_pages; i++) {
4957                 if (eb->pages[i])
4958                         unlock_page(eb->pages[i]);
4959         }
4960
4961         btrfs_release_extent_buffer(eb);
4962         return exists;
4963 }
4964
4965 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4966 {
4967         struct extent_buffer *eb =
4968                         container_of(head, struct extent_buffer, rcu_head);
4969
4970         __free_extent_buffer(eb);
4971 }
4972
4973 /* Expects to have eb->eb_lock already held */
4974 static int release_extent_buffer(struct extent_buffer *eb)
4975 {
4976         WARN_ON(atomic_read(&eb->refs) == 0);
4977         if (atomic_dec_and_test(&eb->refs)) {
4978                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4979                         struct btrfs_fs_info *fs_info = eb->fs_info;
4980
4981                         spin_unlock(&eb->refs_lock);
4982
4983                         spin_lock(&fs_info->buffer_lock);
4984                         radix_tree_delete(&fs_info->buffer_radix,
4985                                           eb->start >> PAGE_CACHE_SHIFT);
4986                         spin_unlock(&fs_info->buffer_lock);
4987                 } else {
4988                         spin_unlock(&eb->refs_lock);
4989                 }
4990
4991                 /* Should be safe to release our pages at this point */
4992                 btrfs_release_extent_buffer_page(eb);
4993 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4994                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
4995                         __free_extent_buffer(eb);
4996                         return 1;
4997                 }
4998 #endif
4999                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5000                 return 1;
5001         }
5002         spin_unlock(&eb->refs_lock);
5003
5004         return 0;
5005 }
5006
5007 void free_extent_buffer(struct extent_buffer *eb)
5008 {
5009         int refs;
5010         int old;
5011         if (!eb)
5012                 return;
5013
5014         while (1) {
5015                 refs = atomic_read(&eb->refs);
5016                 if (refs <= 3)
5017                         break;
5018                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5019                 if (old == refs)
5020                         return;
5021         }
5022
5023         spin_lock(&eb->refs_lock);
5024         if (atomic_read(&eb->refs) == 2 &&
5025             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5026                 atomic_dec(&eb->refs);
5027
5028         if (atomic_read(&eb->refs) == 2 &&
5029             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5030             !extent_buffer_under_io(eb) &&
5031             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5032                 atomic_dec(&eb->refs);
5033
5034         /*
5035          * I know this is terrible, but it's temporary until we stop tracking
5036          * the uptodate bits and such for the extent buffers.
5037          */
5038         release_extent_buffer(eb);
5039 }
5040
5041 void free_extent_buffer_stale(struct extent_buffer *eb)
5042 {
5043         if (!eb)
5044                 return;
5045
5046         spin_lock(&eb->refs_lock);
5047         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5048
5049         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5050             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5051                 atomic_dec(&eb->refs);
5052         release_extent_buffer(eb);
5053 }
5054
5055 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5056 {
5057         unsigned long i;
5058         unsigned long num_pages;
5059         struct page *page;
5060
5061         num_pages = num_extent_pages(eb->start, eb->len);
5062
5063         for (i = 0; i < num_pages; i++) {
5064                 page = eb->pages[i];
5065                 if (!PageDirty(page))
5066                         continue;
5067
5068                 lock_page(page);
5069                 WARN_ON(!PagePrivate(page));
5070
5071                 clear_page_dirty_for_io(page);
5072                 spin_lock_irq(&page->mapping->tree_lock);
5073                 if (!PageDirty(page)) {
5074                         radix_tree_tag_clear(&page->mapping->page_tree,
5075                                                 page_index(page),
5076                                                 PAGECACHE_TAG_DIRTY);
5077                 }
5078                 spin_unlock_irq(&page->mapping->tree_lock);
5079                 ClearPageError(page);
5080                 unlock_page(page);
5081         }
5082         WARN_ON(atomic_read(&eb->refs) == 0);
5083 }
5084
5085 int set_extent_buffer_dirty(struct extent_buffer *eb)
5086 {
5087         unsigned long i;
5088         unsigned long num_pages;
5089         int was_dirty = 0;
5090
5091         check_buffer_tree_ref(eb);
5092
5093         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5094
5095         num_pages = num_extent_pages(eb->start, eb->len);
5096         WARN_ON(atomic_read(&eb->refs) == 0);
5097         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5098
5099         for (i = 0; i < num_pages; i++)
5100                 set_page_dirty(eb->pages[i]);
5101         return was_dirty;
5102 }
5103
5104 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
5105 {
5106         unsigned long i;
5107         struct page *page;
5108         unsigned long num_pages;
5109
5110         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5111         num_pages = num_extent_pages(eb->start, eb->len);
5112         for (i = 0; i < num_pages; i++) {
5113                 page = eb->pages[i];
5114                 if (page)
5115                         ClearPageUptodate(page);
5116         }
5117         return 0;
5118 }
5119
5120 int set_extent_buffer_uptodate(struct extent_buffer *eb)
5121 {
5122         unsigned long i;
5123         struct page *page;
5124         unsigned long num_pages;
5125
5126         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5127         num_pages = num_extent_pages(eb->start, eb->len);
5128         for (i = 0; i < num_pages; i++) {
5129                 page = eb->pages[i];
5130                 SetPageUptodate(page);
5131         }
5132         return 0;
5133 }
5134
5135 int extent_buffer_uptodate(struct extent_buffer *eb)
5136 {
5137         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5138 }
5139
5140 int read_extent_buffer_pages(struct extent_io_tree *tree,
5141                              struct extent_buffer *eb, u64 start, int wait,
5142                              get_extent_t *get_extent, int mirror_num)
5143 {
5144         unsigned long i;
5145         unsigned long start_i;
5146         struct page *page;
5147         int err;
5148         int ret = 0;
5149         int locked_pages = 0;
5150         int all_uptodate = 1;
5151         unsigned long num_pages;
5152         unsigned long num_reads = 0;
5153         struct bio *bio = NULL;
5154         unsigned long bio_flags = 0;
5155
5156         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5157                 return 0;
5158
5159         if (start) {
5160                 WARN_ON(start < eb->start);
5161                 start_i = (start >> PAGE_CACHE_SHIFT) -
5162                         (eb->start >> PAGE_CACHE_SHIFT);
5163         } else {
5164                 start_i = 0;
5165         }
5166
5167         num_pages = num_extent_pages(eb->start, eb->len);
5168         for (i = start_i; i < num_pages; i++) {
5169                 page = eb->pages[i];
5170                 if (wait == WAIT_NONE) {
5171                         if (!trylock_page(page))
5172                                 goto unlock_exit;
5173                 } else {
5174                         lock_page(page);
5175                 }
5176                 locked_pages++;
5177                 if (!PageUptodate(page)) {
5178                         num_reads++;
5179                         all_uptodate = 0;
5180                 }
5181         }
5182         if (all_uptodate) {
5183                 if (start_i == 0)
5184                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5185                 goto unlock_exit;
5186         }
5187
5188         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5189         eb->read_mirror = 0;
5190         atomic_set(&eb->io_pages, num_reads);
5191         for (i = start_i; i < num_pages; i++) {
5192                 page = eb->pages[i];
5193                 if (!PageUptodate(page)) {
5194                         ClearPageError(page);
5195                         err = __extent_read_full_page(tree, page,
5196                                                       get_extent, &bio,
5197                                                       mirror_num, &bio_flags,
5198                                                       READ | REQ_META);
5199                         if (err)
5200                                 ret = err;
5201                 } else {
5202                         unlock_page(page);
5203                 }
5204         }
5205
5206         if (bio) {
5207                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
5208                                      bio_flags);
5209                 if (err)
5210                         return err;
5211         }
5212
5213         if (ret || wait != WAIT_COMPLETE)
5214                 return ret;
5215
5216         for (i = start_i; i < num_pages; i++) {
5217                 page = eb->pages[i];
5218                 wait_on_page_locked(page);
5219                 if (!PageUptodate(page))
5220                         ret = -EIO;
5221         }
5222
5223         return ret;
5224
5225 unlock_exit:
5226         i = start_i;
5227         while (locked_pages > 0) {
5228                 page = eb->pages[i];
5229                 i++;
5230                 unlock_page(page);
5231                 locked_pages--;
5232         }
5233         return ret;
5234 }
5235
5236 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5237                         unsigned long start,
5238                         unsigned long len)
5239 {
5240         size_t cur;
5241         size_t offset;
5242         struct page *page;
5243         char *kaddr;
5244         char *dst = (char *)dstv;
5245         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5246         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5247
5248         WARN_ON(start > eb->len);
5249         WARN_ON(start + len > eb->start + eb->len);
5250
5251         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5252
5253         while (len > 0) {
5254                 page = eb->pages[i];
5255
5256                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5257                 kaddr = page_address(page);
5258                 memcpy(dst, kaddr + offset, cur);
5259
5260                 dst += cur;
5261                 len -= cur;
5262                 offset = 0;
5263                 i++;
5264         }
5265 }
5266
5267 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5268                         unsigned long start,
5269                         unsigned long len)
5270 {
5271         size_t cur;
5272         size_t offset;
5273         struct page *page;
5274         char *kaddr;
5275         char __user *dst = (char __user *)dstv;
5276         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5277         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5278         int ret = 0;
5279
5280         WARN_ON(start > eb->len);
5281         WARN_ON(start + len > eb->start + eb->len);
5282
5283         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5284
5285         while (len > 0) {
5286                 page = eb->pages[i];
5287
5288                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5289                 kaddr = page_address(page);
5290                 if (copy_to_user(dst, kaddr + offset, cur)) {
5291                         ret = -EFAULT;
5292                         break;
5293                 }
5294
5295                 dst += cur;
5296                 len -= cur;
5297                 offset = 0;
5298                 i++;
5299         }
5300
5301         return ret;
5302 }
5303
5304 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5305                                unsigned long min_len, char **map,
5306                                unsigned long *map_start,
5307                                unsigned long *map_len)
5308 {
5309         size_t offset = start & (PAGE_CACHE_SIZE - 1);
5310         char *kaddr;
5311         struct page *p;
5312         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5313         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5314         unsigned long end_i = (start_offset + start + min_len - 1) >>
5315                 PAGE_CACHE_SHIFT;
5316
5317         if (i != end_i)
5318                 return -EINVAL;
5319
5320         if (i == 0) {
5321                 offset = start_offset;
5322                 *map_start = 0;
5323         } else {
5324                 offset = 0;
5325                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
5326         }
5327
5328         if (start + min_len > eb->len) {
5329                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5330                        "wanted %lu %lu\n",
5331                        eb->start, eb->len, start, min_len);
5332                 return -EINVAL;
5333         }
5334
5335         p = eb->pages[i];
5336         kaddr = page_address(p);
5337         *map = kaddr + offset;
5338         *map_len = PAGE_CACHE_SIZE - offset;
5339         return 0;
5340 }
5341
5342 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5343                           unsigned long start,
5344                           unsigned long len)
5345 {
5346         size_t cur;
5347         size_t offset;
5348         struct page *page;
5349         char *kaddr;
5350         char *ptr = (char *)ptrv;
5351         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5352         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5353         int ret = 0;
5354
5355         WARN_ON(start > eb->len);
5356         WARN_ON(start + len > eb->start + eb->len);
5357
5358         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5359
5360         while (len > 0) {
5361                 page = eb->pages[i];
5362
5363                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5364
5365                 kaddr = page_address(page);
5366                 ret = memcmp(ptr, kaddr + offset, cur);
5367                 if (ret)
5368                         break;
5369
5370                 ptr += cur;
5371                 len -= cur;
5372                 offset = 0;
5373                 i++;
5374         }
5375         return ret;
5376 }
5377
5378 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5379                          unsigned long start, unsigned long len)
5380 {
5381         size_t cur;
5382         size_t offset;
5383         struct page *page;
5384         char *kaddr;
5385         char *src = (char *)srcv;
5386         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5387         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5388
5389         WARN_ON(start > eb->len);
5390         WARN_ON(start + len > eb->start + eb->len);
5391
5392         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5393
5394         while (len > 0) {
5395                 page = eb->pages[i];
5396                 WARN_ON(!PageUptodate(page));
5397
5398                 cur = min(len, PAGE_CACHE_SIZE - offset);
5399                 kaddr = page_address(page);
5400                 memcpy(kaddr + offset, src, cur);
5401
5402                 src += cur;
5403                 len -= cur;
5404                 offset = 0;
5405                 i++;
5406         }
5407 }
5408
5409 void memset_extent_buffer(struct extent_buffer *eb, char c,
5410                           unsigned long start, unsigned long len)
5411 {
5412         size_t cur;
5413         size_t offset;
5414         struct page *page;
5415         char *kaddr;
5416         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5417         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5418
5419         WARN_ON(start > eb->len);
5420         WARN_ON(start + len > eb->start + eb->len);
5421
5422         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5423
5424         while (len > 0) {
5425                 page = eb->pages[i];
5426                 WARN_ON(!PageUptodate(page));
5427
5428                 cur = min(len, PAGE_CACHE_SIZE - offset);
5429                 kaddr = page_address(page);
5430                 memset(kaddr + offset, c, cur);
5431
5432                 len -= cur;
5433                 offset = 0;
5434                 i++;
5435         }
5436 }
5437
5438 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5439                         unsigned long dst_offset, unsigned long src_offset,
5440                         unsigned long len)
5441 {
5442         u64 dst_len = dst->len;
5443         size_t cur;
5444         size_t offset;
5445         struct page *page;
5446         char *kaddr;
5447         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5448         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5449
5450         WARN_ON(src->len != dst_len);
5451
5452         offset = (start_offset + dst_offset) &
5453                 (PAGE_CACHE_SIZE - 1);
5454
5455         while (len > 0) {
5456                 page = dst->pages[i];
5457                 WARN_ON(!PageUptodate(page));
5458
5459                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5460
5461                 kaddr = page_address(page);
5462                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5463
5464                 src_offset += cur;
5465                 len -= cur;
5466                 offset = 0;
5467                 i++;
5468         }
5469 }
5470
5471 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5472 {
5473         unsigned long distance = (src > dst) ? src - dst : dst - src;
5474         return distance < len;
5475 }
5476
5477 static void copy_pages(struct page *dst_page, struct page *src_page,
5478                        unsigned long dst_off, unsigned long src_off,
5479                        unsigned long len)
5480 {
5481         char *dst_kaddr = page_address(dst_page);
5482         char *src_kaddr;
5483         int must_memmove = 0;
5484
5485         if (dst_page != src_page) {
5486                 src_kaddr = page_address(src_page);
5487         } else {
5488                 src_kaddr = dst_kaddr;
5489                 if (areas_overlap(src_off, dst_off, len))
5490                         must_memmove = 1;
5491         }
5492
5493         if (must_memmove)
5494                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5495         else
5496                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5497 }
5498
5499 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5500                            unsigned long src_offset, unsigned long len)
5501 {
5502         size_t cur;
5503         size_t dst_off_in_page;
5504         size_t src_off_in_page;
5505         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5506         unsigned long dst_i;
5507         unsigned long src_i;
5508
5509         if (src_offset + len > dst->len) {
5510                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5511                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5512                 BUG_ON(1);
5513         }
5514         if (dst_offset + len > dst->len) {
5515                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5516                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5517                 BUG_ON(1);
5518         }
5519
5520         while (len > 0) {
5521                 dst_off_in_page = (start_offset + dst_offset) &
5522                         (PAGE_CACHE_SIZE - 1);
5523                 src_off_in_page = (start_offset + src_offset) &
5524                         (PAGE_CACHE_SIZE - 1);
5525
5526                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5527                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5528
5529                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5530                                                src_off_in_page));
5531                 cur = min_t(unsigned long, cur,
5532                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5533
5534                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5535                            dst_off_in_page, src_off_in_page, cur);
5536
5537                 src_offset += cur;
5538                 dst_offset += cur;
5539                 len -= cur;
5540         }
5541 }
5542
5543 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5544                            unsigned long src_offset, unsigned long len)
5545 {
5546         size_t cur;
5547         size_t dst_off_in_page;
5548         size_t src_off_in_page;
5549         unsigned long dst_end = dst_offset + len - 1;
5550         unsigned long src_end = src_offset + len - 1;
5551         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5552         unsigned long dst_i;
5553         unsigned long src_i;
5554
5555         if (src_offset + len > dst->len) {
5556                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5557                        "len %lu len %lu\n", src_offset, len, dst->len);
5558                 BUG_ON(1);
5559         }
5560         if (dst_offset + len > dst->len) {
5561                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5562                        "len %lu len %lu\n", dst_offset, len, dst->len);
5563                 BUG_ON(1);
5564         }
5565         if (dst_offset < src_offset) {
5566                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5567                 return;
5568         }
5569         while (len > 0) {
5570                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5571                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5572
5573                 dst_off_in_page = (start_offset + dst_end) &
5574                         (PAGE_CACHE_SIZE - 1);
5575                 src_off_in_page = (start_offset + src_end) &
5576                         (PAGE_CACHE_SIZE - 1);
5577
5578                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5579                 cur = min(cur, dst_off_in_page + 1);
5580                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5581                            dst_off_in_page - cur + 1,
5582                            src_off_in_page - cur + 1, cur);
5583
5584                 dst_end -= cur;
5585                 src_end -= cur;
5586                 len -= cur;
5587         }
5588 }
5589
5590 int try_release_extent_buffer(struct page *page)
5591 {
5592         struct extent_buffer *eb;
5593
5594         /*
5595          * We need to make sure noboody is attaching this page to an eb right
5596          * now.
5597          */
5598         spin_lock(&page->mapping->private_lock);
5599         if (!PagePrivate(page)) {
5600                 spin_unlock(&page->mapping->private_lock);
5601                 return 1;
5602         }
5603
5604         eb = (struct extent_buffer *)page->private;
5605         BUG_ON(!eb);
5606
5607         /*
5608          * This is a little awful but should be ok, we need to make sure that
5609          * the eb doesn't disappear out from under us while we're looking at
5610          * this page.
5611          */
5612         spin_lock(&eb->refs_lock);
5613         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5614                 spin_unlock(&eb->refs_lock);
5615                 spin_unlock(&page->mapping->private_lock);
5616                 return 0;
5617         }
5618         spin_unlock(&page->mapping->private_lock);
5619
5620         /*
5621          * If tree ref isn't set then we know the ref on this eb is a real ref,
5622          * so just return, this page will likely be freed soon anyway.
5623          */
5624         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5625                 spin_unlock(&eb->refs_lock);
5626                 return 0;
5627         }
5628
5629         return release_extent_buffer(eb);
5630 }